International Journal of Advanced Technology and Engineering Exploration, Vol 5(43)

ISSN (Print): 2394-5443 ISSN (Online): 2394-7454

Review Article
http://dx.doi.org/10.19101/IJATEE.2018.543008

A survey on impulse noise removal techniques in image processing

Baby Victoria L.1* and Sathappan S.2
Research Scholar, Erode Arts and Science College, Erode1
Associate Professor, Erode Arts and Science College, Erode2

©2018 ACCENTS

Abstract
In image processing, an essential and most challenging process is removing the noise from the color images. Images are
often corrupted by impulse noise during image acquisition and transmission. Therefore, impulse noise reduction is the
most crucial aspect during image transmission. Over the past decades, several approaches have been proposed for
removing the impulse noise from the images in such a way that the most significant information of the images is
preserved. Hence, the original image quality can be restored efficiently. This paper presents a detailed survey of impulse
noise removal techniques. Initially, different techniques are analysed and its limitations are addressed. Moreover,
performance of all techniques was compared to identify their effectiveness for further improvement on impulse noise
removal techniques. Finally, some future contributions are also provided to improve the impulse noise removal
techniques significantly.

Keywords
Image processing, Impulse noise, Noise removal, Image restoration.

1.Introduction There is a tradeoff between preservation and noise

Any unwanted pixels in the original images are removal. It is more effective in terms of removing
known as noises. Noise is mostly classified into blur impulse noise and preserving the edges and fine
noise and impulse noise. Images are often corrupted information of digital images. Therefore, the primary
by different noises during its acquisition or step before processing the image is the restoration of
transmission [1]. The most well-known noise is the the image by removing noises in the images. The
fat tailed distribution or impulse noise or spike noise. main objective of noise removal technique is
Impulse noise occurs as a result of malfunction of suppressing the noise. The filter can be applied
detector pixels in a digital camera or from missing successfully for reducing heavy noise.
memory components in imaging hardware. There are
two types of impulse noise such as salt-and-pepper For preserving the information, the noise is
noise and random valued noise [2]. The salt-and- eliminated gradually. Such noise removal techniques
pepper noise matches an extreme dynamic range of a are applied depending upon the camera sensors and
pixel value. In this case, noisy pixels can be quite the environmental interferences. When noise is non-
easily detected by an Adaptive Median Filter (AMF). additive, linear filtering techniques are not effective
An image containing salt-and-pepper noise can have in removing the impulse noise. This has led to the
dark pixels i.e., pepper in bright regions and bright utilization of non-linear filter techniques [4]. Over
pixels i.e., salt in dark regions. Random-Valued the past decades, different techniques were applied
Impulse Noise (RVIN) occurs within the dynamic for noise removal process. However, still several
range of an image pixel and it cannot be easily challenges and issues are observed in removing the
detected by AMF. impulse noise from the color images.

Number Nowadays, impulse noise removal is an Hence, the main objective of this article is studying
active research area in image processing. A better the detailed information on different impulse noise
noise filter is required for satisfying two criteria such removal techniques and their drawbacks to further
as noise suppression and useful information improve the noise removal process. Then, based on
preserving in the image with low computational the observed limitations, further improvements on
complexity [3]. impulse noise removal techniques are suggested.

