2010 Second International Conference on Computer Modeling and Simulation

2010 Second International Conference on Computer Modeling and Simulation

A NEW AND EFFICIENT ALGORITHM FOR THE REMOVAL OF HIGH

DENSITY SALT AND PEPPER NOISE IN IMAGES AND VIDEOS

K.Aiswarya V.Jayaraj D.Ebenezer

Dept of ECE Lecturer, Dept of ECE Professor, Dept of ECE
Sri Krishna College of Engg & Tech Sri Krishna College of Engg & Tech Sri Krishna College of Engg & Tech
Coimbatore, India Kuniamuthur, Coimbatore-641008, Kuniamuthur, Coimbatore-641008,
aiswaryak@in.com Tamil Nad u, India Tamil Nadu, India

Abstract— A new and efficient algorith m for high- density salt the effective removal of impulse often leads
Consequently,
and pepper noise removal in images and videos is proposed. to images with blurred and distorted features. Ideally, the
The existing non-linear filter like Standard Median filtering Filter should be applied only to corrupted pixels while
leaving uncorrupted p ixels intact. Applying median filter
(SMF), Adaptive Median Filter (AMF), Decision Based unconditionally across the entire image as practiced in the
Algorithm (DBA) and Robust Estimation Algorithmconventional (RE A) schemes would inevitably alter the intensities
shows better results at low and mediu m noise densities. and remove the signal details of uncorrupted pixels.
At high
Therefore, a noise-detection process to discriminate between
noise densities, their performance is poor. A new algorithm uncorruptedto pixels and the corrupted pixels prior to applying
remove high-density salt and pepper noise using modified nonlinear filtering is highly desirable
Adaptive Median is a “decision-based” or “switching”
shee r sorting method is proposed. The new algorithm has
filter that first identifies possible noisy pixels and then
lower computation time when compared to other standard replaces them using the median filter or its variants, while
algorith ms. Results of the algorithm is compared with various
leaving all other pixels unchanged. T his filter is good at
existing algorithms and it is proved that the new method has detecting noise even at a high noise level. The adaptive
Keywords: Median filter, Midpoint filter, trimmed filter, structure of this filter ensures that most of the impulse noises
better visual appe arance and quantitative me asures
shear sort. are at higher
detected even at a high noise level provided that the
noise densities as high as 90%. window size is large enough. The performance of AMF is
good at lower noise density levels, due to the fact that there
are only fewer corrupted pixels that are replaced by the
I. INTRODUCTION median values [6], [7], [8]. At higher noise densities, the
In the transmission of images over channels, images are number of replacements of corrupted pixel increases
corrupted by salt and pepper noise, due to faulty considerably; increasing window size will provide better
communications[l], [2]. Salt and Pepper noise is also referred noise removal performance; however, the corrupted pixel
to as Impulse noise. The objective of filtering is to remove values and replaced median pixel values are less correlated.
the impulses so that the noise free image is fully recovered As a consequence, the edges are smeared significantly
with minimum signal distortion [3]. Noise removal can be Recently, The DBA processes the corrupted image by
achieved, by using a number of existing linear filtering first detecting the impulse noise. The detection of noisy and
techniques which are popular because of their mathematical noise-free pixels is decided by checking whether the value of
simplicity and the existence of the unifying linear system a processed pixel element lies between the maximum and
theory. A major class of linear filters minimizes the Mean minimum values that occur inside the selected window [9],
Squared Error (MSE) criterion and it provides optimum [10]. This is because the impulse noise pixels can take the
performance among all classes of filters if the noise is maximum and minimum values in the dynamic range (0,
add itive and Gaussian. 255). If the value of the pixel processed is within the range,
The best-known and most widely used non-linear digital then it is an uncorrupted pixel and left unchanged. If the
filters, based on order statistics are median filters. Median value does not lie within this range, then it is a no isy pixel
filters are known for their capability to remove impulse noise and is replaced by the median value of the window or by its
without damaging the edges. Median filters are known for neighborhood values. At higher noise densities, the median
their capability to remove impulse noise as well as preserve value may also be a noisy pixel in which case neighborhood
the edges. The main drawback of a standard median filter pixels are used for replacement; this provides higher
(SMF) [4],[5] is that it is effective only for low noise correlation between the corrupted pixel and neighborhood
densities. At high noise densities, SMFs often exhibit pixel. Higher correlation gives rise to better edge
blurring for large window sizes and insufficient noise preservation. In addition, the DBA uses simple fixed length
suppression for small window sizes. However, most of the window of size 3X3, and hence, it requires significantly
median filters operate uniformly across the image and thus lower processing time compared with AMF and other
tend to modify both noise and noise-free pixels. algorithms. The main drawback of decision-based algorithm
is that streaking occurs at higher noise densities due to

