Content-Based Image Retrieval using Local Binary Curvelet Co-occurrence Pattern—A Multiresolution Technique

Content-Based Image Retrieval using Local Binary Curvelet Co-occurrence Pattern—A... Abstract With the growth of various image-capturing devices, image acquisition is no longer a difficult task. As this technology is flourishing, various types of complex images are being produced. In order to access a large number of images stored in database easily, the images must be properly organized. Field of image retrieval attempts to solve this problem. As the complex images are being produced, processing them using single-resolution techniques is not sufficient as these images may contain varying levels of details. This paper proposes a novel multiresolution descriptor, local binary curvelet co-occurrence pattern, to achieve the task of content-based image retrieval. Curvelet transform of grayscale image is computed followed by computation of local binary pattern of resulting curvelet coefficients. Finally, feature vector is constructed using grey-level co-occurrence matrix which is matched with the feature vector of database images. The proposed descriptor combines the properties of local pattern and multiresolution technique of curvelet transform, and efficiently covers curvilinear and geometrical structures present in the image. Performance of the proposed method is measured in terms of precision and recall and is tested on five benchmark datasets consisting of natural images. The proposed method has been compared with single and multiresolution techniques as well as with some of the other state-of-the-art image retrieval methods. The experimental results clearly demonstrate that the proposed method produces high retrieval accuracy and outperforms other techniques in terms of precision and recall. 1. INTRODUCTION Development of a large number of image-capturing devices has led to the proliferation of huge amount of images. With the development of low-cost hand held devices, capturing of image is much easier than it was a few years ago. This has created a huge repository of unorganized images. Organizing a large number of images is quite challenging and scientists across the world are working on various techniques to accomplish this task. All these issues have made image indexing and retrieval an important problem of computer vision. Primarily, there are two ways of achieving this task. The first approach is annotation of images on the basis of text and keywords. This approach requires annotation of a large number of images to make the task of indexing and retrieval easy. Although this method of retrieval is easy, it requires manual annotation of a large number of images which is a difficult task. Also, this method fails to retrieve visually similar images. The second approach is retrieval of images on the basis of contents of image and it requires no manual tagging and retrieves visually similar images. Content-based image retrieval (CBIR), also named as content-based visual information retrieval (CBVIR), refers to retrieval of images on the basis of features present in the image. In CBIR systems, an image or its sketch is supplied as a query to retrieve images visually similar to the query image. The system extracts features from the image and constructs a feature vector. This feature vector is matched with those of database images to retrieve visually similar images [1]. The construction of efficient feature vector is a challenging problem in image retrieval as the effectiveness of feature vector determines the success of retrieval. Image features can be broadly classified into two categories, primary features such as colour, texture and shape, and semantic features such as type of image. Processing of primary features has been extensively used for image retrieval. CBIR using colour has mostly been used in the form of colour histogram as it is easy to construct [2]. Apart from colour histogram, colour has also been used as a feature in colour correlogram [3] and colour coherence vector [4]. Thus, colour has been a popular feature for retrieval as it is a visible descriptor and invariant to certain geometrical transformations. Texture is another feature which has been used for image indexing [5]. Texture determines properties such as coarseness, roughness and smoothness of a surface. Texture has been exploited through Fourier transform [5] as well as grey-level co-occurrence matrix (GLCM) [6]. Apart from colour and texture, shape is another feature which has been exploited in the form of moments [7] and polygonal shape [8]. Shape features are generally used after segmentation [9]. Lack of good segmentation algorithm makes shape feature less popular than colour and texture. Early image retrieval systems used single feature for feature vector construction and this technique remained popular for a long time. However, as complexity in nature of images started increasing, single feature started proving to be insufficient. Hence, the trend of CBIR shifted to combination of features for feature vector construction. The combination of colour and texture [10], colour and shape [11], texture and shape [12], and colour, texture and shape [13] are some of the approaches that have incorporated multiple features for feature vector construction. Multiple features tend to combine the advantages of more than one feature and are able to extract more features as compared with single feature. Modern methods of CBIR employ human perception analysis and understanding to retrieve visually similar images. Liu et al. [14] described microstructure descriptor (MSD) to extract colour, texture and shape through single microstructure feature. Zhang et al. [15] proposed hybrid information descriptor that combines high-level and low-level features for image retrieval. Liu et al. [16] proposed visual attention model to perform human perception analysis for CBIR. All the methods discussed above have been applied on single resolution of image. An image is generally a complex structure and consists of varying levels of details which the single-resolution processing fails to capture. This fact motivated the researchers to perform CBIR in multiresolution analysis framework. The advantage of performing multiresolution analysis is that it considers varying levels of details for feature vector construction and features that are left undetected at one resolution get considered at another level. A number of multiresolution techniques have been used where multiple levels of image have been exploited to construct feature vector for retrieval. This paper proposes a new multiresolution descriptor, known as local binary curvelet co-occurrence pattern (LBCCP), for image retrieval. The proposed method in this paper exploits local binary pattern (LBP) at multiple scales through curvelet transform followed by construction of GLCM for generation of feature vector. Curvelet, being a multiresolution technique, captures curvilinear structure and sharp edges which wavelet transform fails to capture. The reason for this is that curvelets are highly anisotropic in nature and use fewer coefficients to represent edges unlike wavelets. The feature vector is constructed through GLCM as it provides spatial information by considering occurrence of pixel pairs which other techniques fail to do. In this way, the proposed descriptor exploits spatial information about local feature at multiple resolutions to construct efficient feature vector. The proposed descriptor captures local information at different angles at multiple scales of image which helps in gathering complex details from natural images. This combination of local descriptor along with multiresolution technique such as curvelet helps in detecting sharp curvilinear structures which are not detected through single-resolution processing techniques. For multiresolution processing of image, curvelet transform has been used for the following reasons: first, due to its anisotropic nature, curvelets are able to capture sharp curvilinear structures present in natural images which are not captured by wavelets due to their isotropic nature. Second, natural images are complex in nature and contain varying levels of details as compared with medical and forensic images and consist of complex foreground as well as background details. Such complex details are captured by curvelet transform more effectively than wavelet transform. The construction of feature has been done through GLCM which attempts to capture spatial distribution of intensity values. This helps in providing information about structural arrangement of pixels which are not provided by other features such as histogram. The combination of local pattern with multiresolution technique such as curvelets attempts to capture complex details such as sharp curvilinear structures which single-resolution processing techniques and certain other multiresolution techniques fail to gather. Rest of the paper is organized as follows. Section 2 discusses some of the related work in the field of image retrieval. Section 3 gives brief background of curvelet transform, LBP and GLCM. Section 4 discusses the proposed method. Section 5 discusses experiment and results and Section 6 concludes the paper. 2. RELATED WORK Since the inception of the term image retrieval, several algorithms have been proposed to improve the accuracy of retrieval. Early methods of retrieval were focused mainly on primary features, namely, colour, texture and shape. These features were used either as a single feature or in combination with each other. Most of these methods exploited single resolution of image. A natural image always contains varying levels of details and in order to exploit those varying levels of details, single-resolution processing is not sufficient. This limitation has been overcome by multiresolution processing of image which considers more than one resolution to capture varying levels of details. Several techniques exploiting multiple resolutions of image have been proposed. One of them is wavelet which has been in use for a long time [17]. Wavelet has been combined with colour in the form of wavelet correlogram [18]. This concept has been further improved by adding optimization feature in [19] which not only improves retrieval speed but also retrieval accuracy. Lemard et al. [20] proposed image retrieval based on wavelets by using signatures constructed from wavelet coefficients. Apart from using as a single feature, wavelet has been combined with other features also. Agarwal et al. [21] proposed combination of wavelet and MSD. This method computed MSD of à trous wavelet coefficients to perform retrieval. Fu et al. [22] combined wavelet with shape feature, Zernike moments, for image indexing. The combination of wavelet and shape feature has also been discussed in [23]. Vo et al. [24] exploited the concept of relative phase in complex wavelet domain for texture image retrieval and segmentation. Thus, wavelet transform is a popular multiresolution technique and has been exploited much more than any other multiresolution technique for image retrieval. However, wavelets suffer from certain drawbacks. Due to their non-geometrical nature, wavelets do not effectively represent objects that consist of very sharp edges and curves. Due to this limitation, wavelets fails to effectively exploit regularity of edge curve. Also, due to their non-anisotropic nature, wavelets do not capture details of any region at multiple orientations but consider only limited orientations at all scales. Due to the emergence of various image-capturing devices, more and more complex natural images, consisting of variety of objects, are being produced everyday. Because of the above-mentioned limitations, wavelets fail to gather curvilinear structures and features at edges efficiently. These drawbacks can be overcome by using another multiresolution technique called curvelets [25]. Curvelets are collections of multiscale ridgelets at multiple orientations and scales [25]. Unlike wavelets, curvelets are geometrical in nature and highly anisotropic. These properties enable curvelets to compute coefficients of sharp edges and curves which wavelets fail to achieve. Sumana et al. [26] used curvelet coefficients for image retrieval. This technique applied curvelet transform on texture images and computed low-order statistics to construct feature vector. The method is simple and extracts low-order statistics efficiently, but fails to exploit multiresolution aspect of curvelet transform. Gonde et al. [27] modified basic structure of curvelet transform by introducing Gabor wavelet instead of à trous wavelet and constructed feature vector through vocabulary tree for retrieval. This technique performed better than both original curvelet transform-based technique and wavelet-based technique. However, this method produced low retrieval accuracy. Youssef [28] proposed curvelet-based image retrieval scheme named ICTEDT-CBIR. This method combines colour and texture features along with curvelet transform for image indexing and retrieval. Zhang et al. [29] proposed rotation invariant curvelet feature to exploit texture feature of image. This method applies curvelet transform on region-based colour image to construct feature vector which is rotation invariant. Murtagh and Starck [30] proposed the concept of curvelet moments similar to wavelet moments. Sumana et al. [31] proposed new feature based on texture named generalized Gaussian density texture feature for retrieval. Das et al. [32] compared wavelet and curvelet transform for image retrieval by combining colour and texture features with wavelet and curvelet transform. Curvelet transform has been used not only for image retrieval but also for other image processing applications such as fingerprint recognition [33] and denoising [25]. Multiresolution processing of image has also been applied for texture cartoon analysis of image in [42] and its applications in the field of morphological diversity has been discussed in [43]. A common drawback in most of the above-discussed methods based on curvelet transform is that they fail to discuss and apply the multiresolution property of curvelet transform. Curvelets decompose image into multiple scales and compute coefficients at different orientations. Each scale represents varying levels of details so that features left undetected at one level get considered at another level. This paper attempts to overcome the discussed drawbacks by combining curvelet transform and LBP along with GLCM for feature vector construction, in the form of a novel descriptor, called LBCCP. This descriptor gathers local information at multiple scales and orientations. The feature vector is constructed for each scale separately and retrieval is performed at each scale and then combined at the end thereby producing final retrieval results. This approach considers not only multiresolution aspect of curvelet transform which other methods fail to do but also produces high retrieval accuracy. 3. CURVELET TRANSFORM, LBP, AND GLCM The concept of multiresolution analysis dates back to late 1987 [34]. Although wavelets came into existence early, their multiresolution analysis characteristics were discovered later [34]. Multiresolution analysis has been mainly practiced through Gabor wavelet transform and discrete wavelet transform (DWT). DWT computes coefficients in three directions: horizontal, vertical and diagonal. DWT has proved to be quite useful for various applications of computer vision and has provided promising results [33]. However, wavelets possess certain limitations [25]: Natural images contain highly anisotropic elements which may not be effectively represented by wavelets because they are non-geometrical in nature and fail to represent regularity of edge curve. As the resolution of image becomes finer, a large number of wavelet coefficients are required to represent important edges in the image. This results in repetition of edges of images at multiple scales. At fine scales, a large number of wavelet coefficients are required to reconstruct edges properly. This may cause redundancy of edges scale after scale. Due to this, the features of an image may not be represented effectively. DWT (with Haar wavelet as mother wavelet) computes coefficients in three directions only (horizontal, vertical and diagonal directions). Hence, their ability to resolve directional features is limited. Therefore, wavelets fail to detect curved singularities effectively. To overcome these limitations, other multiresolution techniques such as ridgelet and curvelet were proposed. Ridgelet transform is an important element of curvelet transform. Continuous ridgelet transform provides a sparse representation of smooth as well as of perfectly straight edges. 2D ridgelet transform provides representation of arbitrary bivariate functions f(x1,x2) by superimposition of the elements of the form a−1/2ψ((x1cosθ+x2sinθ−b)/a), where ψ is a wavelet, a > 0 is a scale parameter, θ is an orientation parameter and b is a location scalar parameter. Ridgelets are constant along the lines x1cosθ+x2sinθ, and are wavelets along orthogonal direction. 3.1. Ridgelet transform Let a smooth univariate function ψ:R→R, with sufficient decay and satisfying admissibility condition, be defined as below [25]   ∫|ψˆ(ξ)|2/|ξ|2dξ<∞ (1)which holds if ∫ψ(t)dt=0, that is ψ has a vanishing mean. Let ψ be normalized so that ∫|ψˆ(ξ)|2|ξ|−2dξ=1. Let the bivariate ridgelet ψa,b,θ:R2→R2 be defined as   ψa,b,θ(x)=a−1/2ψ((x1cosθ+x2sinθ−b)/a (2)where a>0, b∈R and θ∈[0,2π). This function is constant along lines (x1cosθ+x2sinθ)=const. Traverse to these ridges, it is a wavelet. For a bivariate function f(x), ridgelet coefficients are defined as   Rf(a,b,θ)=∫ψa,b,θ(x)f(x)dx The reconstruction formula   f(x)=∫02π∫−∞∞∫0∞Rf(a,b,θ)ψa,b,θ(x)daa3dbdθ4π (3)is valid for functions which are both integrable and square integrable. Ridgelet coefficients are calculated by considering ridgelet analysis as wavelet analysis in Radon domain. Radon transform of an object f is defined as collection of line integrals denoted by   Rf(θ,t)=∫f(x1,x2)δ(x1cosθ+x2sinθ−t)dx1dx2 (4)where δ is the Dirac function. The ridgelet coefficients Rf(a,b,θ) of an object f are given by analysis of Radon transform through Rf(a,b,θ)=∫Rf(θ,t)a−1/2ψ((t−b)/a)dt. Edges are curved structure instead of straight lines. The concept of ridgelet transform alone cannot represent edges efficiently. However, edges are almost straight line at fine scales and in order to capture them, ridgelets can be implemented in a localized manner at fine scale. Curvelets are multiscale ridgelets which can accomplish this. Curvelets are defined at all scales, locations and orientations. 3.2. Curvelet transform Discrete curvelet transform of a continuum function f(x1,x2) uses dyadic sequence of scales and a bank of filters (P0f,Δ1f,Δ2f,…) having property that bandpass filter Δs is constructed near frequencies [22s,22s+2]. The curvelet decomposition consists of the following steps [25]: Decomposition of object f into subbands. Smooth windowing of each subband into squares of an appropriate scales. Renormalization of each resulting square to unit scale. Analysis of each square through discrete ridgelet transform. Curvelet transform has the following advantages over wavelet transform: Curvelets are highly anisotropic and efficiently represent sharp edges and curvilinear structures over wavelets. Curvelets use fewer coefficients to represent edges in an image properly. Curvelet transform computes coefficients at multiple scales, locations and orientations. Therefore, curvelets are able to detect curved singularities unlike wavelets. 3.3. LBP LBP was originally proposed by Ojala et al. [35]. LBP operator works in a 3 × 3 pixel block of an image. The pixels are thresholded by the value of centre pixel. The thresholded values are multiplied by the weights given to corresponding pixel values. The LBP operator takes 3 × 3 surrounding of a pixel and Generates a binary 1 if the neighbour value is greater than or equal to centre value. Generates a binary 0 if the neighbour value is less than the centre value. The eight neighbours of the centre can then be represented with an 8-bit number. LBP has the following important properties: It is a simple and efficient local descriptor for describing textures. It encodes the relationship between the grey value of centre pixel and surrounding neighbouring pixels into 0 and 1. It is helpful in extracting local information of an image. As a local feature, when it is combined with global feature acts as a powerful feature vector [36]. 3.4. GLCM GLCM was proposed by Haralick et al. [37]. It is a statistical method for texture analysis of an image. GLCM is a matrix which provides information about how frequently pixel pairs holding specific values and in a specified spatial relationship occur in an image. The occurrence of pixels is in a particular distance and direction. The size of the matrix is determined by maximum grey-level intensity values. GLCM helps in extraction of certain important statistical features. GLCM also helps in determining the spatial arrangement of pixels in an image. Computation of GLCM is demonstrated in Fig. 1 with the help of an example. Figure 1(a) shows the original matrix and Fig. 1(b) shows GLCM constructed for original matrix. In Fig. 1(b), the topmost row and the leftmost column represent pixel values that appear in original matrix. The entries in GLCM represent the number of times pixel pairs appear in original matrix. For example, the pixel pair (1, 1) appears three times in original matrix which is shown in GLCM. Figure 1. View largeDownload slide (a) Original matrix. (b) GLCM for original matrix. Figure 1. View largeDownload slide (a) Original matrix. (b) GLCM for original matrix. 