Date: 24.04.

2025

Final draft of the Paper

Impulse Noise Removal in Images Using Efficient VLSI Architecture- A Comparative Study
A.Evelin Julia Rani.
Department of Electronics and Communication Engineering,
Anand Institute of Higher Technology, Chennai.
India

A R T I C L E I N F O
A B S T R A C T

Keywords:
Image pre-processing is an important operation that is used to redefine an image to
Image de-noising
improve human visual perception and information extraction. To de-noise an image
Noise removal
tainted with impulsive noise, several state-of-the- art methods have been presented.
Impulsive noise
This work examines and compares several de-noising methods based on the median filter
Median filter
Trimmed filter
and its advanced non-linear techniques. The research also focuses on the recommended
Adaptive median imple- mentation methodologies for using deep learning to de-noise impulsive noise.
filter The paper focuses on one approach’s limitations and possible remedies, as well as other
techniques that have been offered. The study also identifies several other difficulties
that have yet to be resolved.

1. Introduction
discontinuities. A non-linear model [2] was designed to solve such
challenges.
Noise is unwelcome data that degrades an image’s image quality.
As a result, non-linear filters became the focus of research [3].
The inconsistent variation of contrasts in an image is caused by noise.
When compared to linear filters, they are superior at handling edges
For future layers of processing, today’s era of great accuracy and
and smoothing images. The greatest example is the median filter [1],
perfection necessitates a very clear image. It is an important area of
which was first proposed by Tukey [4]. The Median Filter, which
research to restore and improve the image quality of a tainted image.
outperforms Linear Filters, has been designed. The median filter
The bulk of researchers in this field are interested in de-noising images
performs wonder- fully in low-density noise, but as the noise density
for better visual perception and information extraction. Image process-
(ND) increases, it stops working. Researchers created various median
ing for meaningful information abstraction is becoming the backbone
filter versions to boost performance. Few work well, and even fewer
of numerous fields, including defence and security, medical
have downsides when it comes to retaining visual details. Some of
diagnostics, astronomical engineering, agriculture, and many more. As
the Median Filter-based algorithms that have been created include
a result, a high-resolution, noise-free image is always preferred.
Filters with a Weighted Median [5], (Switching Median Filter)
However, noise is a component that regularly appears in images. As a
result, de-noising an image to obtain a high-quality image is both (SWMF) [7] Adaptive Median Filter (AMF) [8,9,11–17], Adaptive
necessary and difficult in this sector. Several techniques for de- Switching Median Filter
noising an image have been contributed by several researchers, but (ASMF) [18], Adaptive Switching Median Filter (ASMF) [6,19–25],
there is still room for improve- ment. An image can be contaminated Adaptive Switching Weighted Median Filter (ASWMF) [26], Adaptive
Switching Weighted Median Filter (ASWMF) Adaptive Weighted Mean
for a variety of reasons, and the type of noise that is added to it
Filter [27–31], Alpha Trimmed Mean and Median Filters [32,33], and
deteriorates the image further depending on the noise’s characteristics.
so on. Diffusion-based techniques [34], total variation-based
Gaussian, impulsive, or speckle noise are all possibilities. A
approaches [35], and other image denoising algorithms have been
comparison of numerous de-noising methods for the removal of
developed in recent decades. Wavelet/curvelet-based approaches
impulsive noise from an image is presented in this research. Basically,
[28,29,37–41], and probabilistic based method [42–46,49–53].
linear and non-linear-based methods have been developed for picture A comparative analysis of various de-noising algorithms for impul-
restoration from a corrupted image. The linear model [1] has a lower sive noise removal is offered in this paper. To date, certain
time complexity and works well with increased noise. They stop comparative contributions relating to the evaluation of filtering
working when there is a lot of noise in the image, which causes blur algorithms have been documented [54–58]. This study presents a
and comparative examination of
many filters and their various versions in terms of functionality and
2.3. Switching median filter with boundary discriminative noise detection
relative performance. Grey-scale images are employed to test the algo-
(BDND)
rithms’ adaptability because their application is desired in a variety of
sectors for varied objectives, such as clinical diagnosis, forensics,
This variant of the Median method [8] adapts to extremely polluted
radar, and sonar systems. In comparison to the previously published
photos and mitigates difficulties caused by pixel misjudgement. With the
tech- niques, each algorithm identifies and removes noise in a different
use of a binary decision map, it is designed to identify degraded pixels.
or more advanced manner. The state-of-the-art algorithms stated above
The binary decision map categorizes all pixels into one of three
will be reviewed and contrasted in the next section. Under
groups. Uncorrupted, low-intensity and high-intensity corrupted pixels
