a morphological approach to remove salt and pepper noise in images

Aravinth Chinnapalanichamy et al ,Int.J.Computer Technology & Applications,Vol 3 (6), 1875-1880 

ISSN:2229-6093 

A MORPHOLOGICAL APPROACH TO REMOVE SALT 

AND PEPPER NOISE IN IMAGES 

Aravinth Chinnapalanichamy 

Department of ECE 

College of Engineering Guindy 

Anna University 

Chennai-600025, India 

Arjun Dev Singh R 





Ajith.K.N 





ABSTRACT 

In this paper, a morphological approach based on pixel 

connectivity is proposed for removing impulse noise in 

images. Impulse noise, also called the salt and pepper noise, 

usually occurs in dark and bright regions of the image and the 

presence of this noise degrades the visual quality of the image 

by affecting its texture and fine details. Our algorithm first 

identifies the noisy pixels in image based on 4-connectivity 

between the processing pixel and adjacent noisy pixels in a 

chosen window. Based on the region on which the processing 

is being done, the identified noise is removed and a median 

filter of varying mask size is used for estimating the correct 

value. The proposed method is found to be effective in 

dealing with high density impulse noise and further all the 

fine details and texture of the image is preserved. 

Keywords 

Morphology, Connectivity, Impulse noise, Median filter, 

Mask. 

1. INTRODUCTION 

Often images are corrupted by impulse noise. This 

type of noise occurs generally in digital images due to the 

problems in scanning, video sensor problems, decoding errors, 

etc. Filtering impulse noise is one of the most important pre 

processing procedures that has to be done before feature 

extraction. 

Generally Median filters are used in removing the impulse 

noise. Median filter checks the value which is in the centre of 

the neighbourhood chosen and replaces the centre value with 

the median value of all those neighbourhood values. Some of 

the variants of median filter are Adaptive median filter 

(AMF), Rank order based Adaptive Median Filter (RAMF), 

Switching median filter (SMF), Decision based filter (DBF), 

Hybrid median filter (HMF), etc. However in many cases 

using median filter variants alone is not enough for ensuring 

the better quality of the filtered image. These filters not only 

suppress the noise but also affect the noise free pixels and 

hence, a blurred version of the original image is formed. 

Hence a different method in dealing with the impulse noise 

based on morphological image processing is proposed. This 

approach proves its efficiency for impulse noise removal in 

digital images by using pixel connectivity procedure. The 

removal procedure is of two main stages. The first one is 

noise identification which is followed by image restoration. 

This algorithm is based on how the noisy pixels are connected 

together within a particular window. While processing a 

particular pixel, the chosen window for that pixel includes 

both noise and noise free pixels. A connection is established 

between the noisy pixels alone and based on how many noisy 

pixels are connected; the processing pixel is determined to be 

noise or noise free. 

The proposed algorithm could restore the image even for a 

noise density of 80%. Many iterative procedures are used to 

filter the corrupted image and hence image restoration is 

achieved even in extremely corrupted images. 

2. PROPOSED ALGORITHM 

For processing according to the proposed 

algorithm, the image is divided into three regions as black (0), 

white (255) and gray shade regions (i.e.) intensities other than 

0 and 255. The algorithm consists of two phases. The first 

phase is used for identifying the noisy pixels. This in turn can 

be divided into two sub stages. In sub stage 1, the noisy pixels 

in black and white background regions are found and replaced 

with the correct values. This sub stage involves both noise 

identification and removal. 

In sub stage 2, the noisy pixels in gray shade 

regions are found using the connectivity principle. Pixel 

connectivity is the means by which each pixel can be related 

to their neighbours. In this algorithm, 4-connectivity is used. 

Pixels that are horizontally and vertically adjacent to a pixel is 

said to be the 4- neighbours of that particular pixel. 

In the second phase, the noise identified by sub 

stage 2 is replaced with an estimated value of the correct 

value. This is done by replacing the noise with the median of 

specific values contained within a varying window. In 

general, the second phase is applicable only to sub stage 2 

because sub stage 1 performs both noise identification and 

removal. 

. 

3. PHASE 1- NOISE IDENTIFICATION 

Impulse noise candidates are found in this stage. 