3.5. LBCCP Modern image-capturing devices are capable of producing high-resolution images due to which the images have become more complex more than they were earlier. They contain small as well as large size objects and low- as well as high-resolution objects. In order to perform retrieval of such types of images, single-resolution processing proves to be insufficient as different types of curved structures may be present in the image and processing at single resolution fails to capture them efficiently. This may lead to weak feature vector construction which may result in low retrieval accuracy [7, 23]. The proposed LBCCP descriptor tends to overcome these limitations by gathering local information in an image through LBP at multiple scales with the help of curvelet transform. Natural images have complex and irregular texture. In order to capture these features, LBP proves to be an effective feature descriptor. LBP captures local information from an image by computing relationship of an intensity value with its neighbourhood pixels. This helps in extracting information of structural arrangement of pixels. However, LBP fails to gather directional information since it does not incorporate any technique which is helpful in gathering details from different directions. The combination of LBP with curvelet transform tends to overcome this limitation. Curvelet transform being highly anisotropic in nature gathers sharp curvilinear and geometric structures at multiple orientations and scales. This helps in capturing complex details more effectively than LBP exploited at single resolution of image. Curvelet transform decomposes an image into multiple resolutions for gathering features. The advantage of this is that features left undetected at one scale get detected at another scale. The spatial distribution of LBP codes at different scales of images are gathered by constructing GLCM. This descriptor helps in getting texture information at multiple resolutions by computing LBP of curvilinear structures and sharp edges which other techniques such as single-resolution processing and other multiresolution techniques fail to gather. The proposed LBCCP descriptor has the following important properties which are useful for image retrieval: Due to its anisotropic nature, it effectively represents curvilinear structures and sharp edges. It uses fewer coefficients to reconstruct edges in an image. In this way, it reduces redundancy of coefficients thereby constructing efficient feature vector for retrieval at multiple scales unlike wavelet-based technique. It helps in construction of feature vector at multiple scales, orientations and locations due to which it is capable to capture complex geometrical structure which wavelet-based technique fails to do. It effectively gathers spatial distribution of LBP codes at multiple scales through GLCM which other techniques such as histogram fail to do. 4. THE PROPOSED METHOD The proposed method consists of four steps: Computation of curvelet coefficients of grayscale image. Computation of LBP codes of curvelet coefficients computed in Step 1. Computation of GLCM. Similarity measurement. The schematic diagram for the proposed method is shown in Fig. 2. Figure 2. View largeDownload slide Schematic diagram for the proposed method. Figure 2. View largeDownload slide Schematic diagram for the proposed method. 4.1. Computation of curvelet coefficient The first step of the proposed method is computation of curvelet coefficients. Candes et al. [38] proposed two versions of fast discrete curvelet transforms, namely, unequally spaced fast Fourier transform (USFFT) which is based on unequally fast Fourier transforms, and fast discrete curvelet transform warping (FDCT_WARPING) which is based on wrapping of specially selected Fourier samples. Both versions of curvelet transforms are fast and efficiently compute curvelet coefficients. The proposed method uses USFFT curvelet transform for computation of curvelet coefficients as it is easy to understand and compute. Application of curvelet transform on grayscale image of size 256 × 256 decomposes image into five cells using formula Nscales=ceil(log2(min(N1,N2))−3) [33] from coarse to fine scales:   ({1×1},{1×32},{1×32},{1×64},{1×1}) Here Nscales denotes the number of scales or resolutions of image, N1 and N2 denote the size of image. These five cells further consist of sub-cells that contain curvelet coefficients. These cells denote coefficients computed at different scales, locations and orientations. e.g. {1 × 32} cell denotes that the cell consists of 32 sub-cells of different sizes which hold values of curvelet coefficients computed at different orientations. Table 1 shows the number of sub-cells each cell contains for a 256 × 256 and 128 × 128 image. Table 1. Number of sub-cells each cell contains. Cell  Number of sub-cells  (a) For 256 × 256 image   {1 × 1}  1   {1 × 32}  32   {1 × 32}  32   {1 × 64}  64   {1 × 1}  1  (b) For 128 × 128 image   {1 × 1}  1   {1 × 32}  32   {1 × 64}  64   {1 × 1}  1  Cell  Number of sub-cells  (a) For 256 × 256 image   {1 × 1}  1   {1 × 32}  32   {1 × 32}  32   {1 × 64}  64   {1 × 1}  1  (b) For 128 × 128 image   {1 × 1}  1   {1 × 32}  32   {1 × 64}  64   {1 × 1}  1  When discrete curvelet transform is applied on 2D grayscale image, it decomposes image into different levels of resolution as shown in Table 1. Each cell is considered as a level of resolution and exploited for constructing feature vector. Each level of resolution of curvelet transform consists of coefficients computed at different locations and orientations. The coarse scale consists of low-frequency coefficients and fine scale consists of high-frequency coefficients [32]. The different levels of resolution attempt to gather small as well as large size objects and low as well as high-resolution objects. The LBP codes of resulting curvelet coefficients are computed for each level of resolution separately. The construction of feature vector is done through GLCM. Feature vector for each cell is computed separately and is used for performing retrieval of similar images. This produces a set of similar images for each level separately which are combined to produce final image set. 4.2. Computation of LBP LBP gives description of the surroundings of a pixel by producing a bit-code from binary derivatives of a pixel. LBP takes 3 × 3 neighbourhood of a pixel and generates a binary 1 if the neighbour is greater than or equal to the centre pixel or a binary 0 if the neighbour is less than the centre pixel. This generates an 8-bit code known as LBP code. LBP codes for each cell of curvelet coefficients are computed separately. For a 256 × 256 image, the total number of cells is 5. Each cell further consists of sub-cells which contains curvelet coefficients. LBP codes of coefficients in each case of these sub-cells is computed and stored in separate matrices. 4.3. Computation of GLCM After computation of LBP codes of curvelet coefficients, the next step is construction of feature vector. Feature vector is constructed through GLCM. GLCM is computed for each LBP matrix separately. GLCM determines frequency of co-occurrence of pixels pairs. This gives information about spatial distribution of intensity values in an image which helps in determining structural arrangement of pixels in an image. All these information are not provided by other features such as histogram which provides information about frequency of intensity values only. GLCM in 0° angle with Distance 1 has been used in the proposed method. GLCM has been scaled to size 8 × 8 matrix in the proposed method. 4.4. Similarity measurement Similarity measurement is done for each image of dataset to retrieve visually similar images. The similarity measurement is done by measuring distance between feature vector of query image and database images. The feature vector in the proposed method is constructed through computation of GLCM of LBP codes. Hence, in the proposed method, similarity measurement is done by measuring distance between GLCM of query image and database images. Let the GLCM of query image be denoted by GQ=(GQ1,GQ2,…,GQn) and let the GLCM of database images be denoted by GDB=(GDB1,GDB2,…,GDBn). Then the Euclidean distance between query image and database image is given as   D(GQ,GDB)=∑(GQi−GDBi)2,i=1,2,…,n (5)where D denotes the distance between feature vectors, GQ denotes the feature vector of query image and GDB denotes the feature vector of database image. 5. EXPERIMENT AND RESULTS To perform experiment using the proposed method, images from following five benchmark datasets have been used. These datasets consist of a wide variety of images and are widely used for evaluation of image retrieval. Dataset 1 (Corel-1K) The first dataset used in this experiment is Corel-1K dataset [39]. It consists of 1000 images. The images in this dataset are classified into 10 different categories, namely, Africans, Beaches, Buildings, Buses, Dinosaurs, Elephants, Flowers, Horses, Mountains and Food. Each category consists of 100 images. The size of each image is either 256 × 384 or 384 × 256. Dataset 2 (Olivia-2688) The second dataset used to measure the performance of the proposed method is Olivia-2688 dataset [40]. It consists of 2688 images. The images in this dataset are divided into eight categories, namely, Coast, Forest, Highway, Inside City, Mountain, Open Country, Street and Tall Building. Each category consists of different numbers of images ranging from maximum 410 to minimum 260. The size of each image is 256 × 256. Dataset 3 (Corel-5K) The third dataset used in this experiment is Corel-5K dataset [41]. It consists of 5000 images. The images in this dataset are divided into 50 categories consisting of different types of images in various categories ranging from animals, human beings to sunsets, card, etc. Each category consists of 100 images. The size of each image is either 187 × 128 or 128 × 187. Dataset 4 (Corel-10K) The fourth dataset used for testing the proposed method is Corel-10K dataset [41] which is an extension of Corel-5K dataset. It consists of 10 000 images. The images in this dataset are divided into hundred categories consisting of a wide variety of images. Each category consists of 100 images. Each image is of size 187 × 128 or 128 × 187. Dataset 5 (GHIM-10K) The fifth dataset used in this experiment is GHIM-10K dataset [41]. It consists of 10 000 images. The images in this dataset are divided into 20 categories consisting of various types of images such as horses, insects and flowers. Each category consists of 500 images. The size of each image is either 300 × 400 or 400 × 300. Each image of datasets Corel-1K, Olivia-2688, GHIM-10K has been rescaled to size 256 × 256 (28 × 28) and images of datasets Corel-5K and Corel-10K to size 128 × 128 (27 × 27) to ease the computation. Sample images from each dataset are shown in Fig. 3. Each image of all datasets is taken as query image. If the retrieved images belong to the same category as that of the query image, the retrieval is considered to be successfully, otherwise the retrieval fails. Figure 3. View largeDownload slide Sample images from datasets. Figure 3. View largeDownload slide Sample images from datasets. 5.1. Performance evaluation Performance of the proposed method has been evaluated in terms of precision and recall. Precision is defined as the ratio of the total number of relevant images retrieved to the total number of images retrieved. Mathematically, precision can be formulated as   P=IRTR (6)where IR denotes the total number of relevant images retrieved and TR denotes the total number of images retrieved. Recall is defined as the ratio of total number of relevant images retrieved to the total number of relevant images in the database. Mathematically, recall can be formulated as   R=IRCR (7)where IR denotes the total number of relevant images retrieved and CR denotes the total number of relevant images in the database. In this experiment, TR = 10 and the value of CR varies for different datasets. The value of CR depends on the total number of images in each category of image of datasets. In this experiment, for Corel-1K, Corel-5K and Corel-10K datasets, the value of CR is 100. For GHIM-10K dataset, the value of CR is 500 and for Olivia-2688, the value of CR is different for each category of image depending on the total number of images in each category. 5.2. Retrieval results at multiple resolutions To perform the experiments, the images of datasets Corel-1K, Olivia-2688 and GHIM-10K have been rescaled to size 256 × 256 and images of datasets Corel-5K and Corel-10K have been rescaled to size 128 × 128, to ease the computation. First step of the proposed method is the computation of curvelet transform coefficients of grayscale images followed by computation of LBP codes of resulting coefficients. Finally, feature vector is constructed using GLCM. Application of curvelet transform on grayscale images produces five cells of coefficients for 256 × 256 image and four cells of coefficients for 128 × 128 image. Each of these cells further contains sub-cells of different sizes that hold curvelet coefficients of grayscale image. LBP codes for all sub-cells are computed separately followed by construction of GLCM of each cell. In this experiment, similarity measurement for each of these sub-cells is done separately. This produces five sets of similar images. Union of all these sets is taken to produce a final set of similar images. Recall is computed by counting the total number of relevant images in the final set. Similarly, for precision, top n matches for each set is counted and then union operation is applied on all sets to produce final image set. Mathematically, this can be stated as follows. Let f1,f2,…,fm be the set of similar images obtained from feature vector of a sub-cell. Then, the final set of similar images denoted by fRS is given by   fRS=f1∪f2∪⋯∪fm (8) Similarly, let f1n,f2n,…,fmn be set of top n images obtained from feature vector of a sub-cell. Then, the final set of top n images is denoted by fPSn is given as   fPSn=f1n∪f2n∪⋯∪fmn (9) The above procedure is repeated for all cells of a resolution. In every level, the relevant image set of previous level is also considered and is combined with the current level to produce relevant set for that level. The proposed method has been implemented in MATLAB R2013a on a PC having Windows 8.1 Pro operating system, Intel Core 15-4570 processor at 3.20 GHz and 4 GB of RAM. The average retrieval time taken by the proposed method for 100 images is 0.043 s. Table 2 shows the performance of the proposed method on all datasets used (Corel-1K, Olivia-2688, Corel-5K, Corel-10K, GHIM-10K) at different levels. Figures 4 and 5 show the plots between precision vs. dataset and recall vs. dataset, respectively, for the proposed method on all five datasets for different levels of resolutions. Figure 6 shows the plots between precision and recall for different levels of resolution for the proposed method on all five datasets. Table 2. Average precision and recall for different levels of resolution for all datasets.   Recall (%)  Precision (%)  (a) Corel-1K dataset   Level 1  20.73  37.09   Level 2  86.31  100   Level 3  98.02  100   Level 4  99.02  100   Level 5  99.07  100  (b) Olivia-2688 dataset   Level 1  25.86  45.49   Level 2  91.02  100   Level 3  91.32  100   Level 4  99.96  100   Level 5  99.96  100  (c) Corel-5K dataset   Level 1  9.22  22.86   Level 2  36.41  78.19   Level 3  58.29  98.75   Level 4  61.99  99.22  (d) Corel-10K dataset   Level 1  6.47  18.91   Level 2  21.97  50.60   Level 3  37.70  80.73   Level 4  41.67  84.80  (e) GHIM-10K dataset   Level 1  10.79  25.65   Level 2  62.02  99.69   Level 3  85.29  100   Level 4  97.20  100   Level 5  97.32  100    Recall (%)  Precision (%)  (a) Corel-1K dataset   Level 1  20.73  37.09   Level 2  86.31  100   Level 3  98.02  100   Level 4  99.02  100   Level 5  99.07  100  (b) Olivia-2688 dataset   Level 1  25.86  45.49   Level 2  91.02  100   Level 3  91.32  100   Level 4  99.96  100   Level 5  99.96  100  (c) Corel-5K dataset   Level 1  9.22  22.86   Level 2  36.41  78.19   Level 3  58.29  98.75   Level 4  61.99  99.22  (d) Corel-10K dataset   Level 1  6.47  18.91   Level 2  21.97  50.60   Level 3  37.70  80.73   Level 4  41.67  84.80  (e) GHIM-10K dataset   Level 1  10.79  25.65   Level 2  62.02  99.69   Level 3  85.29  100   Level 4  97.20  100   Level 5  97.32  100  Figure 4. View largeDownload slide Average precision vs. level of resolution for the proposed method on (a) Corel-1K Olivia-2688 and GHIM-10K. (b) Corel-5K and Corel-10K. Figure 4. View largeDownload slide Average precision vs. level of resolution for the proposed method on (a) Corel-1K Olivia-2688 and GHIM-10K. (b) Corel-5K and Corel-10K. Figure 5. View largeDownload slide Average recall vs. level of resolution for the proposed method on (a) Corel-1K. (b) Olivia-2688 and GHIM-10K. (b) Corel-5K and Corel-10K. Figure 5. View largeDownload slide Average recall vs. level of resolution for the proposed method on (a) Corel-1K. (b) Olivia-2688 and GHIM-10K. (b) Corel-5K and Corel-10K. Figure 6. View largeDownload slide Precision vs. recall plot for the proposed method on (a) Corel-1K. (b) Olivia-2688. (c) Corel-5K. (d) Corel-10K. (e) GHIM-10K datasets. Figure 6. View largeDownload slide Precision vs. recall plot for the proposed method on (a) Corel-1K. (b) Olivia-2688. (c) Corel-5K. (d) Corel-10K. (e) GHIM-10K datasets. From the above experimental observations on different datasets, it is clearly observed that the average values of precision and recall increase with the level of resolution. Due to multiresolution analysis, each level covers details which were undetected at previous levels. This phenomenon leads to an increase in the values of precision and recall at different levels of resolution. Curvelets represent curvilinear structures and edges more effectively than wavelets due to their geometrical characteristics. Hence, curvelets construct more effective feature vector as compared to wavelets because they are able to extract more details at edges present in the natural image at multiple scales and directions. These factors result in high retrieval accuracy which can be observed in experimental results. 5.3. Performance comparison among single-resolution technique, wavelet-based technique and the proposed method To show the effectiveness of use of curvelet in the proposed method, the comparison of the proposed descriptor with wavelet-based technique and single resolution has been performed. Processing the single resolution of an image is simple and considers single scale of an image for feature extraction. A number of state-of-the-art methods exploit single resolution of image to construct feature vector. Some of the methods that have exploited single resolution of image for image retrieval produce promising results. However, single resolution fails to gather varying levels of details in an image. Hence, when compared with a multiresolution processing technique, single-resolution processing fails to yield better results. This phenomenon can be observed in Tables 3 and 4 where other multiresolution techniques produce much better results and outperform single-resolution processing technique denoted by SR. Table 3. Performance comparison of the proposed descriptor (LBCCP) with wavelet-based descriptor (LBWCP).   Corel-1K  Olivia-2688  GHIM-10K    LBWCP  LBCCP  LBWCP  LBCCP  LBWCP  LBCCP  (a) In terms of recall (%)   Level 1  44.16  20.73  36.98  25.86  23.75  10.79   Level 2  59.18  86.31  57.36  91.02  35.60  62.02   Level 3  69.24  98.02  77.50  91.32  44.28  85.29   Level 4  77.13  99.02  79.38  99.96  51.94  97.20   Level 5  83.18  99.07  86.15  99.96  58.72  97.32  (b) In terms of precision (%)   Level 1  79.68  37.09  79.32  45.49  64.02  22.65   Level 2  91.43  100  93.48  100  76.66  99.69   Level 3  96.82  100  97.73  100  84.41  100   Level 4  98.65  100  99.23  100  89.79  100   Level 5  99.40  100  99.83  100  93.49  100    Corel-1K  Olivia-2688  GHIM-10K    LBWCP  LBCCP  LBWCP  LBCCP  LBWCP  LBCCP  (a) In terms of recall (%)   Level 1  44.16  20.73  36.98  25.86  23.75  10.79   Level 2  59.18  86.31  57.36  91.02  35.60  62.02   Level 3  69.24  98.02  77.50  91.32  44.28  85.29   Level 4  77.13  99.02  79.38  99.96  51.94  97.20   Level 5  83.18  99.07  86.15  99.96  58.72  97.32  (b) In terms of precision (%)   Level 1  79.68  37.09  79.32  45.49  64.02  22.65   Level 2  91.43  100  93.48  100  76.66  99.69   Level 3  96.82  100  97.73  100  84.41  100   Level 4  98.65  100  99.23  100  89.79  100   Level 5  99.40  100  99.83  100  93.49  100    Corel-5K  Corel-10K    LBWCP  LBCCP  LBWCP  LBCCP  (c) In terms of recall (%)   Level 1  22.75  9.22  16.32  6.47   Level 2  29.28  36.41  20.42  21.97   Level 3  34.23  58.29  23.53  37.70   Level 4  38.44  61.99  26.15  41.67  (d) In terms of precision (%)   Level 1  46.26  22.86  36.45  18.91   Level 2  54.66  78.19  42.04  50.60   Level 3  60.71  98.75  45.96  80.73   Level 4  65.67  99.22  49.12  84.80    Corel-5K  Corel-10K    LBWCP  LBCCP  LBWCP  LBCCP  (c) In terms of recall (%)   Level 1  22.75  9.22  16.32  6.47   Level 2  29.28  36.41  20.42  21.97   Level 3  34.23  58.29  23.53  37.70   Level 4  38.44  61.99  26.15  41.67  (d) In terms of precision (%)   Level 1  46.26  22.86  36.45  18.91   Level 2  54.66  78.19  42.04  50.60   Level 3  60.71  98.75  45.96  80.73   Level 4  65.67  99.22  49.12  84.80  Table 4. Performance analysis of the proposed method LBCCP with other techniques. Dataset  Performance analysis of the proposed method    SR  LBWCP  Corel-1K  Precision  Proposed method outperforms SR by 45.69%  Proposed method outperforms LBWCP by 0.60%  Recall  Proposed method outperforms SR by 68.40%  Proposed method outperforms LBWCP by 15.89%  Olivia-2688  Precision  Proposed method outperforms SR by 44.58%  Proposed method outperforms LBWCP by 0.17%  Recall  Proposed method outperforms SR by 73.77%  Proposed method outperforms LBWCP by 13.81%  Corel-5K  Precision  Proposed method outperforms SR by 64.18%  Proposed method outperforms LBWCP by 33.55%  Recall  Proposed method outperforms SR by 47.74%  Proposed method outperforms LBWCP by 23.55%  Corel-10K  Precision  Proposed method outperforms SR by 57.18%  Proposed method outperforms LBWCP by 35.68%  Recall  Proposed method outperforms SR by 31.60%  Proposed method outperforms LBWCP by 15.52%  GHIM-10K  Precision  Proposed method outperforms SR by 62.68%  Proposed method outperforms LBWCP by 6.51%  Recall  Proposed method outperforms by 71.13%  Proposed method outperforms by 38.60%  Dataset  Performance analysis of the proposed method    SR  LBWCP  Corel-1K  Precision  Proposed method outperforms SR by 45.69%  Proposed method outperforms LBWCP by 0.60%  Recall  Proposed method outperforms SR by 68.40%  Proposed method outperforms LBWCP by 15.89%  Olivia-2688  Precision  Proposed method outperforms SR by 44.58%  Proposed method outperforms LBWCP by 0.17%  Recall  Proposed method outperforms SR by 73.77%  Proposed method outperforms LBWCP by 13.81%  Corel-5K  Precision  Proposed method outperforms SR by 64.18%  Proposed method outperforms LBWCP by 33.55%  Recall  Proposed method outperforms SR by 47.74%  Proposed method outperforms LBWCP by 23.55%  Corel-10K  Precision  Proposed method outperforms SR by 57.18%  Proposed method outperforms LBWCP by 35.68%  Recall  Proposed method outperforms SR by 31.60%  Proposed method outperforms LBWCP by 15.52%  GHIM-10K  Precision  Proposed method outperforms SR by 62.68%  Proposed method outperforms LBWCP by 6.51%  Recall  Proposed method outperforms by 71.13%  Proposed method outperforms by 38.60%  The second method used for comparison is wavelet-based technique which employs discrete wavelet transform (DWT) (using Daubechies1 Haar wavelet) with LBP and GLCM denoted by local binary wavelet co-occurrence pattern (LBWCP). LBWCP has been compared with the proposed descriptor, LBCCP, in terms of precision and recall for five levels of resolution in case of 256 × 256 image and four levels of resolution in case of 128 × 128 image size. LBWCP produces promising results as wavelet gathers directional information at multiple resolution of image. However, wavelets fail to gather details at sharp edges and curvilinear structures due to their non-geometrical nature. Hence, wavelets fail to construct as efficient feature vector as curvelet. It is worth noticing that Level 1 decomposition of DWT produces better retrieval results as compared with Level 1 decomposition of curvelet as shown in Table 3. This is because Level 1 decomposition of DWT computes coefficients in three directions, horizontal, vertical and diagonal, and hence produces three separate matrices for feature vector whereas curvelet transform computes coefficients at single orientation in Level 1 decomposition. Therefore, it produces single matrix for feature vector. However, when compared at successive levels, curvelet transform computes coefficients at multiple orientations and hence is able to extract details at edges and curvilinear structures which wavelet transform fails to do. Hence, the