978-0-7695-3941-6/10 $26.00 © 2010 IEEE 400

409
DOI 10.1109/ICCMS.2010.310
replacement with the neighborhood pixel values. Hence,
details and edges are not recovered satisfactorily, especially
when the noise level is high.
II. SORTING ALGORITHM
A. Shear Sorting Algorithm
Sorting is the most important operation used to find the
median of a window. There are various sorting
algorithms such as binary sort, bubble sort, merge sort,
quick sort etc. In the proposed algorithm, shear sorting
technique is used since it is based on parallel architecture.
In practice the parallel architectures help to reduce the
number of logic cells required for its implementation.
The illustration of shear sorting is shown in Figure. 1- 4.
In the odd phases (1,3,5) even rows are sorted in
descending order and rows are sorted out in ascending
order. In the even phases columns are sorted out
independ ently in ascending order.
Figuer.1.Original matrix before sorting
Figure.2.Step 1- Row sorting
Figure.3.Step 3- Column sorting
401
410
Figure.4. Step 4- Row sorting
Modified Shear Sorting Algorithm:
In order to improve the computational efficiency shear
sorting algorithm is modified as follows:
i)All the three rows of the window are arranged in
ascending order.
ii)Then all the columns are arranged in ascending order.
iii)The right diagonal of the window is now arranged in
ascending order.
In this case, the first element of window is the minimum
value, last element o f the window is the maximum value and
middle element of window is the median value. After the
third stage of sorting itself the median value is obtained
even though all the elements are not arranged in ascending
order. The illustration o f sorting algorithm is shown in the
table containing the comparison of the various sorting
techniques.
Table I
Comparison of the various sorting techniques
No of comparisons
Sorting techniques required to compute
median 3x3 window
Best case Worst
case
Bubble sort 36 36
Shear sort 18 36
Modified shear sort 21 21
III. DECISION BASED UN SYMMETRIC TRIMMED
MEDIAN FILTER(DBUTM)
Decision Based Algorithm (DBA) is a recently proposed
algorithm to remove salt and pepper noise. In DBA each
pixel is processed for de noising using a 3 X 3 window.
During processing if a pixel is ‘0’ or ‘255’ then it is
processed else it is left unchanged. In DBA the corrupted
pixel is replaced by the median of the windo w. At higher
noise densities the median itself will be noisy, and, the
processing pixel will be replaced by the neighborhood-
processed pixel. This repeated replacement of neighborhood
pixels produces streaking effect. In DBUTM, the corrupted
pixels are identified and processed. The DBUTM algorithm
checks whether the left and right extreme values of the
sorted array obtained from the 3x3 window are impulse
values. The corrupted processing pixel is replaced by a
median value of the pixels in the 3 X 3 window after
trimming impulse values. The corrupted pixel is replaced by
the median of the resulting array.
left unchanged. However, in the second window, the center
processing pixel ‘255’ which is noisy is replaced by the
median of the neighborhood pixels which is found by
eliminating ‘0’ and ‘255’ among the neighborhood pixels
(here median=63).

A. Algorithm
The algorithm for DBUTM is as follows:
Step 1: A 2-D window ‘S xy ’ of size 3x3 is selected.
Step 2: The pixel values in the window are sorted in
ascending order, and stored in a 1-D array.
Step 3: If the pixel value in the array is either '0' or ‘255’,
the corresponding pixel values are trimmed (eliminated),
and the median of remaining values is calculated.
Step 4: The pixel being processed is replaced by the median
value calculated.
Move the window by one step, and repeat fro m step 2 to
step 4. The above steps are repeated, until the processing is
completed for entire image.Yes
0<Zx y<255
Noisy image
No
Figure.6. Illustration for DBUTM
Select a 2-D 3 X 3 window
with center element Zxy as B. IMPLEMENTATION FOR VIDEO SEQUENCE
processing pixel
T he video sequence is first converted into frames and
frames
into images. Then DBUTM algorithm is applied to the
Eliminate the elements
images which are separated from frames. After the
with values ‘0’ or ‘255’ in filtering
Algorithm:
Video to frames: The noisy video sequence containing
the 1-D array process, the frames are converted back to the original
impulse noise is converted into avi format, which is an
movie. format and frames are extracted from the
uncompressed
video.
Replace the processing Frames to images: Frames are then converted to images
pixel with the median o f 1- for further processing.
D array Filtering method: The noisy images are de noised using
DBUTM algorithm.