performance evaluation, the performance comparison is done both
are the three classes. The window size is determined by the noise
visually and objectively. Some mathematical formulas are used in the
density, ensuring that at least one-third of the pixels are uncorrupted.
objective analysis, which can be a useful tool for estimating image
Only the degraded pixels are processed by BDND, which reduces the
quality. PSNR, MSE, Mean Absolute Error (MAE), and Structural
time complexity. To separate the pixels into three classes, two well-
Similarity Index are the objective measures used to analyze
performance indicators (SSIM). defined specified limits are required. The precision of classes is,
without a doubt, dependant on the accuracy of recognized boundaries.
Increased window size is required to accommodate one-third of
2. Working principle-based analysis of the state-of-art
uncorrupted pixels in the processing window, which causes additional
blurring. This approach calculates the median value using non-
For a successful image processing activity, images with superior
corrupted pixel ele- ments, un-sharpening the edges in the resulting
vi- sual acuity are always needed. When it comes to achieving the
image [9].
desired image quality, de-noising is crucial. One of the most essential
ap- proaches for image de-noising is filtering. Following is a detailed
2.4. Adaptive median filter (AMF)
ex- amination of some developed algorithms:
To date, other variations and better approaches based on SMF have
2.1. Standard median filter (SMF) been documented, such as AMF [10–16], where the size of the consid-
ered window is adjustable. The AMF is used to increase the filter’s
The Standard Median Filter (SMF) [1] is a simple basic position adaptability, allowing it to modify its size as needed based on the esti-
determination filter that was developed as a non-linear filter. It mation of local noise density. It is effective at removing mixed
operates by processing the centre element of the window in question to impulses with a high likelihood of occurrence while maintaining
reduce impulsive noise. The processing pixel may have a value of "0′′ clarity. Because the median filtering is only applied to the
(black), "255′′ (white), or any other value between 0 and 255, after contaminated pixels, this method is faster to implement. The window
which the pixel is replaced by finding the window’s median. The size of an adaptive median
major flaw in this approach is that all pixels are treated the same way, filter is adaptively extended based on the statistics of the min, median,
whether they are affected or not. Edge preservation is also a problem and max values in the current analysis window. It normally starts with
for SMF. Consider the 3 × 3 window in Fig. 1 to better understand the a window of size 3 and grows to the maximum permitted size until the
functioning principle. median value is in the middle of the two extremes. AMF is frequently
For calculations, the centre element is taken into account. As a utilized by researchers to construct advanced algorithms due to its ease
result, the median for 255 in the centre is derived by ascending the of implementation and high noise removal capacity.
values in the window, i.e. (0, 110,120, 180, 180, 255, 255, 255,
255). The median 2.5. Adaptive switching median filter (ASWM)
value will be used to replace the original centre pixel 255, which is
now the centre element of the formatted values, i.e. 180. Every pixel Unlike the switching median filter, the ASWM [18] does not
element is treated in the same way. require an a priori threshold. The threshold is computed locally from
the pixel intensity in a sliding window in this manner. In the current
window, the weighted mean value and weighted standard deviation are
2.2. Switching median filter (SWMF) determined. The weights are the inverse of the distance between the
considered pixel and the pixel elements’ weighted mean value. As a
SWMF [8] is a significant advancement in the detection of noisy result, impulsive noise does not taint the determination of these data,
pixels with high accuracy. It first detects the faulty pixel before using which are used to calculate the Threshold. The weighted mean is
median filtering. The SWMF is a two-step process. Initially, a test de- assessed iteratively in each window. The weighted standard deviation
termines whether a particular pixel is impulsive or not: a pixel is cor- is also calculated, as well as the Threshold.
rupted if the threshold value is less than the absolute difference
determined between the median of the considered window and the 2.6. Adaptive dynamically weighted median filter (ADWMF)
processed pixel’s intensity [72]. If a contaminated pixel is discovered
after testing, a classic median fitter is used for restoration; if the The weighted median filter [5] is combined with a basic impulse
considered pixel is not found to be contaminated, it is not modified. detector approach in ADWMF [29]. The approach adjusts the window
The main disadvantage of the SWMF method is that it substitutes a
size based on the noise density. The weighted median filter is updated
noisy pixel with a nearby median value without taking into account
once impulse noise is detected by dynamically giving zero weight to
local features such as edges. As a result, some picture and edge
the noisy element in the appropriate frame.
information are not retrieved adequately.
2.7. Noise adaptive fuzzy switching median filter (NAFSM)