overall retrieval result is far better in case of curvelet-based technique as shown in Table 3 (highlighted in bold). Hence, LBCCP performs better than LBWCP descriptor due to its anisotropic nature and ability to extract features at curvature and sharp edges using fewer coefficients. Table 5 shows the performance comparison of the proposed method (LBCCP) with other methods (single-resolution (SR) and wavelet-based technique (LBWCP)) in terms of recall on five datasets (Corel-1K, Olivia-2688, Corel-5K, Corel-10K, GHIM-10K). It can be clearly observed that the proposed method achieves much higher recall values than other techniques on five datasets which have been highlighted in bold. Similarly, Table 6 shows the performance comparison of the proposed method with other techniques in terms of precision. In case of precision also, the proposed method clearly outperforms other techniques. The precision values of the proposed method have been highlighted in bold. It can be observed that the proposed method descriptor LBCCP achieves 100% precision accuracy on certain datasets. For precision, we have considered top 10 matches. The final level of LBCCP obtains all top 10 matches belonging to the same category as that of the query image. Due to this reason, the proposed descriptor achieves 100% precision value in these cases. The performance analysis of the proposed method with other techniques on five datasets is shown in Table 4. Table 5. Comparison of the proposed method (PM) with other techniques in terms of recall (%). Datasets  Single resolution (SR)  LBWCP  LBCCP (PM)  Corel-1K  30.67  83.18  99.07  Olivia-2688  26.19  86.15  99.96  Corel-5K  14.25  38.44  61.99  Corel-10K  10.07  26.15  41.67  GHIM-10K  26.19  58.72  97.32  Datasets  Single resolution (SR)  LBWCP  LBCCP (PM)  Corel-1K  30.67  83.18  99.07  Olivia-2688  26.19  86.15  99.96  Corel-5K  14.25  38.44  61.99  Corel-10K  10.07  26.15  41.67  GHIM-10K  26.19  58.72  97.32  Table 6. Comparison of the proposed method (PM) with other techniques in terms of precision (%). Datasets  Single resolution (SR)  LBWCP  LBCCP (PM)  Corel-1K  54.31  99.40  100  Olivia-2688  55.42  99.83  100  Corel-5K  35.04  65.67  99.22  Corel-10K  27.62  49.12  84.80  GHIM-10K  37.32  93.49  100  Datasets  Single resolution (SR)  LBWCP  LBCCP (PM)  Corel-1K  54.31  99.40  100  Olivia-2688  55.42  99.83  100  Corel-5K  35.04  65.67  99.22  Corel-10K  27.62  49.12  84.80  GHIM-10K  37.32  93.49  100  5.4. Performance comparison of the proposed method with other state-of-the-art methods The performance of the proposed method is compared with other state-of-the-art methods (Srivastava et al. [7], Srivastava et al. [12], Srivastava et al. [23], MSD [14–16]). Table 9. Performance analysis of the proposed method with other state-of-the-art CBIR methods. Dataset  Performance analysis of the proposed method      Srivastava et al. [7]  Srivastava et al. [12]  Srivastava et al. [23]  MSD [14–16]  Corel-1K  Precision  Proposed method outperforms by 64.06%  Proposed method outperforms by 46.30%  Proposed method outperforms by 32.84%  Proposed method outperforms by 24.33%  Recall  Proposed method outperforms by 46.28%  Proposed method outperforms by 27.87%  Proposed method outperforms by 48.98%  Proposed method outperforms by 90.00%  Olivia-2688  Precision  Proposed method outperforms by 86.62%  Proposed method outperforms by 38.87%  Proposed method outperforms by 10.40%  Proposed method outperforms by 52.16%  Recall  Proposed method outperforms by 87.25%  Proposed method outperforms by 63.51%  Proposed method outperforms by 40.85%  Proposed method outperforms by 98.20%  Corel-5K  Precision  Proposed method outperforms by 82.83%  Proposed method outperforms by 67.04%  Proposed method outperforms by 67.27%  Proposed method outperforms by 43.30%  Recall  Proposed method outperforms by 54.67%  Proposed method outperforms by 46.63%  Proposed method outperforms by 42.11%  Proposed method outperforms by 55.28%  Corel-10K  Precision  Proposed method outperforms by 71.65%  Proposed method outperforms by 59.88%  Proposed method outperforms by 62.45%  Proposed method outperforms by 39.18%  Recall  Proposed method outperforms by 37.54%  Proposed method outperforms by 30.73%  Proposed method outperforms by 30.57%  Proposed method outperforms by 36.19%  GHIM-10K  Precision  Proposed method outperforms by 80.66%  Proposed method outperforms by 62.77%  Proposed method outperforms by 50.82%  Proposed method outperforms by 47.98%  Recall  Proposed method outperforms by 88.83%  Proposed method outperforms by 79.68%  Proposed method outperforms by 66.32%  Proposed method outperforms by 94.88%  Dataset  Performance analysis of the proposed method      Srivastava et al. [7]  Srivastava et al. [12]  Srivastava et al. [23]  MSD [14–16]  Corel-1K  Precision  Proposed method outperforms by 64.06%  Proposed method outperforms by 46.30%  Proposed method outperforms by 32.84%  Proposed method outperforms by 24.33%  Recall  Proposed method outperforms by 46.28%  Proposed method outperforms by 27.87%  Proposed method outperforms by 48.98%  Proposed method outperforms by 90.00%  Olivia-2688  Precision  Proposed method outperforms by 86.62%  Proposed method outperforms by 38.87%  Proposed method outperforms by 10.40%  Proposed method outperforms by 52.16%  Recall  Proposed method outperforms by 87.25%  Proposed method outperforms by 63.51%  Proposed method outperforms by 40.85%  Proposed method outperforms by 98.20%  Corel-5K  Precision  Proposed method outperforms by 82.83%  Proposed method outperforms by 67.04%  Proposed method outperforms by 67.27%  Proposed method outperforms by 43.30%  Recall  Proposed method outperforms by 54.67%  Proposed method outperforms by 46.63%  Proposed method outperforms by 42.11%  Proposed method outperforms by 55.28%  Corel-10K  Precision  Proposed method outperforms by 71.65%  Proposed method outperforms by 59.88%  Proposed method outperforms by 62.45%  Proposed method outperforms by 39.18%  Recall  Proposed method outperforms by 37.54%  Proposed method outperforms by 30.73%  Proposed method outperforms by 30.57%  Proposed method outperforms by 36.19%  GHIM-10K  Precision  Proposed method outperforms by 80.66%  Proposed method outperforms by 62.77%  Proposed method outperforms by 50.82%  Proposed method outperforms by 47.98%  Recall  Proposed method outperforms by 88.83%  Proposed method outperforms by 79.68%  Proposed method outperforms by 66.32%  Proposed method outperforms by 94.88%  Srivastava et al. [7] exploit shape feature for retrieval through moments. It uses moments as a single feature, constructs feature vector by dividing image into blocks of different sizes and computing geometric moments of each block. The method is simple and efficiently computes shape feature for retrieval. However, as discussed in the previous section, single feature is insufficient for retrieving similar images. Also, the method exploits single resolution of an image and fails to capture varying levels of details in an image. Hence, this technique produces low retrieval accuracy when compared with the proposed method as shown in Tables 7 and 8 and Figs. 7 and 8. Table 7. Performance comparison of the proposed method (PM) with other state-of-the-art CBIR methods in terms of precision (%). Datasets  Srivastava et al. [7]  Srivastava et al. [12]  Srivastava et al. [23]  MSD [14–16]  LBCCP (PM)  Corel-1K  35.94  53.70  67.16  75.67  100  Olivia-2688  13.38  61.13  89.60  47.84  100  Corel-5K  16.39  32.18  31.95  55.92  99.22  Corel-10K  13.15  24.92  22.35  45.62  84.80  GHIM-10K  19.34  37.23  49.18  52.02  100  Datasets  Srivastava et al. [7]  Srivastava et al. [12]  Srivastava et al. [23]  MSD [14–16]  LBCCP (PM)  Corel-1K  35.94  53.70  67.16  75.67  100  Olivia-2688  13.38  61.13  89.60  47.84  100  Corel-5K  16.39  32.18  31.95  55.92  99.22  Corel-10K  13.15  24.92  22.35  45.62  84.80  GHIM-10K  19.34  37.23  49.18  52.02  100  Table 8. Performance comparison of the proposed method (PM) with other state-of-the-art CBIR methods in terms of recall (%). Datasets  Srivastava et al. [7]  Srivastava et al. [12]  Srivastava et al. [23]  MSD [14–16]  LBCCP (PM)  Corel-1K  52.79  72.09  50.09  9.07  99.07  Olivia-2688  12.71  36.45  59.11  1.76  99.96  Corel-5K  7.32  15.36  19.88  6.71  61.99  Corel-10K  4.13  10.94  11.10  5.48  41.67  GHIM-10K  8.49  17.64  31.00  2.44  97.32  Datasets  Srivastava et al. [7]  Srivastava et al. [12]  Srivastava et al. [23]  MSD [14–16]  LBCCP (PM)  Corel-1K  52.79  72.09  50.09  9.07  99.07  Olivia-2688  12.71  36.45  59.11  1.76  99.96  Corel-5K  7.32  15.36  19.88  6.71  61.99  Corel-10K  4.13  10.94  11.10  5.48  41.67  GHIM-10K  8.49  17.64  31.00  2.44  97.32  Figure 7. View largeDownload slide Performance comparison of the proposed method with other state-of-the-art CBIR methods on five datasets in terms of precision. Figure 7. View largeDownload slide Performance comparison of the proposed method with other state-of-the-art CBIR methods on five datasets in terms of precision. Figure 8. View largeDownload slide Performance comparison of the proposed method with other state-of-the-art CBIR methods on five datasets in terms of recall. Figure 8. View largeDownload slide Performance comparison of the proposed method with other state-of-the-art CBIR methods on five datasets in terms of recall. The second method used for comparison is Srivastava et al. [12] which uses combination of features for retrieval. This method combines local ternary pattern with moments to retrieve visually similar images. The method uses combination of local and global features, thereby, overcoming limitations of single feature. The method extracts more details from an image as compared with single feature. This method exploits combination of features at single resolution and produces good retrieval accuracy on small datasets. However, this method fails to produce high accuracy for large datasets as such datasets consist of a wide variety of images containing varying levels of details. Being exploited at single resolution of image, this method proves to be less efficient when compared with the proposed method which produces high retrieval accuracy for small as well as large datasets. Hence, the proposed method outperforms Srivastava et al. [12] as illustrated in Tables 7 and 8 and Figs. 7 and 8. The third method used for comparison with the proposed method is Srivastava et al. [23]. The method exploits shape feature through moments at multiple resolutions of image using discrete wavelet transform. The method uses multiple resolutions of image to extract shape feature in order to construct feature vector for retrieval. However, as discussed in the previous section, wavelets fail to capture curves and edges efficiently and hence, do not construct as efficient feature vector as constructed by the proposed method. The proposed method employs curvelet transform for constructing feature vector which efficiently captures curvilinear structures and sharp edges due to its anisotropic nature. Hence, the proposed method outperforms Srivastava et al. [23] also as shown in Tables 7 and 8 and Figs. 7 and 8. The fourth method used for comparison with the proposed method is MSD [14]. It is an efficient technique of retrieval which employs edge orientation similarity along with underlying colours in order to construct feature vector. The technique attempts to exploit similar edge orientation for feature vector construction. Also, single MSD feature efficiently extracts colour, texture and shape features simultaneously due to which it proves to be a powerful feature for retrieval. Although MSD feature produces promising results, it too exploits single resolution of image for feature vector construction and fails to consider varying levels of details. Hence, when compared with the proposed technique, MSD [14] technique produces low retrieval results. This fact can be observed in Tables 7 and 8 and Figs. 7 and 8. The methods (Srivastava et al. [7], Srivastava et al. [12] and Srivastava et al. [23]) have been implemented in MATLAB R2013a on a PC having Windows 8.1 Pro operating system, Intel Core 15-4570 processor at 3.20 GHz and 4 GB of RAM. The precision and recall values of MSD [14] have been referred to from [14–16]. The proposed method employing curvelet transform for CBIR outperforms the above-discussed state-of-the-art methods (Srivastava et al. [7], Srivastava et al. [12], Srivastava et al. [23] and MSD [14–16]) due to its anisotropic nature, ability to extract features at curvatures and sharp edges and exploitation of multiple resolutions of image for feature vector construction. These properties of the proposed method help in capturing varying levels of details in an image and produce high retrieval accuracy. Tables 7 and 8 and Figs. 7 and 8 show the performance comparison of the proposed method with other state-of-the-art CBIR methods in terms of precision and recall on five datasets (Corel-1K, Olivia-2688, Corel-5K, Corel-10K and GHIM-10K) used in this paper. Table 8 shows the performance comparison of the proposed method with other state-of-the-art methods in terms of precision. Similarly, Table 8 shows the performance comparison of the proposed method with other state-of-the-art image retrieval methods. It can be clearly observed that the proposed method outperforms other techniques both in terms of precision and recall by a huge margin. The performance analysis of the proposed method with other state-of-the-art CBIR techniques on five datasets is shown in the form of a table in Table 9. 6. CONCLUSION In this paper, we presented a new descriptor, namely, LBCCP for CBIR. The proposed method combines LBP with curvelet transform followed by construction of GLCM for feature vector generation. The feature vector value was compared with those of images in database to retrieve visually similar images. In a nutshell, the proposed method has following advantages: The proposed method exploits multiresolution LBP and gathers local information at multiple scales at curves and edges. Feature vector construction through GLCM is advantageous as it considers spatial arrangement of pixels which the histogram fails to do. The proposed method yields high retrieval accuracy. Performance of the proposed method has been measured in terms of precision and recall. The proposed method, when compared with single-resolution technique and wavelet-based technique as well as other state-of-the-art techniques, outperforms them as demonstrated through experimental results. The proposed combination of LBP and curvelet transform can be useful for other computer vision applications such as face recognition, object recognition and biometric image processing due to useful properties of LBP and curvelet transform. The proposed method can further be improved by combining more features such as shape with local features at multiresolution level to extract more details from an image. REFERENCES 1 Long, F., Zhang, H. and Feng, D.D. ( 2003) Fundamentals of Content-Based Image Retrieval. Multimedia Information Retrieval and Management . Springer, Berlin. 2 Smith, J.R. and Chang, S.F. ( 1996) Tools and Techniques for Color Image Retrieval. Electronic Imaging: Science and Technology , pp. 426– 437. International Society for Optics and Photonics. 3 Huang, J., Kumar, S.R., Mitra, M. and Zhu, W. ( 2001) Image Indexing using Color Correlograms, U.S. Patent 6,246,790. 4 Pass, G., Zabih, R. and Miller, J. ( 1997) Comparing images using color coherence vectors. In Proc. 4th ACM Int. Conf. Multimedia, Boston, MA, USA, November 18–22, pp. 65–73. ACM, New York, USA. 5 Manjunath, B.S. and Ma, W.X. ( 1996) Browsing and retrieval of image data. IEEE. Trans. Pattern. Anal. Mach. Intell. , 18, 837– 842. Google Scholar CrossRef Search ADS   6 Verma, M., Balasubrahmanian, R. and Murala, S. ( 2015) Local extrema co-occurrence pattern using color and texture image retrieval. Neurocomputing. , 165, 255– 269. Google Scholar CrossRef Search ADS   7 Srivastava, P., Binh, N.T. and Khare, A. ( 2013) Content-based image retrieval using moment. In Proc. 2nd Int. Conf. Context-Aware Systems and Applications, Phu Quoc, Vietnam, November 25–26, pp. 221–231. Lecture Notes of ICST, Springer. 8 Andreou, I. and Sgouros, N.M. ( 2005) Computing, explaining visualizing shape similarity in content-based image retrieval. Inf. Process. Manage. , 41, 1121– 1139. Google Scholar CrossRef Search ADS   9 Khare, M., Srivastava, R.K. and Khare, A. ( 2015) Moving object segmentation in daubechies complex wavelet domain. Signal, Image Video Process. , 9, 535– 650. Google Scholar CrossRef Search ADS   10 Lin, C. and Lin, W. ( 2011) Image retrieval based on adaptive color histogram and texture features. Comput. J. , 54, 1136– 1147. Google Scholar CrossRef Search ADS   11 Gevers, T. and Smeulders, A.W. ( 2000) Pictoseek: combining color and shape invariant features for image retrieval. IEEE. Trans. Image. Process. , 9, 102– 119. Google Scholar CrossRef Search ADS PubMed  12 Srivastava, P., Binh, N.T. and Khare, A. ( 2014) Content-based image retrieval using moments of local ternary pattern. Mobile Netw. Appl. , 19, 618– 625. Google Scholar CrossRef Search ADS   13 Wang, X., Yu, Y. and Yang, H. ( 2011) An effective image retrieval scheme using color, texture and shape features. Comput. Stand. Interfaces , 33, 59– 68. Google Scholar CrossRef Search ADS   14 Liu, G., Li, Z., Zhang, L. and Xu, Y. ( 2011) Image retrieval based on microstructure descriptor. Pattern. Recognit. , 44, 2123– 2133. Google Scholar CrossRef Search ADS   15 Zhang, M., Zhang, K., Feng, Q., Wang, J., Kong, J. and Lu, Y. ( 2014) A novel image retrieval method based on hybrid information descriptors. J. Vis. Commun. Image Represent. , 25, 1574– 1587. Google Scholar CrossRef Search ADS   16 Liu, G.H., Yang, J.Y. and Li, Z. ( 2015) Content-based image retrieval using computational attention model. Pattern. Recognit. , 48, 2554– 2566. Google Scholar CrossRef Search ADS   17 Flores-Pulido, L., Starostenko, O., Flores-Quéchol, D., Rodrigues-Flores, J.I., Kirschning, I. and Chávez-Aragón, J.A. ( 2008) Content-based Image Retrieval Using Wavelets. In Proc. 2nd WSEAS Int. Conf. Computer Engineering and Applications, pp. 40–45. 18 Moghaddam, H., Abrishami, T., Khajoie, T., Rouhi, A.H. and Tarzjan, M.S. ( 2005) Wavelet correlogram: A new approach for image indexing and retrieval. Pattern. Recognit. , 38, 2506– 2518. Google Scholar CrossRef Search ADS   19 Saadatmand-Tarzjan, M. and Moghaddam, H.A. ( 2007) A novel evolutionary approach for optimizing content-based image indexing algorithms. IEEE Trans. Syst. Man Cybern. B Cybern. , 37, 139– 153. Google Scholar CrossRef Search ADS PubMed  20 Lamard, M., Cazuguel, G., Quellec, G., Bekri, L., Roux, C. and Cochener, B. ( 2007) Content Based Image Retrieval Based on Wavelet Transform Coefficients Distribution. In Proc. 29th IEEE Annual Int. Conf. Engineering in Medicine and Biology Society, August 22–26, pp. 4532–4535. IEEE. 21 Agarwal, M. and Maheshwari, R.P. ( 2012) Á trous gradient structure descriptor for content-based image retrieval. Int. J. Multimed. Inf. Retr. , 1, 129– 138. Google Scholar CrossRef Search ADS   22 Fu, X., Li, Y., Harrison, R. and Belkasim, S. ( 2006) Content-Based Image Retrieval using Gabor-Zernike Features. In Proc. 18th Int. Conf. Pattern Recognition, Hong Kong, August 20–24, 2, pp. 417–420, IEEE. 23 Srivastava, P., Prakash, O. and Khare, A. ( 2014) Content-Based Image Retrieval using Moments of Wavelet Transform. In Proc. Int. Conf. Control Automation and Information Sciences, Gwangju, South Korea, December 02–05, pp. 159–164. 24 Vo, A. and Oraintara, S. ( 2010) A study of relative phase in complex wavelet domain: Property, statistics and applications in texture image retrieval and segmentation. Signal Process. , 25, 28– 46. 25 Starck, J., Candes, E.J. and Donoho, D.L. ( 2002) The curvelet transform for image denoising. IEEE. Trans. Image. Process. , 11( 6), 670– 684. Google Scholar CrossRef Search ADS PubMed  26 Sumana, I.J., Islam, M.M., Zhang, D. and Liu, G. ( 2008) Content-Based Image Retrieval using Curvelet Transform. In Proc. 10th IEEE Workshop on Multimedia Signal Processing, Cairns, Queensland, Australia, October 8–10, pp. 11–16. IEEE. 27 Gonde, A.B., Maheshwari, R.P. and Balasubramanian, R. ( 2013) Modified curvelet transform with vocabulary tree for content based image retrieval. Digit. Signal Process. , 23, 142– 150. Google Scholar CrossRef Search ADS   28 Youssef, S.F. ( 2012) ICTEDCT-CBIR: Integrating curvelet transform with enhanced dominant colors extraction and texture analysis for efficient content-based image retrieval. Comput. Electr. Eng. , 38, 1358– 1376. Google Scholar CrossRef Search ADS   29 Zhang, D., Islam, M.M., Lu, G. and Sumana, I.J. ( 2012) Rotation invariant curvelet features for region based image retrieval. Int. J. Comput. Vis. , 98, 187– 201. Google Scholar CrossRef Search ADS   30 Murtagh, F. and Starck, J.L. ( 2008) Wavelet and curvelet moments for image classification, Application to aggregate mixture grading. Pattern Recognit. Lett. , 29, 1557– 1564. Google Scholar CrossRef Search ADS   31 Sumana, I.J., Lu, G. and Zhang, D. ( 2012) Comparison of Curvelet and Wavelet Texture Features for Content Based Image Retrieval. In Proc. IEEE Int. Conf. Multimedia and Expo (ICME), Melbourne, Australia, July 9–13, pp. 290–295. IEEE. 32 Das, S., Garg, S. and Sahoo, G. ( 2012) Comparison of content based image retrieval systems using wavelet and curvelet transform. Int. J. Multimed. Appl. , 4, 137. 33 Luo, J., Song, D., Xiu, C., Geng, S. and Dong, T. ( 2014) Fingerprint classification combining curvelet transform and gray-level co-occurrence matrix. Math. Probl. Eng. , 2014, 15. 34 Gonzalez, R.C. and Woods, R.E. ( 2002) Digital Image Processing  ( 2nd edn). Prentice Hall Press. 35 Ojala, T., Pietikainen, M. and Maenpaa, T. ( 2002) Multiresolution gray-scale and rotation invariant texture classification with local binary patterns. IEEE. Trans. Pattern. Anal. Mach. Intell. , 24, 971– 987. Google Scholar CrossRef Search ADS   36 Nigam, S. and Khare, A. ( 2015) Multi-resolution approach for multiple human detection using moments and local binary patterns. Multimed. Tools Appl. , 74, 7037– 7062. Google Scholar CrossRef Search ADS   37 Haralick, R.M., Shanmungam, K. and Dinstein, I. ( 1973) Textural features of image classification. IEEE Trans. Syst., Man Cybern. , 3, 610– 621. Google Scholar CrossRef Search ADS   38 Candes, E., Demanet, L., Donoho, D. and Ying, L. ( 2006) Fast discrete curvelet transforms. Multiscale Model. Simul. , 5, 861– 899. Google Scholar CrossRef Search ADS   39 http://wang.ist.psu.edu/docs/related/. Accessed  April 2014 40 Oliva, A. and Torraiba, A. ( 2001) Modeling the shape of the scene; a holistic representation of the spatial envelope. Int. J. Comput. Vis. , 42, 145– 175. Google Scholar CrossRef Search ADS   41 http://www.ci.gxnu.edu.cn/cbir/. Accessed June 2015. 42 Buades, A. and Lisani, J.L. ( 2016) Directional filters for cartoon + texture image decomposition. Image Process. Line , 6, 75– 88. Google Scholar CrossRef Search ADS   43 Starck, J.L., Murtagh, F. and Fadili, J.M. ( 2010) Sparse Image and Signal Processing: Wavelets, Curvelets, Morphological Diversity . Cambridge University Press. Google Scholar CrossRef Search ADS   Author notes Handling editor: Fionn Murtagh © The British Computer Society 2017. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png The Computer Journal Oxford University Press