For processing according to our algorithm, first the sub stage 

in which the pixel must be processed is determined. In fig 1, 

phase 1 is illustrated using a flowchart. 

IJCTA | Nov-Dec 2012 

Available online@www.ijcta.com 

1875


ISSN:2229-6093 

The steps involved in phase 1 are specified as follows: 

1. The processing pixel f(x, y) is extracted along with 

its neighbourhood pixels. The mask size is chosen 

according to the noise density 

noise pixels occupying a very less area of the bell shaped 

curve of normal distribution (i.e.) the noise occupies only both 

the extremes and the noise free pixels are centred on the mean 

value of the neighbourhood. 

MASK SIZE NOISE 

DENSITY 

3 X 3 0-30% 

5 X 5 30-60% 

7 X 7 60- 90% 

For increasing noise density, the mask size should 

also be linearly increased for better noise 

identification. However increasing the mask size 

also imposes a blurry effect on the restored image. 

G = f x + s, y + t were s, t = −k to k 

and (2k + 1 X 2k + 1) is te mask size 

1. The proposed scheme is applied for an 8 bit image 

(Intensity range-0 to 255). The centre pixel is 

verified for 0 or 255. If it is one of the two values, 

then it is retained as a possible candidate for noise. 

Or else it is considered as noise free and the next 

pixel along with its neighbourhood is taken for 

processing. 

N(x, y) = 

1, f x, y = 0 or 255 

0, else 

2. If N(x, y) =1, then the matrix G is checked for 

values 0 and 255. If all the values in G are 0 and 

255, then sub stage 1 is used and if there are 

different intensity levels (0 to 255), then sub stage 2 

is selected for noise identification. 

G = 

substage 1, G(G ≠ 0 |255 = 0 

substage 2, else 

3.1 Sub stage 1- Image restoration in white 

and black regions 

The window extracted from a white or black region 

from a noise free image will have all its values as either 0 or 

255 .However if impulse noise is present, depending on noise 

density level, noisy 0 values replace some of the 255 values 

in white region and vice versa. But the majority of the pixels 

in a chosen window will be either 0 or 255 based on the 

region. Hence the no of 0’s and 255’s in the window is 

counted and f(x, y) is replaced with the value occurring 

maximum no of times. 

f x, y = 

0, Na > Nb 

255, else 

were Na = no of 0 ′ s in G and Nb = no of 255 ′ s inG. 

3.2 Sub stage 2- Noise Identification in 

Gray shade regions 

The distribution of pixels in a particular 

neighbourhood follows nearly a normal distribution with the 

68% of the total area of the curve is always 

considered as noise for processing according to the algorithm 

although it may also contain some noisy pixels for higher 

noise densities. As proposed earlier, pixel connectivity is used 

to identify a noisy pixel. Hence connectivity is established 

between the pixels contained within the remaining 32% of the 

curve. This portion of the curve may also contain some noise 

free pixels. However, this does not cause significant problem 

for noise identification. 

For establishing connectivity, certain conditions are 

specified based on which the intensity values in the extracted 

window are changed to either 0 (background pixel) or 

1(foreground pixel) and a 4-connectivity is established 

between the foreground pixels (1). The connection thus 

established between the foreground pixels is actually a 

connection among the noisy pixels in the chosen window. 

This procedure is given in the following steps. 

1. A function H is found which is a subset of G with G 

(centre) removed. Then the mean (µ) and standard 

deviation (σ) of H is found. From that, two values (R1 and 

R2) are found. 

R1=µ-σ and R2=µ+σ 

It is seen that almost all the noise free pixels are 

contained within this range and the noisy values are located 

outside the range. However there are certain cases like images 

with high noise density or regions with intensity values closer 

to 0 & 255 where the noisy pixels are also included in the 

range because the mean will be centred on 0 or 255 in either 

case. 

2. Now the values of H are converted to any of the three 

values 0, 255 and constant K based on the following 

conditions. This is based on the assumption that the values 

contained within the range R are noise free and those 

outside the range are noise. 