The detection and filtering phases of NAFSM [22] reduce

impulsive noise. To find the corrupted pixel, the method uses the
histogram of the noisy image. The resultant noisy pixel is then sent to
the filtering stage, which processes the corrupted pixel while leaving
the noise-free pixel elements alone. To find the affected pixel, it looks
for local maxima at either end of the histogram. To identify the noisy
Fig. 1. Considered 3 × 3 window.
2
pixel, a binary noisy masque is constructed. The noisy pixel is given
a value of 0 while the

3
others are given a value of 1. The estimated correction term is used to to replace the pixel in question. If there are no noise-free elements
replace the noisy pixel in the filtering stage, based on the mask’s detected, the mean filter is employed to replace them. APF also does an
marking for noisy pixels.. excellent job of judging and retaining noise-free pixels. If it does not
perform its role as an image applicant, it develops a good relationship
2.8. Modified decision-based un-symmetric trimmed median filter in replacing polluted pixels. The algorithm’s performance degrades as
(MDBUTMF) the noise density rises. When the noise density in an image exceeds
70%, it loses its ability to maintain the image’s characteristics and fine
MDBUTMF [44] was created to solve the problem of all pixels in details. This failure occurs because it employs a mean filtering strategy
the examined window being either 0 or 255. The technique begins by in the absence of noise-free pixels in the affected window.
finding the noisy candidate in an image tainted with impulsive noise.
The associated pixel element is left unmodified if the intensity levels are 3. Performance evaluation based on simulation results
trimmed median. As a result, the majority of pixels in a de-noised
image are unable to estimate the data connected to the original image Under filters based on the median, adaptive median, and advanced
correctly, resulting in image degradation after de-noising. non-linear filter, several simulations are done using the state-of-the-art
methods outlined above. The datasets were obtained from reputable
websites such as imageprocessingplace.com and sipi.usc.edu. The
2.9. Adaptive probability filter (APF)
BSDS300 dataset and images of Lena, Lady, House, Pepper, Mandrill are
used. The photographs have been reduced to 160 by 120 pixels so that
The author used a probabilistic technique to construct a novel
they can be used in all of the state-of-the-art. The median filter tech-
detection module in the APF [48]. To distinguish the impulsive noise,
niques, as well as an enhanced version of it, are implemented in the
two extreme grey levels of the image are used, coupled with the
MATLAB R2013b environment. The suggested method’s performance is
spread of noise. For noise detection, the evaluated processed pixel is
assessed using quantitative measures such as mean square error
compared to the adjacent pixel to determine its suitability as an image
(MSE), peak signal-to-noise error (PSNR), mean absolute error
element. A noise pixel is highlighted if it is detected during the
(MAE), and structural similarity index measure (SSIM) that are
detection step. The processing pixel is the one that has been marked.
formulated as follows:
To replace the pro- cessing pixel, the number of noise-free pixel ∑m,n )2
elements is calculated and compared to the estimated threshold value Zi j — Zd
between the two extreme grey values. MDBUTMF processes pixels with a ,
MSE = i,j i,j (3)
value between the two extremes. m×n

2.10. A new method based on pixel density in salt and pepper noise PSN = (255
2
) (4)
removal (BPDF) R 10log10 MS

∑ ⃒⃒ ⃒⃒
E

The most repeating noise-free pixel is used to get the median value in
m,nZi,j — Zd
i,j i,j
MAE (5)
this method [46]. To develop the algorithm, a range between the two = m × n
extremes is chosen. If a processing pixel is identified to be noisy, the
algorithm looks for at least one pixel within that range and at least one where ‘Z’ denotes the clean image and ‘Zd’ denotes de-noised image. The
pixel beyond that range. Outside the range, the pixel is assumed to be size of the image is m × n.
noise-free. If this is the case, the most repeating pixel intensity is i(2μXμY + C1) + (2σXY +
searched once more. The processing pixel is replaced with the median SSIM(X,
C 2) Y) = 2 (6)
of i(
C2)μ + μ2 + c1) + (σ2 + σ2 +
the repeated pixel. Because the range chosen is superficial, the algorithm X Y X Y

built fails to keep the image’s details.

