Machine Vision and Applications (2018) 29:415–432
Introducing spectral moment features in analyzing the SpecTex
hyperspectral texture database
Received: 9 May 2017 / Accepted: 2 November 2017 / Published online: 29 November 2017
© Springer-Verlag GmbH Germany, part of Springer Nature 2017
Hyperspectral imaging provides more information than conventional RGB images. However, its high dimensionality prevents
its adaptation to the existing image processing techniques. Deﬁning full-band spectral feature is the ﬁrst missing step, which
is currently dealt with indirectly by band selection or dimension reduction. This article proposes a spectral feature extraction
method using the mathematical moments to quantify the shape of the reﬂectance spectrum from different aspects. A whole
family of features is presented by changing the moment attributes. All the features and their combinations are extensively
tested in texture analysis of a new hyperspectral image database from textile samples (SpecTex). Two supervised experiments
are performed: image patch classiﬁcation and pixel-wise mosaic image segmentation. The proposed features are compared to
four other features: the grayscale intensity, the RGB and CIELab values, and the principal components. Also, three analysis
methods are tested: co-occurrence matrix, Gabor ﬁlter bank, and local binary pattern. In all cases, the moment features
outperformed the opponents. Notably, combining the moment features with complementary attributes remarkably improved
the performance. The most discriminative combinations are studied and formulated in this article.
Keywords Feature extraction · Moments · Hyperspectral imaging · Texture analysis · Image database
Spectral imaging has gained signiﬁcant applications in
remote sensing [1,2], cultural heritage , face recognition
, medical diagnosis , and color management and print-
ing . However, an impediment to its widespread employ-
ment is the lack of feature extraction methods. Grayscale
image processing techniques rely on the order that the scalar
pixel intensity naturally exerts on the feature space. Similar
order does not exist in a spectral vector space. Extending
grayscale techniques to RGB images by introducing heuris-
tic ordering to the RGB color space has been studied [7–11].
However, for spectral vector spaces with high dimensions,
these methods fail either theoretically or computationally.
Current methods deal with the problem indirectly in three
ways. The ﬁrst way is to apply a dimension reduction step
This study was funded by the Finnish Funding Agency for Innovation
(TEKES), funding decision 3268/31/2015.
School of Computing, University of Eastern Finland, P.O.
Box 111, 80101 Joensuu, Finland
before the feature extraction and project the data to a low-
dimensional subspace [12,13]. Second is to quantize the
high-dimensional space and identify a set of endmembers
to represent the original intractable manifold and order the
endmembers based on a criterion [9,14]. The third method is
band selection, where one or a few discriminating bands are
extracted and the rest are discarded [8,15,16]. Methods that
directly handle the original full bands are missing.
In this work, novel spectral features based on the mathe-
matical moments of the spectra are proposed. Mathematical
moments are long known for invariant feature extraction
from grayscale images . In that application, moments
express the spatial variation of the neighboring pixels’
intensities, whereas here, they describe the variations in
a pixel’s spectrum as a function of wavelength, i.e., the
shape of the spectrum. In this article, two types of moments
are deﬁned. First one describes the spectrum as a func-
tion of wavelength, and the second as the probability
function of reﬂectance factors independent of wavelength.
These moments are deﬁned for both the original spec-
trum and the L
-normalized version. Also, raw and central
moments, and the moment’s order are studied. Based on these