VLSI Architecture of Switching Median Filter for Salt and Pepper Noise Removal

Transcription

1 VLSI Architecture of Switching Median Filter for Salt and Pepper Noise Removal V. R. Vijaykumar, G. Santhanamari, S. Elango Abstract In this paper, VLSI architecture of new switching based median filter to remove high density salt and pepper noise in digital images is proposed. The absolute difference between center pixel and the median of trimmed array obtained from a 3 x 3 sliding window is compared with the predefined threshold value to identify the pixel is noisy or not. In the filtering stage, the noisy pixels are replaced by median of noise free pixels in the 3 x 3 filtering window. The experimental results for various test images show that the performance of the proposed algorithm is superior to existing algorithms, namely SMF, ACWMF, TMF, PWMAD, ARWMF, REBF, MDBUTMF and NAWMF in terms of visual quality and edge preservation. The proposed algorithm is also implemented with VHDL and simulated using Xilinx The quantitative analysis in terms of logic elements, power and delay are observed in Altera Quartus II and compared with existing state of art algorithms, namely SMF, DBA, Parallel sorting, REBF, MDBUTMF and NAWMF. Index Terms Impulse Noise, edge preservation, Median Filter, Noise Detector, VLSI implementation, I I. INTRODUCTION N the process of image acquisition and transmission over the channel, the images are frequently influenced by some external environment and corrupted with impulses. The fixed valued impulse noise corrupts the true intensity value in random position with corruptive values in the extreme ranges, ie 0 (pepper) and 255 (salt) called as salt and pepper noise. Another type of impulse noise is random valued impulse noise which affects the true intensity value with corruptive values in the range [0, 255], which is also the dynamic range of the image. The objective of filtering is to restore the original image from the noise corrupted image. Generally, linear filters can remove the impulse noise, but it blurs the image. Hence the best known nonlinear filter namely standard median filter (SMF) [1] is widely used due to its simplicity and computational efficiency. But it exhibits blurring effect for larger window size and less noise suppression for smaller Manuscript received November 27, 2014; revised August04, 2015 V. R. Vijaykumar is currently working as an associate professor in the department of ECE, Anna University-Regional Center, Coimbatore. ( vr_vijay@yahoo.com) G. Santhanamari is currently working as an assistant professor in the department of ECE, Tamilnadu College of Engineering, Coimbatore. ( santhanam88@yahoo.co.in) S. Elango is currently working as an assistant professor in the department of ECE, Bannari Amman Institute Technology, Sathiyamangalam. ( eceelango@gmail.com) window size, at higher noise density. The weighted median filter (WMF) [2], center weighted median filter (CWMF) [3] and adaptive center weighted median filter (ACWMF) [4] are proposed to improve the performance of the standard median filter by giving more weight to some selected pixels in the filtering window. However, most of the median based filters are applied to all the pixels in the noisy image which affects both noise and noise free pixel intensity which leads to blurring of an output image. Hence, many filtering algorithms with a switching strategy which discriminates the corrupted pixels and uncorrupted pixels namely, tristate median filter (TMF)[5], advanced impulse detection based on pixel wise median of absolute deviation (PWMAD)[6] and new decision based algorithm (DBA) [7] are proposed. In TMF, the corrupted pixel is replaced by either the median value or the center weighted median value based on the threshold value and noise free pixels are left unaltered. But it performs well for the images corrupted with slightly higher impulse noise ranges to 50% only. The DBA used 3 x 3 fixed window for detection and filtering process. If the processing pixel is either 0 or 255, then it is replaced with a median value of local neighborhood pixels in the 3 x 3 sliding window, otherwise retained. At higher noise level all the pixels in the selected window are corrupted, and the median value may also be a noisy value. In that case, the left neighborhood pixel is used to replace the corrupted center pixel which produces the streaking effect. Also the edges are not recovered satisfactorily, since the local feature in filtering window is not taking into account. To overcome the above drawback, adaptive recursive weighted median filter (ARWMF) [8] and robust estimation based filter (REBF) [9] are proposed. The ARWMF used median controlled algorithm for weight calculation to achieve a high degree of noise suppression and edge preservation. The REBF used an influence function based on the local estimate of image standard deviation to calculate the estimated value of corrupted pixel which gives better restoration results. The decision based unsymmetric trimmed median filter (DBUTMF) proposed in [10] uses the trimmed median value to replace the noisy pixel. At higher noise level if the selected window contains all the pixels as noisy pixels, then the trimmed median value cannot be obtained. In addition to that, DBUTMF does not provide better restoration results when the noise level is more than 60%. The modified decision based unsymmetric trimmed median filter

