Journal of Real-Time Image Processing
ORIGINAL RESEARCH PAPER
Real‑time impulse noise removal
· Cem Kalyoncu
Received: 15 December 2017 / Accepted: 16 May 2018
© Springer-Verlag GmbH Germany, part of Springer Nature 2018
An adaptive interpolation-based impulse noise removal (AIBINR) algorithm is proposed to remove impulse noise from
color and gray-scale images in real time. AIBINR works fast and has no need for parameter tuning to remove ﬁxed-valued
impulse noise. A GPU application has been developed to demonstrate the speed and inherent parallelization capabilities of
the proposed method. Using the high-speed implementation, we have shown that AIBINR can denoise color images fast
enough to be used in real-time video denoising, while having comparable denoising performance when compared to the
state-of-the-art methods without any modiﬁcation to its parameters.
Keywords Impulse noise removal · Cuda · Real-time denoising · High deﬁnition
Hardware and communication faults, or faulty memory loca-
tions create impulse noise in pictures . Impulse noise
is a problem for image-processing applications. Moreover,
images containing noise may cause performance reduction
in computer vision applications such as motion detection
[5, 11, 12]. Most of these tasks require real-time denois-
ing, which has been studied extensively for many diﬀerent
noise types [1, 10, 14, 17, 23, 32]. On the other hand, real-
time impulse noise removal has gained minimal attention.
However, as a common noise type, real-time denoising of
impulse noise should be desired.
The real-time systems receive data, process them,
and return the output quick enough to aﬀect the environ-
ment . In other words, real-time systems produce correct
output within the deﬁned time period .
Median ﬁlter is commonly used for removing impulse
noise, however, it is not suitable for high noise densities.
This has led researchers to focus on developing median-
based ﬁlters [20, 21, 26, 30, 33–35, 37]. There are many
techniques for removing high-density impulse noise such as
adaptive median ﬁlter (AMF) , cloud model ﬁlter (CMF)
, modiﬁed decision-based unsymmetric trimmed median
filter (MDBUTMF) , boundary discriminative noise
detection (BDND)  , improved boundary discriminative
noise detection (IBDND) , interpolation-based impulse
noise removal (IBINR)  and unbiased weighted mean
ﬁlter (UWMF) .
Ng and Ma proposed a BDND algorithm that performs
eﬃcient noise removal up to 80% . This method uses
switching ﬁlter strategy and ﬁnds noisy pixels using the
local statistics. In , authors propose a modiﬁcation on
the BDND algorithm to change expansion of window size
in ﬁltering step. IBDND applies 21x21 ﬁlter in the detec-
tion stage that has costly operations on this large window.
This method applies adaptive weighted mean ﬁlter in the
A method is proposed by Zhou  uses Cloud Model
 to detect noisy pixels. This method uses randomness
and fuzziness that exist in impulse noise. Cloud Model
identiﬁes pixels as good or corrupted pixel candidates and
applies weighted fuzzy mean ﬁlter (WFM) to the pixels
which are classiﬁed as noisy. Cloud model is successful to
ﬁnd noisy pixels with the classiﬁcation rate more than 99.9%
at higher noise densities.
Modiﬁed decision unsymmetric trimmed median ﬁlter
(MDBUTF)  detects non-noisy pixels in the ﬁltering win-
dow and takes the trimmed median value as the new pixel
value. If all pixels are detected as noise; mean value is used
as the new center pixel. In , adaptive iterative fuzzy ﬁlter
(AIFF) is proposed. Detection of noisy pixels is done by an
* Cem Kalyoncu
Department of Computer Engineering, European University
of Lefke, Lefke, Northern Cyprus, TR-10 Mersin, Turkey