c1 = (K1L) 2
(7)
2.12. Probabilistic decision based filter (PDBF)
c2 = (K2L)2 (8)
Trimmed Median Filter (TMF) and Patch Else Trimmed Median Filter where, ‘µX’, ‘µY’ denote the average and ‘σX’, ‘σY’ represent the
(PETMF) are fused independently for low and high ND to de-noise a variance of image ‘X’ and ‘Y’, respectively; ‘σXY’ represents the co-
picture contaminated with impulsive noise in the PDBF [47]. When the variance. ‘K1’ and ‘K2’ are the constant. The value of ‘L’ in an 8-bit
ND is less than 50%, TMF is used, but PEITMF is used when the ND grayscale picture is
is larger than 50%. When compared to the great majority of other 255. The standard pictures, BSDS300 dataset, and BBBC041 dataset
state-of-the-art, the PDBF performs wonderfully, but it lacks a are used to simulate algorithms based on median filters and their
powerful arrangement when the NFP is even included when enhanced version, and the results are summarized. Tables 1 and 2
computing the contain all of the obtained results. Figs. 3–5 shows the
corresponding graphical

4
Table 1
additional benefits that can be seen in the findings. In comparison to
Performance comparison of all considered state-of-the-art algorithms in terms of
PSNR using image Lena in various noise densities. other state-of-the-art algorithms, a superior algorithm should have a
lower MSE, a higher SSIM, and a higher PSNR value. As shown in
Algorithms Noise Density
Ta- bles 1 and 2 and Figs. 2 and 3, the algorithms adopting a
10% 30% 50% 70% 90% probabilistic method have a higher PSNR and SSIM value than the
APF 43.55 40.77 36.14 32.64 25.53 previously pro- posed algorithms. Because the algorithms built
BPDF 41.17 38.34 34.17 27.63 22.78 utilizing a probabilistic method have a stronger noise identification
PDBF 42.87 39.89 35.72 31.11 24.15
strategy, they perform better. They devised a reliable method for
MDBUTMF 40.76 34.82 32.98 26.46 21.18
DBA 40.11 34.17 31.68 25.45 20.71
determining whether a pixel contains information from the original
UTMF 36.18 31.85 29.77 24.88 19.48 image or is noise. Trimmed Median Filter (TMF) and Patch Else Trimmed
TMF 28.63 25.85 21.44 17.17 15.66 Median Filter (PETMF) are fused independently for low and high ND to
NAFSM 39.36 34.52 31.88 25.66 20.11 de-noise a picture contaminated with impulsive noise in the PDBF
ADWMF 38.68 34.44 28.68 24.46 21.33
[47]. When the ND is less than 50%, TMF is used, but PEITMF is
ASWM 37.11 33.75 24.45 20.14 18.66
AMF 36.77 33.25 22.11 17.52 16.88 used when the ND is larger than 50%. When compared to the
SWMF-BDND 34.22 32.35 27.84 23.54 20.23 great majority of other state-of-the-art, the PDBF performs
SWMF 29.57 25.82 24.21 17.11 8.12 wonderfully, but it lacks a powerful arrangement when the NFP is
SMF 27.63 24.11 23.47 16.16 7.78 even included when computing the trimmed median. As a result, the
majority of pixels in a de-noised image are unable to estimate the data
connected to the original image correctly, resulting in image
Table 2 degradation after de-noising. As a result, they replace each pixel with
Performance comparison of all considered state-of-the-art algorithm in terms of arbitrary values that have no relationship to neighboring or local pixels.
SSIM using image 13,026 from BSDS300 Dataset. Advanced non-linear filters, such as NAFSM, MDBUTMF, BPDF, PDBF,
Algorithms Noise Density and APF, have a higher de-noising ability. Due to the creation of a good
10% 30% 50% 70% 90%
detection technique, APF outperforms the others. These filters can
provide some edge break and blurring effects in high noise densities. It
APF 0.9526 0.9266 0.8496 0.8145 0.7841
BPDF 0.8687 0.8487 0.8092 0.7787 0.6901
can be seen in Figs. 4 and 5. The APF may be further improved by
PDBF 0.8714 0.8622 0.8133 0.7814 0.6918 utilizing some improved algorithm of median filter instead of the basic
MDBUTMF 0.9014 0.8615 0.8291 0.7912 0.7211 median filter.
DBA 0.8614 0.8461 0.8072 0.7796 0.7011 The medical images most often get intact with impulse noise due to
UTMF 0.8512 0.8321 0.7921 0.7632 0.6922
improper imaging. The APF filter in its present form can be utilized in
TMF 0.8141 0.7955 0.7628 0.6836 0.5814
NAFSM 0.9289 0.8711 0.8116 0.7754 0.6722
the medical field for enhancing the image quality for further diagnosis.
ADWMF 0.9255 0.8689 0.7869 0.6933 0.6741
ASWM 0.9235 0.8424 0.7087 0.6478 0.6187 4. Conclusion
AMF 0.9155 0.8387 0.6926 0.6087 0.5974
SWMF-BDND 0.9089 0.8231 0.7769 0.7034 0.6543
This work gives an examination based on a comparison of the
SWMF 0.8098 0.7326 0.7011 0.6339 0.4428
SMF 0.7787 0.7002 0.6813 0.6241 0.4316 approach and performance of a state-of-the-art algorithm for removing
impulsive noise from a grayscale image that was developed. Extensive
simulations are run, and the results are tabulated as well as a graphical
representations for a better understanding. The PSNR value continu- depiction. For visual analysis, the obtained denoised photos are also
ously increases with the incorporation and development of advanced displayed. The benefits, drawbacks, and solutions to such drawbacks
approaches based on the median filter, as seen in Table 1 and Fig. 2. by other established algorithms are discussed. Several variants were
Furthermore, algorithms that use noise detection techniques offer created by combining the adaptive filtering approach with the median
to alter the window size. This method has gained popularity, however,
the main issue of removal in high noise density, edge preservation, and
blurring remains.