*Author for correspondence

160
 International Journal of Advanced Technology and Engineering Exploration, Vol 5(43)

2.Survey on impulse noise removal hyper-spectral images and also deal with the spatial
techniques redundancy and spectral correlation of such images
A novel and unique principle of adaptive dual by using individual dictionaries. Initially, the issues
threshold median filter [5] was proposed for detecting of denoising hyper-spectral images while they were
a random valued impulse noise. In this method, the corrupted by impulse noise were investigated. To
thresholds were computed by averaging process and a overcome such issues, BCS was proposed such that it
brilliant quality of recovered image was provided by empirically learns the spatial and spectral sparsifying
using simple Median (MED) filter which removes dictionaries during image denoising.
noise from the image. A weighted couple sparse
representation model [6] was proposed for removing Localised rank-ordered differences vector filter [10]
impulse noise. In this approach, the complicated was proposed for suppression of high-density
relationships between the reconstructed and the noisy impulse noise in color images. This method was
images were exploited for making the coding proposed based on an impulsive noise detector for
coefficients more suitable in order to recover the greyscale images that has been adapted in a localized
noise-free image. Additionally, the image pixels were fashion using geometric information for processing
classified into clear, slightly corrupted and heavily color images. According to this statistic, a filtering
corrupted. Then, various data fidelity regularizations between the identity and a non-linear vector filter
were applied to various pixels for further improving was proposed. By using this method, each pixel was
the denoising performance. Here, the dictionary was checked for whether it was noisy or not based on the
directly trained on the noisy raw data through fixed window mask size for all ratio noise intensities.
addressing a weighted rank-one minimization A finite number of sub-windows were defined which
problem. are utilized for removing impulsive noise while
preserving fine information. Once a corrupted pixel
A low-rank prior in small oriented noise-free image was detected, a non-linear filtering was performed.
patches was introduced for removing impulse noise
[7]. A low-rank matrix approximation was significant A two stage quaternion switching vector filter [11]
for preserving the texture information in the was proposed for removing impulse noise in the color
optimally oriented path by considering an oriented images. Initially, an effective color distance measure
patch as a matrix. According to this prior, a single- method was proposed by using quaternion
patch method was proposed within a generalized joint representation. Then, the directional samples along
low-rank and sparse matrix recovery method for with four directions were utilized by this filter
simultaneously detecting and removing nonpointwise according to the new color distance measure for
random-valued impulse noise. A weighting matrix classifying the image pixels into possible noisy and
was integrated for encoding an initial estimate of the noise-free pixels. For possible noisy pixels, the idea
spatial noise distributions. An accelerated proximal of peer group was modified and extended to the
gradient method was adopted for estimating the directional samples for further detecting whether they
optimal noise-free image patches. were corrupted by impulse noise or not. Finally, a
weighted vector median filter was performed only on
A novel denoising scheme was proposed [8] in which the pixels that were identified as noisy.
all noisy pixels were mutually restored through non-
uniform sampling and supervised piecewise A combination of adaptive vector median filter
autoregressive modelling based super-resolution. In (VMF) and weighted mean filter was proposed [12]
this framework, the noisy pixels were equally to remove high-density impulse noise from color
estimated in groups by solving a well-designed images. In this approach, the noisy and non-noisy
optimization issue in which image structure feature pixels were classified according to the non-causal
was assumed as a significant limitation. Another linear prediction error. The adaptive VMF was
objective was that piecewise autoregressive model processed over the noisy pixel where the window size
was not only adopted but also designed therefore all was adapted based on the availability of high-quality
noise-free pixels can be utilized for supervising the pixels. On the other hand, a non-noisy pixel was
model training and optimization issue solving for substituted with the weighted mean of high-quality
achieving higher accuracy. pixels of the processing window. A new edge
preserving contextual model based image restoration
A blind compressed sensing (BCS) technique was technique [13] was proposed for images affected by
proposed [9] to eliminate sparse impulse noise from impulse noise. This method consists of two processes