Content-Based Image Retrieval using Local Binary Curvelet Co-occurrence Pattern—A Multiresolution Technique

Loading next page...
 
/lp/ou_press/content-based-image-retrieval-using-local-binary-curvelet-co-aN7WLlc2aK
Publisher
Oxford University Press
Copyright
© The British Computer Society 2017. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com
ISSN
0010-4620
eISSN
1460-2067
D.O.I.
10.1093/comjnl/bxx086
Publisher site
See Article on Publisher Site

Abstract

Abstract With the growth of various image-capturing devices, image acquisition is no longer a difficult task. As this technology is flourishing, various types of complex images are being produced. In order to access a large number of images stored in database easily, the images must be properly organized. Field of image retrieval attempts to solve this problem. As the complex images are being produced, processing them using single-resolution techniques is not sufficient as these images may contain varying levels of details. This paper proposes a novel multiresolution descriptor, local binary curvelet co-occurrence pattern, to achieve the task of content-based image retrieval. Curvelet transform of grayscale image is computed followed by computation of local binary pattern of resulting curvelet coefficients. Finally, feature vector is constructed using grey-level co-occurrence matrix which is matched with the feature vector of database images. The proposed descriptor combines the properties of local pattern and multiresolution technique of curvelet transform, and efficiently covers curvilinear and geometrical structures present in the image. Performance of the proposed method is measured in terms of precision and recall and is tested on five benchmark datasets consisting of natural images. The proposed method has been compared with single and multiresolution techniques as well as with some of the other state-of-the-art image retrieval methods. The experimental results clearly demonstrate that the proposed method produces high retrieval accuracy and outperforms other techniques in terms of precision and recall. 1. INTRODUCTION Development of a large number of image-capturing devices has led to the proliferation of huge amount of images. With the development of low-cost hand held devices, capturing of image is much easier than it was a few years ago. This has created a huge repository of unorganized images. Organizing a large number of images is quite challenging and scientists across the world are working on various techniques to accomplish this task. All these issues have made image indexing and retrieval an important problem of computer vision. Primarily, there are two ways of achieving this task. The first approach is annotation of images on the basis of text and keywords. This approach requires annotation of a large number of images to make the task of indexing and retrieval easy. Although this method of retrieval is easy, it requires manual annotation of a large number of images which is a difficult task. Also, this method fails to retrieve visually similar images. The second approach is retrieval of images on the basis of contents of image and it requires no manual tagging and retrieves visually similar images. Content-based image retrieval (CBIR), also named as content-based visual information retrieval (CBVIR), refers to retrieval of images on the basis of features present in the image. In CBIR systems, an image or its sketch is supplied as a query to retrieve images visually similar to the query image. The system extracts features from the image and constructs a feature vector. This feature vector is matched with those of database images to retrieve visually similar images [1]. The construction of efficient feature vector is a challenging problem in image retrieval as the effectiveness of feature vector determines the success of retrieval. Image features can be broadly classified into two categories, primary features such as colour, texture and shape, and semantic features such as type of image. Processing of primary features has been extensively used for image retrieval. CBIR using colour has mostly been used in the form of colour histogram as it is easy to construct [2]. Apart from colour histogram, colour has also been used as a feature in colour correlogram [3] and colour coherence vector [4]. Thus, colour has been a popular feature for retrieval as it is a visible descriptor and invariant to certain geometrical transformations. Texture is another feature which has been used for image indexing [5]. Texture determines properties such as coarseness, roughness and smoothness of a surface. Texture has been exploited through Fourier transform [5] as well as grey-level co-occurrence matrix (GLCM) [6]. Apart from colour and texture, shape is another feature which has been exploited in the form of moments [7] and polygonal shape [8]. Shape features are generally used after segmentation [9]. Lack of good segmentation algorithm makes shape feature less popular than colour and texture. Early image retrieval systems used single feature for feature vector construction and this technique remained popular for a long time. However, as complexity in nature of images started increasing, single feature started proving to be insufficient. Hence, the trend of CBIR shifted to combination of features for feature vector construction. The combination of colour and texture [10], colour and shape [11], texture and shape [12], and colour, texture and shape [13] are some of the approaches that have incorporated multiple features for feature vector construction. Multiple features tend to combine the advantages of more than one feature and are able to extract more features as compared with single feature. Modern methods of CBIR employ human perception analysis and understanding to retrieve visually similar images. Liu et al. [14] described microstructure descriptor (MSD) to extract colour, texture and shape through single microstructure feature. Zhang et al. [15] proposed hybrid information descriptor that combines high-level and low-level features for image retrieval. Liu et al. [16] proposed visual attention model to perform human perception analysis for CBIR. All the methods discussed above have been applied on single resolution of image. An image is generally a complex structure and consists of varying levels of details which the single-resolution processing fails to capture. This fact motivated the researchers to perform CBIR in multiresolution analysis framework. The advantage of performing multiresolution analysis is that it considers varying levels of details for feature vector construction and features that are left undetected at one resolution get considered at another level. A number of multiresolution techniques have been used where multiple levels of image have been exploited to construct feature vector for retrieval. This paper proposes a new multiresolution descriptor, known as local binary curvelet co-occurrence pattern (LBCCP), for image retrieval. The proposed method in this paper exploits local binary pattern (LBP) at multiple scales through curvelet transform followed by construction of GLCM for generation of feature vector. Curvelet, being a multiresolution technique, captures curvilinear structure and sharp edges which wavelet transform fails to capture. The reason for this is that curvelets are highly anisotropic in nature and use fewer coefficients to represent edges unlike wavelets. The feature vector is constructed through GLCM as it provides spatial information by considering occurrence of pixel pairs which other techniques fail to do. In this way, the proposed descriptor exploits spatial information about local feature at multiple resolutions to construct efficient feature vector. The proposed descriptor captures local information at different angles at multiple scales of image which helps in gathering complex details from natural images. This combination of local descriptor along with multiresolution technique such as curvelet helps in detecting sharp curvilinear structures which are not detected through single-resolution processing techniques. For multiresolution processing of image, curvelet transform has been used for the following reasons: first, due to its anisotropic nature, curvelets are able to capture sharp curvilinear structures present in natural images which are not captured by wavelets due to their isotropic nature. Second, natural images are complex in nature and contain varying levels of details as compared with medical and forensic images and consist of complex foreground as well as background details. Such complex details are captured by curvelet transform more effectively than wavelet transform. The construction of feature has been done through GLCM which attempts to capture spatial distribution of intensity values. This helps in providing information about structural arrangement of pixels which are not provided by other features such as histogram. The combination of local pattern with multiresolution technique such as curvelets attempts to capture complex details such as sharp curvilinear structures which single-resolution processing techniques and certain other multiresolution techniques fail to gather. Rest of the paper is organized as follows. Section 2 discusses some of the related work in the field of image retrieval. Section 3 gives brief background of curvelet transform, LBP and GLCM. Section 4 discusses the proposed method. Section 5 discusses experiment and results and Section 6 concludes the paper. 2. RELATED WORK Since the inception of the term image retrieval, several algorithms have been proposed to improve the accuracy of retrieval. Early methods of retrieval were focused mainly on primary features, namely, colour, texture and shape. These features were used either as a single feature or in combination with each other. Most of these methods exploited single resolution of image. A natural image always contains varying levels of details and in order to exploit those varying levels of details, single-resolution processing is not sufficient. This limitation has been overcome by multiresolution processing of image which considers more than one resolution to capture varying levels of details. Several techniques exploiting multiple resolutions of image have been proposed. One of them is wavelet which has been in use for a long time [17]. Wavelet has been combined with colour in the form of wavelet correlogram [18]. This concept has been further improved by adding optimization feature in [19] which not only improves retrieval speed but also retrieval accuracy. Lemard et al. [20] proposed image retrieval based on wavelets by using signatures constructed from wavelet coefficients. Apart from using as a single feature, wavelet has been combined with other features also. Agarwal et al. [21] proposed combination of wavelet and MSD. This method computed MSD of à trous wavelet coefficients to perform retrieval. Fu et al. [22] combined wavelet with shape feature, Zernike moments, for image indexing. The combination of wavelet and shape feature has also been discussed in [23]. Vo et al. [24] exploited the concept of relative phase in complex wavelet domain for texture image retrieval and segmentation. Thus, wavelet transform is a popular multiresolution technique and has been exploited much more than any other multiresolution technique for image retrieval. However, wavelets suffer from certain drawbacks. Due to their non-geometrical nature, wavelets do not effectively represent objects that consist of very sharp edges and curves. Due to this limitation, wavelets fails to effectively exploit regularity of edge curve. Also, due to their non-anisotropic nature, wavelets do not capture details of any region at multiple orientations but consider only limited orientations at all scales. Due to the emergence of various image-capturing devices, more and more complex natural images, consisting of variety of objects, are being produced everyday. Because of the above-mentioned limitations, wavelets fail to gather curvilinear structures and features at edges efficiently. These drawbacks can be overcome by using another multiresolution technique called curvelets [25]. Curvelets are collections of multiscale ridgelets at multiple orientations and scales [25]. Unlike wavelets, curvelets are geometrical in nature and highly anisotropic. These properties enable curvelets to compute coefficients of sharp edges and curves which wavelets fail to achieve. Sumana et al. [26] used curvelet coefficients for image retrieval. This technique applied curvelet transform on texture images and computed low-order statistics to construct feature vector. The method is simple and extracts low-order statistics efficiently, but fails to exploit multiresolution aspect of curvelet transform. Gonde et al. [27] modified basic structure of curvelet transform by introducing Gabor wavelet instead of à trous wavelet and constructed feature vector through vocabulary tree for retrieval. This technique performed better than both original curvelet transform-based technique and wavelet-based technique. However, this method produced low retrieval accuracy. Youssef [28] proposed curvelet-based image retrieval scheme named ICTEDT-CBIR. This method combines colour and texture features along with curvelet transform for image indexing and retrieval. Zhang et al. [29] proposed rotation invariant curvelet feature to exploit texture feature of image. This method applies curvelet transform on region-based colour image to construct feature vector which is rotation invariant. Murtagh and Starck [30] proposed the concept of curvelet moments similar to wavelet moments. Sumana et al. [31] proposed new feature based on texture named generalized Gaussian density texture feature for retrieval. Das et al. [32] compared wavelet and curvelet transform for image retrieval by combining colour and texture features with wavelet and curvelet transform. Curvelet transform has been used not only for image retrieval but also for other image processing applications such as fingerprint recognition [33] and denoising [25]. Multiresolution processing of image has also been applied for texture cartoon analysis of image in [42] and its applications in the field of morphological diversity has been discussed in [43]. A common drawback in most of the above-discussed methods based on curvelet transform is that they fail to discuss and apply the multiresolution property of curvelet transform. Curvelets decompose image into multiple scales and compute coefficients at different orientations. Each scale represents varying levels of details so that features left undetected at one level get considered at another level. This paper attempts to overcome the discussed drawbacks by combining curvelet transform and LBP along with GLCM for feature vector construction, in the form of a novel descriptor, called LBCCP. This descriptor gathers local information at multiple scales and orientations. The feature vector is constructed for each scale separately and retrieval is performed at each scale and then combined at the end thereby producing final retrieval results. This approach considers not only multiresolution aspect of curvelet transform which other methods fail to do but also produces high retrieval accuracy. 3. CURVELET TRANSFORM, LBP, AND GLCM The concept of multiresolution analysis dates back to late 1987 [34]. Although wavelets came into existence early, their multiresolution analysis characteristics were discovered later [34]. Multiresolution analysis has been mainly practiced through Gabor wavelet transform and discrete wavelet transform (DWT). DWT computes coefficients in three directions: horizontal, vertical and diagonal. DWT has proved to be quite useful for various applications of computer vision and has provided promising results [33]. However, wavelets possess certain limitations [25]: Natural images contain highly anisotropic elements which may not be effectively represented by wavelets because they are non-geometrical in nature and fail to represent regularity of edge curve. As the resolution of image becomes finer, a large number of wavelet coefficients are required to represent important edges in the image. This results in repetition of edges of images at multiple scales. At fine scales, a large number of wavelet coefficients are required to reconstruct edges properly. This may cause redundancy of edges scale after scale. Due to this, the features of an image may not be represented effectively. DWT (with Haar wavelet as mother wavelet) computes coefficients in three directions only (horizontal, vertical and diagonal directions). Hence, their ability to resolve directional features is limited. Therefore, wavelets fail to detect curved singularities effectively. To overcome these limitations, other multiresolution techniques such as ridgelet and curvelet were proposed. Ridgelet transform is an important element of curvelet transform. Continuous ridgelet transform provides a sparse representation of smooth as well as of perfectly straight edges. 2D ridgelet transform provides representation of arbitrary bivariate functions f(x1,x2) by superimposition of the elements of the form a−1/2ψ((x1cosθ+x2sinθ−b)/a), where ψ is a wavelet, a > 0 is a scale parameter, θ is an orientation parameter and b is a location scalar parameter. Ridgelets are constant along the lines x1cosθ+x2sinθ, and are wavelets along orthogonal direction. 