1876


ISSN:2229-6093 

 

If R1 is –ve and R2 is +ve, then all the values 

except 0 contained within this range are 

converted to a constant (k) and the values greater 

than R2 is converted to 255[con.1] 

H = 

k, R1 < H < R2and H ≠ 0 

0, H = 0 

255, H > R2 

If R1 is +ve and R2 is +ve but less than 255 

(highest possible intensity value), then all the 

values within this range are converted to a 

constant K. The values less than R1 are converted 

to zero and the values greater than R2 are 

converted to 255.[con.2] 

H = 

0, H < R1 

k, R1 < H < R2 

255, H > R2 

 

If R1 is +ve and R2 is +v2 but greater than 

maximum possible intensity value (255), then all 

the values except 255 within the range are 

converted to a constant K. And the values less 

than R1 are converted to 0.[con.3] 

H = 

0, H < R1 

k, R1 < H < R2 and H ≠ 255 

255, H = 255 

3. At this stage, the function H contains only the above 

mentioned three values. Now, the Centre pixel of G (i.e.) 

f(x, y) is added to H to make it a (2k+1 x 2k+1) matrix. H 

values are again altered based on f(x, y). 



1877


ISSN:2229-6093 

 

If f(x, y) =0, first all H values equal to zero are 

changed to 1 and then all the other values are 

changed to 0. The order in which the values are 

changed is important in this case. 

If f(x, y) =255, then all the H values equal to 255 

are changed to 1 and all the other values are 

changed to 0. 

3. All the above conditions and change of values are done 

only to facilitate finding the connectivity between 

pixels. The 4- connectivity between foreground pixels 

(value 1) in matrix H is found (i.e.) no of adjacent 

pixels (N) connected to the centre pixel is counted. 

4. If this number (N) is greater than a particular value, 

say A, then the processing pixel is said to be noise free. 

If it is less than A, then it is noise. The value A 

depends on the mask size. For a mask of size( 2k+1 x 

2k+1 ), the minimum no of pixels that should be 

connected together so that the centre pixel is noise free 

is (k+1)*(2k+1) 

f(x, y) = noisefree, N > A 

noise, N ≤ A 

were N is te maximum no of adjacent pixels connected 

to centre pixel based on 4 − connectivity and A = 

k + 1 ∗ (2k + 1) 

the median of those values are calculated and that median 

value replaces f(x, y). The steps of phase 2 is shown in Fig.2 

1. For a mask of size (2k+1 x 2k+1), the following values 

are extracted. 

f ′ = f(x + s, y + t) 

were s, t = 0 to 1 

2. The values other than 0 and 255 are found in f’. Then 

the median of those values is calculated which in turn is 

the estimated correct value. 

f x, y = med(f ′ f ′ ≠ 0 255 

3. If all the values of f’ are either 0 or 255, then the upper 

limit of sand t is incremented by 1. If still the function 

f’ has no values other than 0 and 255, then the limit is 

again incremented by 1 and this increment happens till 

K. 

f ′ = f(x + s, y + t) 

were s, t = 0 to p, p = 0,1, … k 

And f x, y = med(f ′ f ′ ≠ 0 255 . 

4. PHASE 2- NOISE REMOVAL USING 

MEDIAN FILTER OF VARYING MASK 

SIZE 

Once the pixel has been identified as noise, then it is 

replaced with an estimated value of the correct intensity level. 

From the G matrix, certain elements are extracted and then 

5. EXPERIMENTAL RESULTS 

Simulations are done on various standard images at 

different noise levels. The performance of this scheme is 

measured by using the parameter PSNR and compared with 

the PSNR values of other methods and is shown that our 

method is superior to other methods. 