Fig. 2. Graphical comparison of all considered state-of-the-art in terms of PSNR

using image Lena.
Fig. 3. Graphical comparison of all considered state-of-the-art in terms of SSIM
using image 13,026 from BSDS300 Dataset.
5
Fig. 4. Visual Comparison of the obtained images using considered algorithms: (a) Original image, (b) Noisy image with 50% ND, (c) SMF, (d) SWMF, (e) SWMF-
BDND, (f) AMF, (g) ASWF, (h) ADWMF, (i) TMF, (j) NAFSM, (k) UTMF, (l) DBA, (m) MDBUTMF, (n) PDBF, (o) BPDF, and (p) APF.

Trimmed mean and median were also established, which

previous sections, it can be enhanced further to create a more robust
eliminated the problem of the average being pulled to an arbitrary
algorithm.
figure. This development was a significant step toward resolving the
blurring problem. Researchers began to utilize a probabilistic
technique in conjunction with median variants. Both PDBF and APF Ethics approval
perform well in terms of denoising, however, PDBF has trouble with
This article’s work has been tested on the already available data in
the issue of an even number of noise-free pixels, which causes images
the research community.
to blur at high noise density. Along with good detection, APF uses a
mean filter that fails to keep the original information at high noise
density. As a result, a robust approach to address the difficulties in
PDBF and APF can be devised in the future. Because this filter has
some of the constraints stated in the

6
Fig. 5. Visual Comparison of the obtained Lena images using considered algorithms: (a) Original image, (b) Noisy image with 50% ND, (c) SMF, (d) SWMF, (e)
SWMF-BDND, (f) AMF, (g) ASWF, (h) ADWMF, (i) TMF, (j) NAFSM, (k) UTMF, (l) DBA, (m) MDBUTMF, (n) PDBF.

References

[1] R. Gonzalez and R. Woods, Digital Image Processing, 2nd ed. Prentice Hall. G. Thanakumar, Absolute difference based progressive switching median filter for
[2] A. Gellert, R. Brad, Context-based prediction filtering of impulse noise images, IET efficient impulse noise removal, Int. J. Eng. Trends Technol. 41 (5) (2016)
Image Process. 10 (6) (2016) 429–437.
[3] E.A.D. Poularikas, “Nonlinear digital filtering,” Handb. Formulas Tables SignProcess.,
1999.
[4] J.W. Tukey, Exploratory Data Analysis, Addison-Wesley Publishing Company,
1977.
[5] L. Yin, R. Yang, M. Gabbouj, S. Member, Weighted median filters : a tutorial, IEEE
Trans. Acoust. Speech, Signal Process. 43 (6) (1996) 157–192. Trans. Circuits Syst.
Analog Digit. Signal Process.
[6] S. Zhang, M.A. Karim, A new impulse detector for switching median filters, IEEE
Signal Process. Lett. 9 (11) (2002) 360–363.
[7] Z. Wang, D. Zhang, Progressive switching median filter for the removal of impulse
noise from highly corrupted images, IEEE Trans. Circuits Syst. Analog Digit. Signal
Process. 46 (1) (1999) 78–80.
[8] P.E. Ng, K. Ma, Discriminative noise detection for extremely, IEEE Trans.
ImageProcess. 15 (6) (2006) 1506–1516.
7
8
9