3.1. Ridgelet transform Let a smooth univariate function ψ:R→R, with sufficient decay and satisfying admissibility condition, be defined as below [25]   ∫|ψˆ(ξ)|2/|ξ|2dξ<∞ (1)which holds if ∫ψ(t)dt=0, that is ψ has a vanishing mean. Let ψ be normalized so that ∫|ψˆ(ξ)|2|ξ|−2dξ=1. Let the bivariate ridgelet ψa,b,θ:R2→R2 be defined as   ψa,b,θ(x)=a−1/2ψ((x1cosθ+x2sinθ−b)/a (2)where a>0, b∈R and θ∈[0,2π). This function is constant along lines (x1cosθ+x2sinθ)=const. Traverse to these ridges, it is a wavelet. For a bivariate function f(x), ridgelet coefficients are defined as   Rf(a,b,θ)=∫ψa,b,θ(x)f(x)dx The reconstruction formula   f(x)=∫02π∫−∞∞∫0∞Rf(a,b,θ)ψa,b,θ(x)daa3dbdθ4π (3)is valid for functions which are both integrable and square integrable. Ridgelet coefficients are calculated by considering ridgelet analysis as wavelet analysis in Radon domain. Radon transform of an object f is defined as collection of line integrals denoted by   Rf(θ,t)=∫f(x1,x2)δ(x1cosθ+x2sinθ−t)dx1dx2 (4)where δ is the Dirac function. The ridgelet coefficients Rf(a,b,θ) of an object f are given by analysis of Radon transform through Rf(a,b,θ)=∫Rf(θ,t)a−1/2ψ((t−b)/a)dt. Edges are curved structure instead of straight lines. The concept of ridgelet transform alone cannot represent edges efficiently. However, edges are almost straight line at fine scales and in order to capture them, ridgelets can be implemented in a localized manner at fine scale. Curvelets are multiscale ridgelets which can accomplish this. Curvelets are defined at all scales, locations and orientations. 3.2. Curvelet transform Discrete curvelet transform of a continuum function f(x1,x2) uses dyadic sequence of scales and a bank of filters (P0f,Δ1f,Δ2f,…) having property that bandpass filter Δs is constructed near frequencies [22s,22s+2]. The curvelet decomposition consists of the following steps [25]: Decomposition of object f into subbands. Smooth windowing of each subband into squares of an appropriate scales. Renormalization of each resulting square to unit scale. Analysis of each square through discrete ridgelet transform. Curvelet transform has the following advantages over wavelet transform: Curvelets are highly anisotropic and efficiently represent sharp edges and curvilinear structures over wavelets. Curvelets use fewer coefficients to represent edges in an image properly. Curvelet transform computes coefficients at multiple scales, locations and orientations. Therefore, curvelets are able to detect curved singularities unlike wavelets. 3.3. LBP LBP was originally proposed by Ojala et al. [35]. LBP operator works in a 3 × 3 pixel block of an image. The pixels are thresholded by the value of centre pixel. The thresholded values are multiplied by the weights given to corresponding pixel values. The LBP operator takes 3 × 3 surrounding of a pixel and Generates a binary 1 if the neighbour value is greater than or equal to centre value. Generates a binary 0 if the neighbour value is less than the centre value. The eight neighbours of the centre can then be represented with an 8-bit number. LBP has the following important properties: It is a simple and efficient local descriptor for describing textures. It encodes the relationship between the grey value of centre pixel and surrounding neighbouring pixels into 0 and 1. It is helpful in extracting local information of an image. As a local feature, when it is combined with global feature acts as a powerful feature vector [36]. 3.4. GLCM GLCM was proposed by Haralick et al. [37]. It is a statistical method for texture analysis of an image. GLCM is a matrix which provides information about how frequently pixel pairs holding specific values and in a specified spatial relationship occur in an image. The occurrence of pixels is in a particular distance and direction. The size of the matrix is determined by maximum grey-level intensity values. GLCM helps in extraction of certain important statistical features. GLCM also helps in determining the spatial arrangement of pixels in an image. Computation of GLCM is demonstrated in Fig. 1 with the help of an example. Figure 1(a) shows the original matrix and Fig. 1(b) shows GLCM constructed for original matrix. In Fig. 1(b), the topmost row and the leftmost column represent pixel values that appear in original matrix. The entries in GLCM represent the number of times pixel pairs appear in original matrix. For example, the pixel pair (1, 1) appears three times in original matrix which is shown in GLCM. Figure 1. View largeDownload slide (a) Original matrix. (b) GLCM for original matrix. Figure 1. View largeDownload slide (a) Original matrix. (b) GLCM for original matrix. 3.5. LBCCP Modern image-capturing devices are capable of producing high-resolution images due to which the images have become more complex more than they were earlier. They contain small as well as large size objects and low- as well as high-resolution objects. In order to perform retrieval of such types of images, single-resolution processing proves to be insufficient as different types of curved structures may be present in the image and processing at single resolution fails to capture them efficiently. This may lead to weak feature vector construction which may result in low retrieval accuracy [7, 23]. The proposed LBCCP descriptor tends to overcome these limitations by gathering local information in an image through LBP at multiple scales with the help of curvelet transform. Natural images have complex and irregular texture. In order to capture these features, LBP proves to be an effective feature descriptor. LBP captures local information from an image by computing relationship of an intensity value with its neighbourhood pixels. This helps in extracting information of structural arrangement of pixels. However, LBP fails to gather directional information since it does not incorporate any technique which is helpful in gathering details from different directions. The combination of LBP with curvelet transform tends to overcome this limitation. Curvelet transform being highly anisotropic in nature gathers sharp curvilinear and geometric structures at multiple orientations and scales. This helps in capturing complex details more effectively than LBP exploited at single resolution of image. Curvelet transform decomposes an image into multiple resolutions for gathering features. The advantage of this is that features left undetected at one scale get detected at another scale. The spatial distribution of LBP codes at different scales of images are gathered by constructing GLCM. This descriptor helps in getting texture information at multiple resolutions by computing LBP of curvilinear structures and sharp edges which other techniques such as single-resolution processing and other multiresolution techniques fail to gather. The proposed LBCCP descriptor has the following important properties which are useful for image retrieval: Due to its anisotropic nature, it effectively represents curvilinear structures and sharp edges. It uses fewer coefficients to reconstruct edges in an image. In this way, it reduces redundancy of coefficients thereby constructing efficient feature vector for retrieval at multiple scales unlike wavelet-based technique. It helps in construction of feature vector at multiple scales, orientations and locations due to which it is capable to capture complex geometrical structure which wavelet-based technique fails to do. It effectively gathers spatial distribution of LBP codes at multiple scales through GLCM which other techniques such as histogram fail to do. 4. THE PROPOSED METHOD The proposed method consists of four steps: Computation of curvelet coefficients of grayscale image. Computation of LBP codes of curvelet coefficients computed in Step 1. Computation of GLCM. Similarity measurement. The schematic diagram for the proposed method is shown in Fig. 2. Figure 2. View largeDownload slide Schematic diagram for the proposed method. Figure 2. View largeDownload slide Schematic diagram for the proposed method. 4.1. Computation of curvelet coefficient The first step of the proposed method is computation of curvelet coefficients. Candes et al. [38] proposed two versions of fast discrete curvelet transforms, namely, unequally spaced fast Fourier transform (USFFT) which is based on unequally fast Fourier transforms, and fast discrete curvelet transform warping (FDCT_WARPING) which is based on wrapping of specially selected Fourier samples. Both versions of curvelet transforms are fast and efficiently compute curvelet coefficients. The proposed method uses USFFT curvelet transform for computation of curvelet coefficients as it is easy to understand and compute. Application of curvelet transform on grayscale image of size 256 × 256 decomposes image into five cells using formula Nscales=ceil(log2(min(N1,N2))−3) [33] from coarse to fine scales:   ({1×1},{1×32},{1×32},{1×64},{1×1}) Here Nscales denotes the number of scales or resolutions of image, N1 and N2 denote the size of image. These five cells further consist of sub-cells that contain curvelet coefficients. These cells denote coefficients computed at different scales, locations and orientations. e.g. {1 × 32} cell denotes that the cell consists of 32 sub-cells of different sizes which hold values of curvelet coefficients computed at different orientations. Table 1 shows the number of sub-cells each cell contains for a 256 × 256 and 128 × 128 image. Table 1. Number of sub-cells each cell contains. Cell  Number of sub-cells  (a) For 256 × 256 image   {1 × 1}  1   {1 × 32}  32   {1 × 32}  32   {1 × 64}  64   {1 × 1}  1  (b) For 128 × 128 image   {1 × 1}  1   {1 × 32}  32   {1 × 64}  64   {1 × 1}  1  Cell  Number of sub-cells  (a) For 256 × 256 image   {1 × 1}  1   {1 × 32}  32   {1 × 32}  32   {1 × 64}  64   {1 × 1}  1  (b) For 128 × 128 image   {1 × 1}  1   {1 × 32}  32   {1 × 64}  64   {1 × 1}  1  When discrete curvelet transform is applied on 2D grayscale image, it decomposes image into different levels of resolution as shown in Table 1. Each cell is considered as a level of resolution and exploited for constructing feature vector. Each level of resolution of curvelet transform consists of coefficients computed at different locations and orientations. The coarse scale consists of low-frequency coefficients and fine scale consists of high-frequency coefficients [32]. The different levels of resolution attempt to gather small as well as large size objects and low as well as high-resolution objects. The LBP codes of resulting curvelet coefficients are computed for each level of resolution separately. The construction of feature vector is done through GLCM. Feature vector for each cell is computed separately and is used for performing retrieval of similar images. This produces a set of similar images for each level separately which are combined to produce final image set. 4.2. Computation of LBP LBP gives description of the surroundings of a pixel by producing a bit-code from binary derivatives of a pixel. LBP takes 3 × 3 neighbourhood of a pixel and generates a binary 1 if the neighbour is greater than or equal to the centre pixel or a binary 0 if the neighbour is less than the centre pixel. This generates an 8-bit code known as LBP code. LBP codes for each cell of curvelet coefficients are computed separately. For a 256 × 256 image, the total number of cells is 5. Each cell further consists of sub-cells which contains curvelet coefficients. LBP codes of coefficients in each case of these sub-cells is computed and stored in separate matrices. 4.3. Computation of GLCM After computation of LBP codes of curvelet coefficients, the next step is construction of feature vector. Feature vector is constructed through GLCM. GLCM is computed for each LBP matrix separately. GLCM determines frequency of co-occurrence of pixels pairs. This gives information about spatial distribution of intensity values in an image which helps in determining structural arrangement of pixels in an image. All these information are not provided by other features such as histogram which provides information about frequency of intensity values only. GLCM in 0° angle with Distance 1 has been used in the proposed method. GLCM has been scaled to size 8 × 8 matrix in the proposed method. 4.4. Similarity measurement Similarity measurement is done for each image of dataset to retrieve visually similar images. The similarity measurement is done by measuring distance between feature vector of query image and database images. The feature vector in the proposed method is constructed through computation of GLCM of LBP codes. Hence, in the proposed method, similarity measurement is done by measuring distance between GLCM of query image and database images. Let the GLCM of query image be denoted by GQ=(GQ1,GQ2,…,GQn) and let the GLCM of database images be denoted by GDB=(GDB1,GDB2,…,GDBn). Then the Euclidean distance between query image and database image is given as   D(GQ,GDB)=∑(GQi−GDBi)2,i=1,2,…,n (5)where D denotes the distance between feature vectors, GQ denotes the feature vector of query image and GDB denotes the feature vector of database image. 5. EXPERIMENT AND RESULTS To perform experiment using the proposed method, images from following five benchmark datasets have been used. These datasets consist of a wide variety of images and are widely used for evaluation of image retrieval. Dataset 1 (Corel-1K) The first dataset used in this experiment is Corel-1K dataset [39]. It consists of 1000 images. The images in this dataset are classified into 10 different categories, namely, Africans, Beaches, Buildings, Buses, Dinosaurs, Elephants, Flowers, Horses, Mountains and Food. Each category consists of 100 images. The size of each image is either 256 × 384 or 384 × 256. Dataset 2 (Olivia-2688) The second dataset used to measure the performance of the proposed method is Olivia-2688 dataset [40]. It consists of 2688 images. The images in this dataset are divided into eight categories, namely, Coast, Forest, Highway, Inside City, Mountain, Open Country, Street and Tall Building. Each category consists of different numbers of images ranging from maximum 410 to minimum 260. The size of each image is 256 × 256. Dataset 3 (Corel-5K) The third dataset used in this experiment is Corel-5K dataset [41]. It consists of 5000 images. The images in this dataset are divided into 50 categories consisting of different types of images in various categories ranging from animals, human beings to sunsets, card, etc. Each category consists of 100 images. The size of each image is either 187 × 128 or 128 × 187. Dataset 4 (Corel-10K) The fourth dataset used for testing the proposed method is Corel-10K dataset [41] which is an extension of Corel-5K dataset. It consists of 10 000 images. The images in this dataset are divided into hundred categories consisting of a wide variety of images. Each category consists of 100 images. Each image is of size 187 × 128 or 128 × 187. Dataset 5 (GHIM-10K) The fifth dataset used in this experiment is GHIM-10K dataset [41]. It consists of 10 000 images. The images in this dataset are divided into 20 categories consisting of various types of images such as horses, insects and flowers. Each category consists of 500 images. The size of each image is either 300 × 400 or 400 × 300. Each image of datasets Corel-1K, Olivia-2688, GHIM-10K has been rescaled to size 256 × 256 (28 × 28) and images of datasets Corel-5K and Corel-10K to size 128 × 128 (27 × 27) to ease the computation. Sample images from each dataset are shown in Fig. 3. Each image of all datasets is taken as query image. If the retrieved images belong to the same category as that of the query image, the retrieval is considered to be successfully, otherwise the retrieval fails. Figure 3. View largeDownload slide Sample images from datasets. Figure 3. View largeDownload slide Sample images from datasets. 5.1. Performance evaluation Performance of the proposed method has been evaluated in terms of precision and recall. Precision is defined as the ratio of the total number of relevant images retrieved to the total number of images retrieved. Mathematically, precision can be formulated as   P=IRTR (6)where IR denotes the total number of relevant images retrieved and TR denotes the total number of images retrieved. Recall is defined as the ratio of total number of relevant images retrieved to the total number of relevant images in the database. Mathematically, recall can be formulated as   R=IRCR (7)where IR denotes the total number of relevant images retrieved and CR denotes the total number of relevant images in the database. In this experiment, TR = 10 and the value of CR varies for different datasets. The value of CR depends on the total number of images in each category of image of datasets. In this experiment, for Corel-1K, Corel-5K and Corel-10K datasets, the value of CR is 100. For GHIM-10K dataset, the value of CR is 500 and for Olivia-2688, the value of CR is different for each category of image depending on the total number of images in each category. 