The standard images chosen for testing are Lena, Boat and 

gold hill. All the images are 8 bit gray scale images with size 

512x512. The mask size is chosen accordingly to the noise 

density level. For instance, while testing with Lena image 

corrupted by impulse noise level of 70%, the mask size 

chosen is 5x5. This implies that the value of K is 2. Further 

while identifying a noisy pixel in sub stage 2, the value of A is 

set as 15. While processing a particular pixel of the image, if 



1878


ISSN:2229-6093 

the no of adjacent pixels connected to the processing pixel 

after window modification exceeds 15, then the pixel is 

determined as noise free. To illustrate the quality of the 

filtered image, the parameter peak signal to noise ratio 

(PSNR) is used. The PSNR value found in this case is found 

to be 30.627 db 

The results of our experiment are summarised in the following 

graphs and table. In Fig.3 (a), original Lena image is shown 

and images corrupted with 70%, 80% noise density and the 

restored images are shown in latter figs. It is seen that the 

proposed scheme not only filters out the noise but also 

preserves the edge and other fine image details. The filtered 

images also have visually good quality. 

. 

Tab.1 shows the PSNR value calculated for Lena image at 

various noise ratios using different schemes. It is clearly 

evident from the table that our algorithm is superior to others. 



1879


ISSN:2229-6093 

Simulation results of the other images Boat and Gold hill 

can be seen in Fig.4 

6. CONCLUSION 

A morphological approach based on pixel 

connectivity is used in dealing with high density salt and 

pepper noise. Connectivity is established between centre and 

adjacent pixels and based on how many no of pixels are 

connected, noise is determined. Performance of this scheme is 

tested on various standard images. The proposed algorithm 

gave good results when to compare to others. Further, it is 

seen that all the texture and edge information of the image is 

preserved in this method. 

. 

7. REFERENCES 

[1] W. K. Pratt, Digital Image Processing. New York: Wiley 

Interscience, 1991. 

[2] Toh, K.K.V. and Isa, N.A.M., 2010. Noise Adaptive 

Fuzzy Switching Median Filter for Salt-and-Pepper 

Noise Reduction. IEEE Signal Processing Letters, vol. 

17, no. 3, pp. 281-284. 

[3] Nair, M.S., Revathy, K., and Tatavarti, R., 2008. 

Removal of Salt-and-Pepper Noise in Images: A New 

Decision-Based Algorithm. In Proceedings of the 

International MultiConference of Engineers and 

Computer Scientists, vol. I. 

[4] T. Nodes and N. Gallagher, “Median filters: Some 

modifications and their properties,” IEEE Trans. Acoust., 

Speech, Signal Process., vol. ASSP-30, no. 5, pp. 739– 

746, Oct. 1982. 

[5] W. Luo, “An efficient detail-preserving approach for 

removing impulse noise in images,” IEEE Signal 

Process. Lett., vol. 13, no. 7, pp.413–416, Jul. 2006. 

[6] P. Civicioglu, “Using uncorrupted neighbourhoods of the 

pixels for impulsive noise suppression with ANFIS,” 

IEEE Trans. Image Process., vol. 16, no. 3, pp. 759–773, 

Mar. 2007 

[7] Esakkirajan, S., Veerakumar, T., Subramanyam, A.N. 

and PremChand, C.H., 2011. Removal of High Density 

Salt and Pepper Noise Through Modified Decision Based 

Unsymmetric Trimmed Median Filter, IEEE Signal 

Processing Letters, vol. 18, no. 5, pp. 287-290. 

[8] Gonzalez, R.C., and Woods, R.E., 2004. Digital Image 

processing (2nd edition), Pearson Education. 

[9] Xin Zangju, Wang Shoujue, Deng Haojiang, Luo Yujing, 

“A new filtering algorithm based on extremum and 

median value”, Journal of Image and Graphics, 

2001,6(6),pp.25-28. 

[10] Gou Zhongkui, Zhang Shaojun, Li Zhongfu, Jin Jian, 

“New adaptive median filter algorithm based on extreme 

value”, Infrared and LaserEngineering, 2005, 34(1), 

pp.98-101. 

[11] H. Chan, H. Chung-wa, and M. Mikolova, “Salt and 

pepper noise removal by median type noise detectors and 

detail-preserving regularization,” IEEE Transactions on 

Image Processing, vol. 14, pp.1479-1485, Oct. 2005. 

[12] A. K. Jain, “Fundamentals of Digital Image Processing,” 

Englewood Cliffs, NJ: Prentice-Hall, 1989. 



1880

a morphological approach to remove salt and pepper noise in images

Create successful ePaper yourself

Delete template?

Save as template?