161
Victoria L and Sathappan S

such as identification of noisy pixels and restoration This section illustrates an overview of merits and
of the original image. Initially, an absolute directional demerits of different impulse noise removal
difference of the neighborhood pixels was followed techniques whose functional scenarios are discussed
for identifying the pixels those were affected by in above section. Through the literature survey on
impulse noise. Also, an edge preserving contextual impulse noise removal techniques, the following
model was proposed for restoring the identified noisy limitations are observed.
pixels. Here, the parameters of the contextual  When number of uncorrupted pixels was zero, the
window were obtained by using a Gaussian kernel. size of filtering window was considered as
This approach was depending on the context model inadequate by using improved boundary
of the noise-free pixels in the selected window. discriminative noise detection (BDND) algorithm.
 In some cases, noise-free pixels were detected as
A fast non-locally centralized sparse representation corrupted pixels and the quality of performance
(FNCSR) algorithm [14] was proposed for image was image dependent.
denoising. In this algorithm, a dictionary for each  Computational complexity and time edge
image was obtained by using dictionary learning preserving method must be maintained at lower
approach such as k-means and principal component level for preserving the edge information and
analysis (PCA). After that, Peak Signal-to-Noise removing the noise at high density levels.
Ratio (PSNR) index was applied for assessing the  Weighted couple sparse representation model
image quality of the reconstructed images based on requires additional non-local self-similarity priors
such dictionaries. Moreover, quality-aware features to preserve more information of image and texture
and support vector machine (SVM) were employed which improves the output image quality.
for constructing a fast noise level estimator (NLE) to  The original patches were over-smoothened by
estimate the noise level from a single noisy image. low-rank prior in small oriented noise-free image
According to the estimate noise level, the parameters patches and only low level information in the
such as search window and search step were selected images was considered.
automatically.  The computation complexity of non-uniform
sampling and autoregressive modelling, two stage
Sparse and low-rank decomposition of Hankel quaternion switching vector filter and
structured matrix [15] was proposed for impulse CAVMFWMF was high.
noise removal. This approach was proposed based on
 Reconstruction error of BCS method was high. As
the annihilating filter-based low-rank Hankel matrix
well, noisy pixels were considered for preserving
(ALOHA). Thus, it was known as robust ALOHA
the original images using LRODF method.
according to the observation that an image corrupted
 Computational feasibility of FNCSR algorithm
with the impulse noise can be modeled as sparse
was limited.
components whereas underlying image can be
 Moreover, sparse and low-rank decomposition of
modeled using a low-rank Hankel structured matrix.
Hankel structured matrix requires an automatic
The sparse and low-rank matrix decomposition
optimal patch size selection method for filtering
problem was solved by alternating direction
process.
technique of multiplier approach including initial
factorized matrices coming from a low-rank matrix-
From the following Table 1, the most challenging
fitting algorithm. This algorithm was applied in a
issues in impulse noise removal techniques are
patch-by-patch fashion for adapting the local image
observed and an ideal solution is identified to
statistics that have different spectral distributions.
overcome those issues in image processing.
3.Discussions
Table 1 Comparison of different impulse noise removal techniques (Considering lena image)
Ref. Methods Merits Demerits Performance metrics
No
[5] Adaptive dual threshold Better PSNR. Quality of performance was (Noise Density=10%)
median filter image dependent. PSNR=41.02dB
[6] Weighted couple sparse Better performance in noise Requires an integration of non- (Noise Density=30%):
representation model reduction. local self-similarity priors for PSNR=34.26dB,
preserving more image and Structural Similarity
texture information to obtain Index Metric
the higher quality outputs. (SSIM)=0.9442
162
 International Journal of Advanced Technology and Engineering Exploration, Vol 5(43)

Ref. Methods Merits Demerits Performance metrics

No
[7] Low-rank prior in small Better performance in noise Inevitably over-smooth the (Mixed sizes from 1×1
oriented noise-free removal. original patches and it does not to 3×3):
image patches for consider any high level PSNR=30.58dB
nonpointwise impulse information in the images.
noise removal
[8] Non-uniform sampling High restoration accuracy Computational complexity was (Noise Density=70%):
and autoregressive and efficiently removing the very high. PSNR=32.08dB,
modelling based super- detected impulse noises SSIM=0.9744
resolution for high when image edges and CPU time=4.3sec
quality impulse noise information were preserved.
removal
[9] BCS technique based It can be used for Reconstruction error was high. WDC mall image
impulse denoising reconstructing hyperspectral (Noise Density=10%):
images from compressive PSNR=55.24dB
measurements in the
presence of impulse noise.
[10] LRODF based high- Robust to noise, easy to A group of noisy pixels were (Noise Density=10%):
quality impulse noise implement. assumed as information to PSNR=38.77dB
removal preserve during removal of Normalized Color
impulse noise. Difference
(NCD)=0.46×102,
Color Multiscale SSIM
(CMSSIM)=0.999
[11] Two stage quaternion It can detect noisy pixels on Computational complexity was (Noise Density=10%):
switching vector filter gradient edges. high. PSNR=34.16dB,
for impulse noise Mean Absolute Error
removal (MAE)=1.61,
NCD=0.0143,
Feature Similarity Index
Metric (FSIM)=0.9919
[12] CAVMFWMF for high- Better performance for both Computational complexity (Noise Density=10%):
density impulse noise low and high-density of during detection process was PSNR=42.27dB,
removal impulse noise. high. SSIM=0.883,
FSIM=0.9983,
CMSSIM=0.9988
[13] Edge preserving Better denoising ability. Computational time was high. (Noise Density=40%)
contextual model PSNR=34.91,
Mean SSIM=0.93,
Non-Shifted Edge Ratio
(NSER)=0.61,
Correlation Factor
(CF)=0.996,
Computational
time=275.09seconds
[14] FNCSR algorithm High computational Computational feasibility was (Noise level=5)
efficiency. limited. Running
time=35.92seconds,
PSNR=38.55dB,
SSIM=0.9433
[15] Sparse and low-rank High robust to noise. It requires an automatic (Noise Density=10%):
decomposition of Hankel selection of optimal patch size PSNR=38.91dB
structured matrix for for filtering process.
impulse noise removal