5.2. Retrieval results at multiple resolutions To perform the experiments, the images of datasets Corel-1K, Olivia-2688 and GHIM-10K have been rescaled to size 256 × 256 and images of datasets Corel-5K and Corel-10K have been rescaled to size 128 × 128, to ease the computation. First step of the proposed method is the computation of curvelet transform coefficients of grayscale images followed by computation of LBP codes of resulting coefficients. Finally, feature vector is constructed using GLCM. Application of curvelet transform on grayscale images produces five cells of coefficients for 256 × 256 image and four cells of coefficients for 128 × 128 image. Each of these cells further contains sub-cells of different sizes that hold curvelet coefficients of grayscale image. LBP codes for all sub-cells are computed separately followed by construction of GLCM of each cell. In this experiment, similarity measurement for each of these sub-cells is done separately. This produces five sets of similar images. Union of all these sets is taken to produce a final set of similar images. Recall is computed by counting the total number of relevant images in the final set. Similarly, for precision, top n matches for each set is counted and then union operation is applied on all sets to produce final image set. Mathematically, this can be stated as follows. Let f1,f2,…,fm be the set of similar images obtained from feature vector of a sub-cell. Then, the final set of similar images denoted by fRS is given by   fRS=f1∪f2∪⋯∪fm (8) Similarly, let f1n,f2n,…,fmn be set of top n images obtained from feature vector of a sub-cell. Then, the final set of top n images is denoted by fPSn is given as   fPSn=f1n∪f2n∪⋯∪fmn (9) The above procedure is repeated for all cells of a resolution. In every level, the relevant image set of previous level is also considered and is combined with the current level to produce relevant set for that level. The proposed method has been implemented in MATLAB R2013a on a PC having Windows 8.1 Pro operating system, Intel Core 15-4570 processor at 3.20 GHz and 4 GB of RAM. The average retrieval time taken by the proposed method for 100 images is 0.043 s. Table 2 shows the performance of the proposed method on all datasets used (Corel-1K, Olivia-2688, Corel-5K, Corel-10K, GHIM-10K) at different levels. Figures 4 and 5 show the plots between precision vs. dataset and recall vs. dataset, respectively, for the proposed method on all five datasets for different levels of resolutions. Figure 6 shows the plots between precision and recall for different levels of resolution for the proposed method on all five datasets. Table 2. Average precision and recall for different levels of resolution for all datasets.   Recall (%)  Precision (%)  (a) Corel-1K dataset   Level 1  20.73  37.09   Level 2  86.31  100   Level 3  98.02  100   Level 4  99.02  100   Level 5  99.07  100  (b) Olivia-2688 dataset   Level 1  25.86  45.49   Level 2  91.02  100   Level 3  91.32  100   Level 4  99.96  100   Level 5  99.96  100  (c) Corel-5K dataset   Level 1  9.22  22.86   Level 2  36.41  78.19   Level 3  58.29  98.75   Level 4  61.99  99.22  (d) Corel-10K dataset   Level 1  6.47  18.91   Level 2  21.97  50.60   Level 3  37.70  80.73   Level 4  41.67  84.80  (e) GHIM-10K dataset   Level 1  10.79  25.65   Level 2  62.02  99.69   Level 3  85.29  100   Level 4  97.20  100   Level 5  97.32  100    Recall (%)  Precision (%)  (a) Corel-1K dataset   Level 1  20.73  37.09   Level 2  86.31  100   Level 3  98.02  100   Level 4  99.02  100   Level 5  99.07  100  (b) Olivia-2688 dataset   Level 1  25.86  45.49   Level 2  91.02  100   Level 3  91.32  100   Level 4  99.96  100   Level 5  99.96  100  (c) Corel-5K dataset   Level 1  9.22  22.86   Level 2  36.41  78.19   Level 3  58.29  98.75   Level 4  61.99  99.22  (d) Corel-10K dataset   Level 1  6.47  18.91   Level 2  21.97  50.60   Level 3  37.70  80.73   Level 4  41.67  84.80  (e) GHIM-10K dataset   Level 1  10.79  25.65   Level 2  62.02  99.69   Level 3  85.29  100   Level 4  97.20  100   Level 5  97.32  100  Figure 4. View largeDownload slide Average precision vs. level of resolution for the proposed method on (a) Corel-1K Olivia-2688 and GHIM-10K. (b) Corel-5K and Corel-10K. Figure 4. View largeDownload slide Average precision vs. level of resolution for the proposed method on (a) Corel-1K Olivia-2688 and GHIM-10K. (b) Corel-5K and Corel-10K. Figure 5. View largeDownload slide Average recall vs. level of resolution for the proposed method on (a) Corel-1K. (b) Olivia-2688 and GHIM-10K. (b) Corel-5K and Corel-10K. Figure 5. View largeDownload slide Average recall vs. level of resolution for the proposed method on (a) Corel-1K. (b) Olivia-2688 and GHIM-10K. (b) Corel-5K and Corel-10K. Figure 6. View largeDownload slide Precision vs. recall plot for the proposed method on (a) Corel-1K. (b) Olivia-2688. (c) Corel-5K. (d) Corel-10K. (e) GHIM-10K datasets. Figure 6. View largeDownload slide Precision vs. recall plot for the proposed method on (a) Corel-1K. (b) Olivia-2688. (c) Corel-5K. (d) Corel-10K. (e) GHIM-10K datasets. From the above experimental observations on different datasets, it is clearly observed that the average values of precision and recall increase with the level of resolution. Due to multiresolution analysis, each level covers details which were undetected at previous levels. This phenomenon leads to an increase in the values of precision and recall at different levels of resolution. Curvelets represent curvilinear structures and edges more effectively than wavelets due to their geometrical characteristics. Hence, curvelets construct more effective feature vector as compared to wavelets because they are able to extract more details at edges present in the natural image at multiple scales and directions. These factors result in high retrieval accuracy which can be observed in experimental results. 5.3. Performance comparison among single-resolution technique, wavelet-based technique and the proposed method To show the effectiveness of use of curvelet in the proposed method, the comparison of the proposed descriptor with wavelet-based technique and single resolution has been performed. Processing the single resolution of an image is simple and considers single scale of an image for feature extraction. A number of state-of-the-art methods exploit single resolution of image to construct feature vector. Some of the methods that have exploited single resolution of image for image retrieval produce promising results. However, single resolution fails to gather varying levels of details in an image. Hence, when compared with a multiresolution processing technique, single-resolution processing fails to yield better results. This phenomenon can be observed in Tables 3 and 4 where other multiresolution techniques produce much better results and outperform single-resolution processing technique denoted by SR. Table 3. Performance comparison of the proposed descriptor (LBCCP) with wavelet-based descriptor (LBWCP).   Corel-1K  Olivia-2688  GHIM-10K    LBWCP  LBCCP  LBWCP  LBCCP  LBWCP  LBCCP  (a) In terms of recall (%)   Level 1  44.16  20.73  36.98  25.86  23.75  10.79   Level 2  59.18  86.31  57.36  91.02  35.60  62.02   Level 3  69.24  98.02  77.50  91.32  44.28  85.29   Level 4  77.13  99.02  79.38  99.96  51.94  97.20   Level 5  83.18  99.07  86.15  99.96  58.72  97.32  (b) In terms of precision (%)   Level 1  79.68  37.09  79.32  45.49  64.02  22.65   Level 2  91.43  100  93.48  100  76.66  99.69   Level 3  96.82  100  97.73  100  84.41  100   Level 4  98.65  100  99.23  100  89.79  100   Level 5  99.40  100  99.83  100  93.49  100    Corel-1K  Olivia-2688  GHIM-10K    LBWCP  LBCCP  LBWCP  LBCCP  LBWCP  LBCCP  (a) In terms of recall (%)   Level 1  44.16  20.73  36.98  25.86  23.75  10.79   Level 2  59.18  86.31  57.36  91.02  35.60  62.02   Level 3  69.24  98.02  77.50  91.32  44.28  85.29   Level 4  77.13  99.02  79.38  99.96  51.94  97.20   Level 5  83.18  99.07  86.15  99.96  58.72  97.32  (b) In terms of precision (%)   Level 1  79.68  37.09  79.32  45.49  64.02  22.65   Level 2  91.43  100  93.48  100  76.66  99.69   Level 3  96.82  100  97.73  100  84.41  100   Level 4  98.65  100  99.23  100  89.79  100   Level 5  99.40  100  99.83  100  93.49  100    Corel-5K  Corel-10K    LBWCP  LBCCP  LBWCP  LBCCP  (c) In terms of recall (%)   Level 1  22.75  9.22  16.32  6.47   Level 2  29.28  36.41  20.42  21.97   Level 3  34.23  58.29  23.53  37.70   Level 4  38.44  61.99  26.15  41.67  (d) In terms of precision (%)   Level 1  46.26  22.86  36.45  18.91   Level 2  54.66  78.19  42.04  50.60   Level 3  60.71  98.75  45.96  80.73   Level 4  65.67  99.22  49.12  84.80    Corel-5K  Corel-10K    LBWCP  LBCCP  LBWCP  LBCCP  (c) In terms of recall (%)   Level 1  22.75  9.22  16.32  6.47   Level 2  29.28  36.41  20.42  21.97   Level 3  34.23  58.29  23.53  37.70   Level 4  38.44  61.99  26.15  41.67  (d) In terms of precision (%)   Level 1  46.26  22.86  36.45  18.91   Level 2  54.66  78.19  42.04  50.60   Level 3  60.71  98.75  45.96  80.73   Level 4  65.67  99.22  49.12  84.80  Table 4. Performance analysis of the proposed method LBCCP with other techniques. Dataset  Performance analysis of the proposed method    SR  LBWCP  Corel-1K  Precision  Proposed method outperforms SR by 45.69%  Proposed method outperforms LBWCP by 0.60%  Recall  Proposed method outperforms SR by 68.40%  Proposed method outperforms LBWCP by 15.89%  Olivia-2688  Precision  Proposed method outperforms SR by 44.58%  Proposed method outperforms LBWCP by 0.17%  Recall  Proposed method outperforms SR by 73.77%  Proposed method outperforms LBWCP by 13.81%  Corel-5K  Precision  Proposed method outperforms SR by 64.18%  Proposed method outperforms LBWCP by 33.55%  Recall  Proposed method outperforms SR by 47.74%  Proposed method outperforms LBWCP by 23.55%  Corel-10K  Precision  Proposed method outperforms SR by 57.18%  Proposed method outperforms LBWCP by 35.68%  Recall  Proposed method outperforms SR by 31.60%  Proposed method outperforms LBWCP by 15.52%  GHIM-10K  Precision  Proposed method outperforms SR by 62.68%  Proposed method outperforms LBWCP by 6.51%  Recall  Proposed method outperforms by 71.13%  Proposed method outperforms by 38.60%  Dataset  Performance analysis of the proposed method    SR  LBWCP  Corel-1K  Precision  Proposed method outperforms SR by 45.69%  Proposed method outperforms LBWCP by 0.60%  Recall  Proposed method outperforms SR by 68.40%  Proposed method outperforms LBWCP by 15.89%  Olivia-2688  Precision  Proposed method outperforms SR by 44.58%  Proposed method outperforms LBWCP by 0.17%  Recall  Proposed method outperforms SR by 73.77%  Proposed method outperforms LBWCP by 13.81%  Corel-5K  Precision  Proposed method outperforms SR by 64.18%  Proposed method outperforms LBWCP by 33.55%  Recall  Proposed method outperforms SR by 47.74%  Proposed method outperforms LBWCP by 23.55%  Corel-10K  Precision  Proposed method outperforms SR by 57.18%  Proposed method outperforms LBWCP by 35.68%  Recall  Proposed method outperforms SR by 31.60%  Proposed method outperforms LBWCP by 15.52%  GHIM-10K  Precision  Proposed method outperforms SR by 62.68%  Proposed method outperforms LBWCP by 6.51%  Recall  Proposed method outperforms by 71.13%  Proposed method outperforms by 38.60%  The second method used for comparison is wavelet-based technique which employs discrete wavelet transform (DWT) (using Daubechies1 Haar wavelet) with LBP and GLCM denoted by local binary wavelet co-occurrence pattern (LBWCP). LBWCP has been compared with the proposed descriptor, LBCCP, in terms of precision and recall for five levels of resolution in case of 256 × 256 image and four levels of resolution in case of 128 × 128 image size. LBWCP produces promising results as wavelet gathers directional information at multiple resolution of image. However, wavelets fail to gather details at sharp edges and curvilinear structures due to their non-geometrical nature. Hence, wavelets fail to construct as efficient feature vector as curvelet. It is worth noticing that Level 1 decomposition of DWT produces better retrieval results as compared with Level 1 decomposition of curvelet as shown in Table 3. This is because Level 1 decomposition of DWT computes coefficients in three directions, horizontal, vertical and diagonal, and hence produces three separate matrices for feature vector whereas curvelet transform computes coefficients at single orientation in Level 1 decomposition. Therefore, it produces single matrix for feature vector. However, when compared at successive levels, curvelet transform computes coefficients at multiple orientations and hence is able to extract details at edges and curvilinear structures which wavelet transform fails to do. Hence, the overall retrieval result is far better in case of curvelet-based technique as shown in Table 3 (highlighted in bold). Hence, LBCCP performs better than LBWCP descriptor due to its anisotropic nature and ability to extract features at curvature and sharp edges using fewer coefficients. Table 5 shows the performance comparison of the proposed method (LBCCP) with other methods (single-resolution (SR) and wavelet-based technique (LBWCP)) in terms of recall on five datasets (Corel-1K, Olivia-2688, Corel-5K, Corel-10K, GHIM-10K). It can be clearly observed that the proposed method achieves much higher recall values than other techniques on five datasets which have been highlighted in bold. Similarly, Table 6 shows the performance comparison of the proposed method with other techniques in terms of precision. In case of precision also, the proposed method clearly outperforms other techniques. The precision values of the proposed method have been highlighted in bold. It can be observed that the proposed method descriptor LBCCP achieves 100% precision accuracy on certain datasets. For precision, we have considered top 10 matches. The final level of LBCCP obtains all top 10 matches belonging to the same category as that of the query image. Due to this reason, the proposed descriptor achieves 100% precision value in these cases. The performance analysis of the proposed method with other techniques on five datasets is shown in Table 4. Table 5. Comparison of the proposed method (PM) with other techniques in terms of recall (%). Datasets  Single resolution (SR)  LBWCP  LBCCP (PM)  Corel-1K  30.67  83.18  99.07  Olivia-2688  26.19  86.15  99.96  Corel-5K  14.25  38.44  61.99  Corel-10K  10.07  26.15  41.67  GHIM-10K  26.19  58.72  97.32  Datasets  Single resolution (SR)  LBWCP  LBCCP (PM)  Corel-1K  30.67  83.18  99.07  Olivia-2688  26.19  86.15  99.96  Corel-5K  14.25  38.44  61.99  Corel-10K  10.07  26.15  41.67  GHIM-10K  26.19  58.72  97.32  Table 6. Comparison of the proposed method (PM) with other techniques in terms of precision (%). Datasets  Single resolution (SR)  LBWCP  LBCCP (PM)  Corel-1K  54.31  99.40  100  Olivia-2688  55.42  99.83  100  Corel-5K  35.04  65.67  99.22  Corel-10K  27.62  49.12  84.80  GHIM-10K  37.32  93.49  100  Datasets  Single resolution (SR)  LBWCP  LBCCP (PM)  Corel-1K  54.31  99.40  100  Olivia-2688  55.42  99.83  100  Corel-5K  35.04  65.67  99.22  Corel-10K  27.62  49.12  84.80  GHIM-10K  37.32  93.49  100  5.4. Performance comparison of the proposed method with other state-of-the-art methods The performance of the proposed method is compared with other state-of-the-art methods (Srivastava et al. [7], Srivastava et al. [12], Srivastava et al. [23], MSD [14–16]). Table 9. Performance analysis of the proposed method with other state-of-the-art CBIR methods. Dataset  Performance analysis of the proposed method      Srivastava et al. [7]  Srivastava et al. [12]  Srivastava et al. [23]  MSD [14–16]  Corel-1K  Precision  Proposed method outperforms by 64.06%  Proposed method outperforms by 46.30%  Proposed method outperforms by 32.84%  Proposed method outperforms by 24.33%  Recall  Proposed method outperforms by 46.28%  Proposed method outperforms by 27.87%  Proposed method outperforms by 48.98%  Proposed method outperforms by 90.00%  Olivia-2688  Precision  Proposed method outperforms by 86.62%  Proposed method outperforms by 38.87%  Proposed method outperforms by 10.40%  Proposed method outperforms by 52.16%  Recall  Proposed method outperforms by 87.25%  Proposed method outperforms by 63.51%  Proposed method outperforms by 40.85%  Proposed method outperforms by 98.20%  Corel-5K  Precision  Proposed method outperforms by 82.83%  Proposed method outperforms by 67.04%  Proposed method outperforms by 67.27%  Proposed method outperforms by 43.30%  Recall  Proposed method outperforms by 54.67%  Proposed method outperforms by 46.63%  Proposed method outperforms by 42.11%  Proposed method outperforms by 55.28%  Corel-10K  Precision  Proposed method outperforms by 71.65%  Proposed method outperforms by 59.88%  Proposed method outperforms by 62.45%  Proposed method outperforms by 39.18%  Recall  Proposed method outperforms by 37.54%  Proposed method outperforms by 30.73%  Proposed method outperforms by 30.57%  Proposed method outperforms by 36.19%  GHIM-10K  Precision  Proposed method outperforms by 80.66%  Proposed method outperforms by 62.77%  Proposed method outperforms by 50.82%  Proposed method outperforms by 47.98%  Recall  Proposed method outperforms by 88.83%  Proposed method outperforms by 79.68%  Proposed method outperforms by 66.32%  Proposed method outperforms by 94.88%  Dataset  Performance analysis of the proposed method      Srivastava et al. [7]  Srivastava et al. [12]  Srivastava et al. [23]  MSD [14–16]  Corel-1K  Precision  Proposed method outperforms by 64.06%  Proposed method outperforms by 46.30%  Proposed method outperforms by 32.84%  Proposed method outperforms by 24.33%  Recall  Proposed method outperforms by 46.28%  Proposed method outperforms by 27.87%  Proposed method outperforms by 48.98%  Proposed method outperforms by 90.00%  Olivia-2688  Precision  Proposed method outperforms by 86.62%  Proposed method outperforms by 38.87%  Proposed method outperforms by 10.40%  Proposed method outperforms by 52.16%  Recall  Proposed method outperforms by 87.25%  Proposed method outperforms by 63.51%  Proposed method outperforms by 40.85%  Proposed method outperforms by 98.20%  Corel-5K  Precision  Proposed method outperforms by 82.83%  Proposed method outperforms by 67.04%  Proposed method outperforms by 67.27%  Proposed method outperforms by 43.30%  Recall  Proposed method outperforms by 54.67%  Proposed method outperforms by 46.63%  Proposed method outperforms by 42.11%  