Restored image
processed frames are finally converted back
Frames to movie: After completing the entire process, the

Figure.5. Flow Chart For DBUTM into original

Illustration: movie.
The figure 6 shows an image matrix. In the first window at
the left end, the processing pixel is ‘57’ which lies between
‘0’ and ‘255’. Hence the processing pixel is uncorrupted and

402
411
 performance of the algorithm is better when compared to the
Avi to frame other architecture of this type.
conversio n

REFERENCES
Read in frames
a b c d e f a b c d e f
1. W.K.Pratt, Digital Image Processing (New York:
Figure 13. Results for 50% noise corrupted Barbara image (a) original Figure 9. Results for 90% noise corrupted lena image (a) original lena
image (b ) 30% noise corrupted image. Restoration results of (c) SMF (d) Wiley(b)1978).
image 90% noise corrupted image. Restoration results of (c) SMF (d)
AMF (e) DBA (f) DBUTM AMF2. (e) I. Pitas
DBA (f) and A.N. Venetsanopoulous, Non-linear
DBUTM
DBUTM
Digital Filter Principles and Applications, (Boston: Kluwer
Academic publishers, 1990).
3. J. Astola and P. Kuosmanen, Fundamentals of Non-
Linear Digital Filtering, BocRaton, CRC, 1997
Write out frames 4. N.C.Gallagher, Jr. and G.L.Wise, “A Theoretical
Analysis of the Properties of Median Filters,” IEEE Trans.
Acoust., Speech and Signal Processing, vol.ASSP-29,
a b c d e f pp.1136-1141, Dec.1981.
Frames to avi
Figure 14. Results for 90% noise corrupted Parrot image (a) original image 5. T.A.Nodes and N.C.Gallagher, “Median Filters:
(b) 90% noise corrupted image. Restoration results of (c) SMF (d) AMF (e)
DBA (f) DBUTM Some Modifications and their properties,” IEEE Trans.
Figure 7. Flowchart for processing a Video Sequence by DBUTM Acoust., Speech and Signal Processing, vol. ASSP-30,
algorithm pp.739-746, April 1987.
T able II Figure 10. Comparison graph of PSNR at different Noise densities for ‘lena’
IV. results
RESULTS AND DISCUSSION image
Quantitative of various 6. S.Manikandan, O.Uma Maheswari, D.Ebenezer,
filters for 30% corrupted
Barbara image "An Adaptive Recursive Weighted Median Filter with
The developed algorithms are tested usin g 512X512, Improved8- Performance in Impulsive Noisy Environment",
bits/pixel image Lena (Gray), Parrot (color), BarbaraWSEAS Transactions on Electronics, Issue 3, Vol.1, July
METRICS SMF AMF DBA DBUTM 2004.
(color). The performance o f the proposed algorithm is tested
7. Ho-Ming Lin, "Median Filters with Adaptive
for various levels of noise corruption and compared with
PSNR 23.84 28.36 28.09 29.37 Length", IEEE Transactions o n Circuits and Systems, Vol.
35, No. 6, June 1988
standard filters
MSE 268.74 94.94namely standard median filter (SMF),
100.98 75.13
adaptive median filter (AMF) and decision based dimensional Median Filtering algorithm" ,
8. T.S. Huang, G.J. Yang, and G.Y. Tang, "Fast two
algorithm
IEF 22.01 62.4 58.6 78.82
(DBA). Each time the test image is corrupted byIEEE salt and
pepper noise of different T able III density ranging from 10 Transactions
to11.Comparison
Figure
image
90 with graph onof MSE
Acoustics,
at different NoiseSpeech and Signal
densities for ‘lena’
anQuantitative
increment of 10
results and itfilters
of various willforbe90%
applied to various
corrupted filters.
Processing,
Parrot image to the visual quality, the perfo rmance of the
In addition
developed algorithm and other standard algorithms 1(1995),
are pp. 499-502.
quantitatively
METRICS SMF AMF measured
DBA DBUTMby the follo wing parameters 9. Pok, such as G. Jyh-Charn Liu, " Decision based
peakPSNRsignal-to-noise 23.68 (PSNR), Mean squaremedian
8.26 17.11 19.03ratio error
(MSE) and Image Enhancement Factor (IEF). All the filters
MSE 9699.2 1264.2 812.77 278.82 filter improved by predictions", ICIP
are implemented in MATLAB 7.1 on a PC equipped with
IEF 1.9 14.59 22.71 66.45
2.4 GHz CPU and 1 GB RAM memory for the Figure 99(1999)ofproceedings
evaluation
12. Comparison graph of IEF at different Noise densities for ‘lena’
computation time of all algorithms. The quantitative vol 2 pp 410-413
image

performances inV. terms CONCLUSION

of PSNR, MSE and IEF for 10.all theSrinivasan K. S. and Ebenezer D., "A
An efficient
algorithms non-linear
are algorithm
given in to remove
Table II high-density
and Table salt
III.
and p epper noise is proposed. The modified sheer sorting New Fast and
architecture reduces the computational time required for
a b c d e
finding the median. This increases the efficiency of the
f
Figure 8. Results for 70% noise corrupted lena image (a) original lena
E fficient Decision- Based Algorithm for
system. T he algorithm removes noise even at higher noise
image (b) 70% noise corrupted image. Restoration results of (c) SMF (d)
densities
AMF and(f)preserves
(e) DBA DBUTM the edges and fine details. The
Removal of High-
Density Impulse Noises" , IEEE signal
processing letters,
403Vol. 14, No. 3, pp.189 -192, 2007
404
412
413