4.Conclusion and future work presented. It is obvious all researchers have tried in
In this article, a detailed survey on impulse noise different techniques for removing the impulse noise
removal techniques in image processing was from color images to achieve better results than the
other filtering-based noise removal techniques. Based
163
Victoria L and Sathappan S

on the analysis, robust ALOHA has better [10] Roig B, Estruch VD. Localised rank-ordered
performance than the other impulse noise removal differences vector filter for suppression of high-
techniques. This technique was derived based on the density impulse noise in colour images. IET Image
consideration that the spectrum of a noiseless image Processing. 2016; 10(1):24-33.
patch is sparse in Fourier domain. Moreover in the [11] Jin L, Zhu Z, Xu X, Li X. Two-stage quaternion
switching vector filter for color impulse noise
future contribution, this technique can be further
removal. Signal Processing. 2016; 128:171-85.
improved by automatically selecting an optimal patch
[12] Roy A, Singha J, Manam L, Laskar RH. Combination
size and reducing the computational complexity. of adaptive vector median filter and weighted mean
Additionally, it can be extended for image patches filter for removal of high-density impulse noise from
which are sparse in the other transform domains i.e., colour images. IET Image Processing. 2017;
wavelet, frequency, etc. 11(6):352-61.
[13] Veerakumar T, Subudhi BN, Esakkirajan S, Pradhan
Acknowledgment PK. Context model based edge preservation filter for
None. impulse noise removal. Expert Systems with
Applications. 2017; 88:29-44.
Conflicts of interest [14] Xu S, Yang X, Jiang S. A fast nonlocally centralized
The authors have no conflicts of interest to declare. sparse representation algorithm for image denoising.
Signal Processing. 2017; 131:99-112.
References [15] Jin KH, Ye JC. Sparse and low-rank decomposition of
[1] Davis RR, Clavier O. Impulsive noise: a brief review. a hankel structured matrix for impulse noise removal.
Hearing Research. 2017; 349:34-6. IEEE Transactions on Image Processing. 2018;
[2] Koli M, Balaji S. Literature survey on impulse noise 27(3):1448-61.
reduction. Signal & Image Processing. 2013; 4(5):75-
95. Baby Victoria received the MCA
[3] Suganthi A, Senthilmurugan M. Comparative study of degree from Bharathidasan University,
various impulse noise reduction techniques. India in the year 2003 and MPhil
International Journal of Engineering Research and degree in Computer Science from
Application. 2013; 3(5):1302-6. Bharathiyar University, India in the
year 2008 respectively. Currently, She
[4] Pritamdas K, Singh KM, Singh LL. A summary on
is an Assistant Professor of Computer
various impulse noise removal techniques.
Science, Erode Arts and Science
International Journal of Science and Research. 2017;
College, Erode, affiliated to Bharathiyar University. She
6(3):941-54.
has a total experience of over 10 years. She has published 4
[5] Gupta V, Chaurasia V, Shandilya M. Random-valued papers in Conferences and 2 papers in Journals. Her area of
impulse noise removal using adaptive dual threshold interest image processing.
median filter. Journal of Visual Communication and Email: victoriamca456@gmail.com
Image Representation. 2015; 26:296-304.
[6] Chen CL, Liu L, Chen L, Tang YY, Zhou Y.
Weighted couple sparse representation with classified
Sathappan received the MSc degree in
regularization for impulse noise removal. IEEE
Applied Mathematics from
Transactions on Image Processing. 2015; 24(11):4014-
Bharathidasan University, India in the
26.
year 1984, the MPhil and PhD degrees
[7] Wang R, Pakleppa M, Trucco E. Low-rank prior in in computer science from Bharathiyar
single patches for nonpointwise impulse noise University, India in the year 1996 and
removal. IEEE Transactions on Image Processing. 2012 respectively. Currently, He is an
2015; 24(5):1485-96. Associate Professor of Computer
[8] Wang X, Shi G, Zhang P, Wu J, Li F, Wang Y, et al. Science, Erode Arts and Science College, Erode, affiliated
High quality impulse noise removal via non-uniform to Bharathiyar University. He has been a supervisor for
sampling and autoregressive modelling based super- several students of MCA/MPhil., programs over the past
resolution. IET Image Processing. 2016; 10(4):304-13. several years. He has a total experience of over 32 years.
[9] Majumdar A, Ansari N, Aggarwal H, Biyani P. He has published 32 papers in Conferences and 25 papers
Impulse denoising for hyper-spectral images: a blind in Journal. He is also a chairman of Board of QP setters in
compressed sensing approach. Signal Processing. various universities. His areas of interest include computer
2016; 119:136-41. simulation and image processing.

164