Proposed method outperforms by 55.28%  Corel-10K  Precision  Proposed method outperforms by 71.65%  Proposed method outperforms by 59.88%  Proposed method outperforms by 62.45%  Proposed method outperforms by 39.18%  Recall  Proposed method outperforms by 37.54%  Proposed method outperforms by 30.73%  Proposed method outperforms by 30.57%  Proposed method outperforms by 36.19%  GHIM-10K  Precision  Proposed method outperforms by 80.66%  Proposed method outperforms by 62.77%  Proposed method outperforms by 50.82%  Proposed method outperforms by 47.98%  Recall  Proposed method outperforms by 88.83%  Proposed method outperforms by 79.68%  Proposed method outperforms by 66.32%  Proposed method outperforms by 94.88%  Srivastava et al. [7] exploit shape feature for retrieval through moments. It uses moments as a single feature, constructs feature vector by dividing image into blocks of different sizes and computing geometric moments of each block. The method is simple and efficiently computes shape feature for retrieval. However, as discussed in the previous section, single feature is insufficient for retrieving similar images. Also, the method exploits single resolution of an image and fails to capture varying levels of details in an image. Hence, this technique produces low retrieval accuracy when compared with the proposed method as shown in Tables 7 and 8 and Figs. 7 and 8. Table 7. Performance comparison of the proposed method (PM) with other state-of-the-art CBIR methods in terms of precision (%). Datasets  Srivastava et al. [7]  Srivastava et al. [12]  Srivastava et al. [23]  MSD [14–16]  LBCCP (PM)  Corel-1K  35.94  53.70  67.16  75.67  100  Olivia-2688  13.38  61.13  89.60  47.84  100  Corel-5K  16.39  32.18  31.95  55.92  99.22  Corel-10K  13.15  24.92  22.35  45.62  84.80  GHIM-10K  19.34  37.23  49.18  52.02  100  Datasets  Srivastava et al. [7]  Srivastava et al. [12]  Srivastava et al. [23]  MSD [14–16]  LBCCP (PM)  Corel-1K  35.94  53.70  67.16  75.67  100  Olivia-2688  13.38  61.13  89.60  47.84  100  Corel-5K  16.39  32.18  31.95  55.92  99.22  Corel-10K  13.15  24.92  22.35  45.62  84.80  GHIM-10K  19.34  37.23  49.18  52.02  100  Table 8. Performance comparison of the proposed method (PM) with other state-of-the-art CBIR methods in terms of recall (%). Datasets  Srivastava et al. [7]  Srivastava et al. [12]  Srivastava et al. [23]  MSD [14–16]  LBCCP (PM)  Corel-1K  52.79  72.09  50.09  9.07  99.07  Olivia-2688  12.71  36.45  59.11  1.76  99.96  Corel-5K  7.32  15.36  19.88  6.71  61.99  Corel-10K  4.13  10.94  11.10  5.48  41.67  GHIM-10K  8.49  17.64  31.00  2.44  97.32  Datasets  Srivastava et al. [7]  Srivastava et al. [12]  Srivastava et al. [23]  MSD [14–16]  LBCCP (PM)  Corel-1K  52.79  72.09  50.09  9.07  99.07  Olivia-2688  12.71  36.45  59.11  1.76  99.96  Corel-5K  7.32  15.36  19.88  6.71  61.99  Corel-10K  4.13  10.94  11.10  5.48  41.67  GHIM-10K  8.49  17.64  31.00  2.44  97.32  Figure 7. View largeDownload slide Performance comparison of the proposed method with other state-of-the-art CBIR methods on five datasets in terms of precision. Figure 7. View largeDownload slide Performance comparison of the proposed method with other state-of-the-art CBIR methods on five datasets in terms of precision. Figure 8. View largeDownload slide Performance comparison of the proposed method with other state-of-the-art CBIR methods on five datasets in terms of recall. Figure 8. View largeDownload slide Performance comparison of the proposed method with other state-of-the-art CBIR methods on five datasets in terms of recall. The second method used for comparison is Srivastava et al. [12] which uses combination of features for retrieval. This method combines local ternary pattern with moments to retrieve visually similar images. The method uses combination of local and global features, thereby, overcoming limitations of single feature. The method extracts more details from an image as compared with single feature. This method exploits combination of features at single resolution and produces good retrieval accuracy on small datasets. However, this method fails to produce high accuracy for large datasets as such datasets consist of a wide variety of images containing varying levels of details. Being exploited at single resolution of image, this method proves to be less efficient when compared with the proposed method which produces high retrieval accuracy for small as well as large datasets. Hence, the proposed method outperforms Srivastava et al. [12] as illustrated in Tables 7 and 8 and Figs. 7 and 8. The third method used for comparison with the proposed method is Srivastava et al. [23]. The method exploits shape feature through moments at multiple resolutions of image using discrete wavelet transform. The method uses multiple resolutions of image to extract shape feature in order to construct feature vector for retrieval. However, as discussed in the previous section, wavelets fail to capture curves and edges efficiently and hence, do not construct as efficient feature vector as constructed by the proposed method. The proposed method employs curvelet transform for constructing feature vector which efficiently captures curvilinear structures and sharp edges due to its anisotropic nature. Hence, the proposed method outperforms Srivastava et al. [23] also as shown in Tables 7 and 8 and Figs. 7 and 8. The fourth method used for comparison with the proposed method is MSD [14]. It is an efficient technique of retrieval which employs edge orientation similarity along with underlying colours in order to construct feature vector. The technique attempts to exploit similar edge orientation for feature vector construction. Also, single MSD feature efficiently extracts colour, texture and shape features simultaneously due to which it proves to be a powerful feature for retrieval. Although MSD feature produces promising results, it too exploits single resolution of image for feature vector construction and fails to consider varying levels of details. Hence, when compared with the proposed technique, MSD [14] technique produces low retrieval results. This fact can be observed in Tables 7 and 8 and Figs. 7 and 8. The methods (Srivastava et al. [7], Srivastava et al. [12] and Srivastava et al. [23]) have been implemented in MATLAB R2013a on a PC having Windows 8.1 Pro operating system, Intel Core 15-4570 processor at 3.20 GHz and 4 GB of RAM. The precision and recall values of MSD [14] have been referred to from [14–16]. The proposed method employing curvelet transform for CBIR outperforms the above-discussed state-of-the-art methods (Srivastava et al. [7], Srivastava et al. [12], Srivastava et al. [23] and MSD [14–16]) due to its anisotropic nature, ability to extract features at curvatures and sharp edges and exploitation of multiple resolutions of image for feature vector construction. These properties of the proposed method help in capturing varying levels of details in an image and produce high retrieval accuracy. Tables 7 and 8 and Figs. 7 and 8 show the performance comparison of the proposed method with other state-of-the-art CBIR methods in terms of precision and recall on five datasets (Corel-1K, Olivia-2688, Corel-5K, Corel-10K and GHIM-10K) used in this paper. Table 8 shows the performance comparison of the proposed method with other state-of-the-art methods in terms of precision. Similarly, Table 8 shows the performance comparison of the proposed method with other state-of-the-art image retrieval methods. It can be clearly observed that the proposed method outperforms other techniques both in terms of precision and recall by a huge margin. The performance analysis of the proposed method with other state-of-the-art CBIR techniques on five datasets is shown in the form of a table in Table 9. 6. CONCLUSION In this paper, we presented a new descriptor, namely, LBCCP for CBIR. The proposed method combines LBP with curvelet transform followed by construction of GLCM for feature vector generation. The feature vector value was compared with those of images in database to retrieve visually similar images. In a nutshell, the proposed method has following advantages: The proposed method exploits multiresolution LBP and gathers local information at multiple scales at curves and edges. Feature vector construction through GLCM is advantageous as it considers spatial arrangement of pixels which the histogram fails to do. The proposed method yields high retrieval accuracy. Performance of the proposed method has been measured in terms of precision and recall. The proposed method, when compared with single-resolution technique and wavelet-based technique as well as other state-of-the-art techniques, outperforms them as demonstrated through experimental results. The proposed combination of LBP and curvelet transform can be useful for other computer vision applications such as face recognition, object recognition and biometric image processing due to useful properties of LBP and curvelet transform. The proposed method can further be improved by combining more features such as shape with local features at multiresolution level to extract more details from an image. REFERENCES 1 Long, F., Zhang, H. and Feng, D.D. ( 2003) Fundamentals of Content-Based Image Retrieval. Multimedia Information Retrieval and Management . Springer, Berlin. 2 Smith, J.R. and Chang, S.F. ( 1996) Tools and Techniques for Color Image Retrieval. Electronic Imaging: Science and Technology , pp. 426– 437. International Society for Optics and Photonics. 3 Huang, J., Kumar, S.R., Mitra, M. and Zhu, W. ( 2001) Image Indexing using Color Correlograms, U.S. Patent 6,246,790. 4 Pass, G., Zabih, R. and Miller, J. ( 1997) Comparing images using color coherence vectors. In Proc. 4th ACM Int. Conf. Multimedia, Boston, MA, USA, November 18–22, pp. 65–73. ACM, New York, USA. 5 Manjunath, B.S. and Ma, W.X. ( 1996) Browsing and retrieval of image data. IEEE. Trans. Pattern. Anal. Mach. Intell. , 18, 837– 842. Google Scholar CrossRef Search ADS   6 Verma, M., Balasubrahmanian, R. and Murala, S. ( 2015) Local extrema co-occurrence pattern using color and texture image retrieval. Neurocomputing. , 165, 255– 269. Google Scholar CrossRef Search ADS   7 Srivastava, P., Binh, N.T. and Khare, A. ( 2013) Content-based image retrieval using moment. In Proc. 2nd Int. Conf. Context-Aware Systems and Applications, Phu Quoc, Vietnam, November 25–26, pp. 221–231. Lecture Notes of ICST, Springer. 8 Andreou, I. and Sgouros, N.M. ( 2005) Computing, explaining visualizing shape similarity in content-based image retrieval. Inf. Process. Manage. , 41, 1121– 1139. Google Scholar CrossRef Search ADS   9 Khare, M., Srivastava, R.K. and Khare, A. ( 2015) Moving object segmentation in daubechies complex wavelet domain. Signal, Image Video Process. , 9, 535– 650. Google Scholar CrossRef Search ADS   10 Lin, C. and Lin, W. ( 2011) Image retrieval based on adaptive color histogram and texture features. Comput. J. , 54, 1136– 1147. Google Scholar CrossRef Search ADS   11 Gevers, T. and Smeulders, A.W. ( 2000) Pictoseek: combining color and shape invariant features for image retrieval. IEEE. Trans. Image. Process. , 9, 102– 119. Google Scholar CrossRef Search ADS PubMed  12 Srivastava, P., Binh, N.T. and Khare, A. ( 2014) Content-based image retrieval using moments of local ternary pattern. Mobile Netw. Appl. , 19, 618– 625. Google Scholar CrossRef Search ADS   13 Wang, X., Yu, Y. and Yang, H. ( 2011) An effective image retrieval scheme using color, texture and shape features. Comput. Stand. Interfaces , 33, 59– 68. Google Scholar CrossRef Search ADS   14 Liu, G., Li, Z., Zhang, L. and Xu, Y. ( 2011) Image retrieval based on microstructure descriptor. Pattern. Recognit. , 44, 2123– 2133. Google Scholar CrossRef Search ADS   15 Zhang, M., Zhang, K., Feng, Q., Wang, J., Kong, J. and Lu, Y. ( 2014) A novel image retrieval method based on hybrid information descriptors. J. Vis. Commun. Image Represent. , 25, 1574– 1587. Google Scholar CrossRef Search ADS   16 Liu, G.H., Yang, J.Y. and Li, Z. ( 2015) Content-based image retrieval using computational attention model. Pattern. Recognit. , 48, 2554– 2566. Google Scholar CrossRef Search ADS   17 Flores-Pulido, L., Starostenko, O., Flores-Quéchol, D., Rodrigues-Flores, J.I., Kirschning, I. and Chávez-Aragón, J.A. ( 2008) Content-based Image Retrieval Using Wavelets. In Proc. 2nd WSEAS Int. Conf. Computer Engineering and Applications, pp. 40–45. 18 Moghaddam, H., Abrishami, T., Khajoie, T., Rouhi, A.H. and Tarzjan, M.S. ( 2005) Wavelet correlogram: A new approach for image indexing and retrieval. Pattern. Recognit. , 38, 2506– 2518. Google Scholar CrossRef Search ADS   19 Saadatmand-Tarzjan, M. and Moghaddam, H.A. ( 2007) A novel evolutionary approach for optimizing content-based image indexing algorithms. IEEE Trans. Syst. Man Cybern. B Cybern. , 37, 139– 153. Google Scholar CrossRef Search ADS PubMed  20 Lamard, M., Cazuguel, G., Quellec, G., Bekri, L., Roux, C. and Cochener, B. ( 2007) Content Based Image Retrieval Based on Wavelet Transform Coefficients Distribution. In Proc. 29th IEEE Annual Int. Conf. Engineering in Medicine and Biology Society, August 22–26, pp. 4532–4535. IEEE. 21 Agarwal, M. and Maheshwari, R.P. ( 2012) Á trous gradient structure descriptor for content-based image retrieval. Int. J. Multimed. Inf. Retr. , 1, 129– 138. Google Scholar CrossRef Search ADS   22 Fu, X., Li, Y., Harrison, R. and Belkasim, S. ( 2006) Content-Based Image Retrieval using Gabor-Zernike Features. In Proc. 18th Int. Conf. Pattern Recognition, Hong Kong, August 20–24, 2, pp. 417–420, IEEE. 23 Srivastava, P., Prakash, O. and Khare, A. ( 2014) Content-Based Image Retrieval using Moments of Wavelet Transform. In Proc. Int. Conf. Control Automation and Information Sciences, Gwangju, South Korea, December 02–05, pp. 159–164. 24 Vo, A. and Oraintara, S. ( 2010) A study of relative phase in complex wavelet domain: Property, statistics and applications in texture image retrieval and segmentation. Signal Process. , 25, 28– 46. 25 Starck, J., Candes, E.J. and Donoho, D.L. ( 2002) The curvelet transform for image denoising. IEEE. Trans. Image. Process. , 11( 6), 670– 684. Google Scholar CrossRef Search ADS PubMed  26 Sumana, I.J., Islam, M.M., Zhang, D. and Liu, G. ( 2008) Content-Based Image Retrieval using Curvelet Transform. In Proc. 10th IEEE Workshop on Multimedia Signal Processing, Cairns, Queensland, Australia, October 8–10, pp. 11–16. IEEE. 27 Gonde, A.B., Maheshwari, R.P. and Balasubramanian, R. ( 2013) Modified curvelet transform with vocabulary tree for content based image retrieval. Digit. Signal Process. , 23, 142– 150. Google Scholar CrossRef Search ADS   28 Youssef, S.F. ( 2012) ICTEDCT-CBIR: Integrating curvelet transform with enhanced dominant colors extraction and texture analysis for efficient content-based image retrieval. Comput. Electr. Eng. , 38, 1358– 1376. Google Scholar CrossRef Search ADS   29 Zhang, D., Islam, M.M., Lu, G. and Sumana, I.J. ( 2012) Rotation invariant curvelet features for region based image retrieval. Int. J. Comput. Vis. , 98, 187– 201. Google Scholar CrossRef Search ADS   30 Murtagh, F. and Starck, J.L. ( 2008) Wavelet and curvelet moments for image classification, Application to aggregate mixture grading. Pattern Recognit. Lett. , 29, 1557– 1564. Google Scholar CrossRef Search ADS   31 Sumana, I.J., Lu, G. and Zhang, D. ( 2012) Comparison of Curvelet and Wavelet Texture Features for Content Based Image Retrieval. In Proc. IEEE Int. Conf. Multimedia and Expo (ICME), Melbourne, Australia, July 9–13, pp. 290–295. IEEE. 32 Das, S., Garg, S. and Sahoo, G. ( 2012) Comparison of content based image retrieval systems using wavelet and curvelet transform. Int. J. Multimed. Appl. , 4, 137. 33 Luo, J., Song, D., Xiu, C., Geng, S. and Dong, T. ( 2014) Fingerprint classification combining curvelet transform and gray-level co-occurrence matrix. Math. Probl. Eng. , 2014, 15. 34 Gonzalez, R.C. and Woods, R.E. ( 2002) Digital Image Processing  ( 2nd edn). Prentice Hall Press. 35 Ojala, T., Pietikainen, M. and Maenpaa, T. ( 2002) Multiresolution gray-scale and rotation invariant texture classification with local binary patterns. IEEE. Trans. Pattern. Anal. Mach. Intell. , 24, 971– 987. Google Scholar CrossRef Search ADS   36 Nigam, S. and Khare, A. ( 2015) Multi-resolution approach for multiple human detection using moments and local binary patterns. Multimed. Tools Appl. , 74, 7037– 7062. Google Scholar CrossRef Search ADS   37 Haralick, R.M., Shanmungam, K. and Dinstein, I. ( 1973) Textural features of image classification. IEEE Trans. Syst., Man Cybern. , 3, 610– 621. Google Scholar CrossRef Search ADS   38 Candes, E., Demanet, L., Donoho, D. and Ying, L. ( 2006) Fast discrete curvelet transforms. Multiscale Model. Simul. , 5, 861– 899. Google Scholar CrossRef Search ADS   39 http://wang.ist.psu.edu/docs/related/. Accessed  April 2014 40 Oliva, A. and Torraiba, A. ( 2001) Modeling the shape of the scene; a holistic representation of the spatial envelope. Int. J. Comput. Vis. , 42, 145– 175. Google Scholar CrossRef Search ADS   41 http://www.ci.gxnu.edu.cn/cbir/. Accessed June 2015. 42 Buades, A. and Lisani, J.L. ( 2016) Directional filters for cartoon + texture image decomposition. Image Process. Line , 6, 75– 88. Google Scholar CrossRef Search ADS   43 Starck, J.L., Murtagh, F. and Fadili, J.M. ( 2010) Sparse Image and Signal Processing: Wavelets, Curvelets, Morphological Diversity . Cambridge University Press. Google Scholar CrossRef Search ADS   Author notes Handling editor: Fionn Murtagh © The British Computer Society 2017. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com

Journal

The Computer JournalOxford University Press

Published: Mar 1, 2018

There are no references for this article.

You’re reading a free preview. Subscribe to read the entire article.


DeepDyve is your
personal research library

It’s your single place to instantly
discover and read the research
that matters to you.

Enjoy affordable access to
over 12 million articles from more than
10,000 peer-reviewed journals.

All for just $49/month

Explore the DeepDyve Library

Unlimited reading

Read as many articles as you need. Full articles with original layout, charts and figures. Read online, from anywhere.

Stay up to date

Keep up with your field with Personalized Recommendations and Follow Journals to get automatic updates.

Organize your research

It’s easy to organize your research with our built-in tools.

Your journals are on DeepDyve

Read from thousands of the leading scholarly journals from SpringerNature, Elsevier, Wiley-Blackwell, Oxford University Press and more.

All the latest content is available, no embargo periods.

See the journals in your area

Monthly Plan

  • Read unlimited articles
  • Personalized recommendations
  • No expiration
  • Print 20 pages per month
  • 20% off on PDF purchases
  • Organize your research
  • Get updates on your journals and topic searches

$49/month

Start Free Trial

14-day Free Trial

Best Deal — 39% off

Annual Plan

  • All the features of the Professional Plan, but for 39% off!
  • Billed annually
  • No expiration
  • For the normal price of 10 articles elsewhere, you get one full year of unlimited access to articles.

$588

$360/year

billed annually
Start Free Trial

14-day Free Trial