2 (MDBUTMF) is proposed [11] as a remedy for the above drawback. When all the pixels in the selected window are corrupted at higher noise density, it takes the mean value of all the pixels in the sliding window to replace the corrupted center pixel which may also be a noisy value. In addition to that, unlike the median filter, mean filter, smoothens the image. To overcome the above said issue, the threshold based noise detection mechanism is proposed in this paper, instead of direct detection. Recently a new adaptive weighted mean filter (NAWMF) has been proposed in [12]. Over the years, hardware implementation of median filter has been attempted in software and also available in the DSP processor environment. Since the VLSI implementation of median filters employs a sorting technique, it is the major concern to implement the median filter in hardware for real time application. In [13], an architecture of rank based 2D median filter is implemented in FPGA. The high throughput VLSI architecture for an existing median filter introduced in [14] and pipelined median filter architecture introduced in [15] reduce the cell count, but they have not processed the real time image. The optimized sorting architecture for the median filter introduced in [16] is an efficient architecture, but it is not implemented in image processing applications. The high speed pipelined architecture for adaptive median filtering algorithm proposed in [17] uses parallel sorting architecture to find the median value. Though parallel sorting architecture increases the speed of operation, hardware complexity is also increased. The major challenges in various kinds of the architecture oriented median filtering algorithms are their computational time and hardware cost. In this paper, an efficient VLSI architecture for the proposed switching based median filter is also presented. The rest of the paper is organized as follows. In Section II, the proposed median filtering algorithm is described. The VLSI architecture of proposed median filter is discussed in Section III. Section IV presents the simulation and implementation results. Finally, the conclusion is given in Section V. II. PROPOSED SWITCHING MEDIAN FILTER The proposed algorithm process each and every pixel in the noisy image to detect the presence of noise and filtering is applied if it is corrupted otherwise left unaltered. The fixed valued impulse noise corrupted pixels can take either maximum (S max ) or minimum (S min ) intensity values in the dynamic range [0, 255]. If the processing pixel lies within the range S min < X ij < S max, then it is noise free and actual value is retained. If it is identified as noisy as described in the following algorithm steps, then the median value of neighborhood pixels is used to replace the noisy intensity value. The proposed algorithm is described in the following steps. Algorithm Steps: Step 1: Apply a 3 x 3 filtering mask S ij as shown in figure.1, centered about the processing pixel X ij in the noisy image. The minimum intensity values (S min ) and maximum intensity value (S max ) are determined by sorting the elements in S ij Step 2: If S min = 0 or S max = 255, then continue, else go to step6. Step 3: Get the noise free pixel intensity values in the filtering mask whose coordinates are defined in equation (1) into an array U. If the array U is empty, then going to step5, else continue. S ij = X i+k,j+l k,l;-1:1 S min^ S max (1) Step 4: Median of the array U is found. If the absolute difference X ij median (U) > T then X ij is the impulse corrupted pixel and it is replaced with the median (U) otherwise left unaltered and go to step7. X i-1,j-1 X i-1,j X i-1,j+1 X i,j-1 X i,j X i,j+1 X i+1,j-1 X i+1,j X i+1,j+1 Fig 1. 3 x 3 filtering mask Step 5: The median of S ij as defined in equation (2) is used to replace corrupted X ij and go to step7. Median (S ij ) = Median (X i+k,j+l -1< k,l < 1 ) (2) Step 6: The processing pixel X ij is uncorrupted and actual intensity value is retained as such. Step 7: Repeat the process from step1 for all the pixels in the noisy image to restore the corrupted pixel intensities. The proposed algorithm is tested on different images with different characteristic like, images with smooth regions and images with high frequency details. The optimum threshold (T) value is obtained based on trial and error approach and found that, it is in the range of 25 T 30 for better restoration and edge preservation. III. ARCHITECTURE OF PROPOSED MEDIAN FILTER The architecture of proposed median filter mainly consists of noise detector, sorting network and switching stage as shown in figure 2. A. Noise Detectors Generally the images are corrupted with salt and pepper noise during the image acquisition and transmission process. Due to the addition of salt and pepper noise also called as fixed valued impulse noise, the pixel value gets modified to Fig 2. Block Diagram of Proposed Median Filter

3 either minimum gray scale value (0) or maximum gray scale value (255). The two types of noise detector namely salt noise detector (all one detector) which detects the maximum gray scale value and pepper noise detector (all zero detector) which detects the minimum gray scale value are used to detect the fixed valued impulse noise. B. Sorting network Let us consider a sorting of, 9 elements in an array, where I represents the current pixel intensity value to be sorted and J represents the rest of the elements in an array. (i) Salt Noise detector Let us consider an input A=A 7, A6 A 0 f (1)= A 7 & A 6 & A 5 & A 4 & A 3 & A 2 & A 1 & A 0 ; If f (1) = 1 then A is corrupted pixel (Salt Noise detected); Else A is uncorrupted Pixel; The figure 3 shows the logic diagram and boolean expression of the salt noise detector. It detects whether all the bits of a pixel intensity value are one (255) or not. Algorithm Step 1: The value I is compared with the rest of the J values. Step 2: If I is less than J, then shift the I value to the temporary register t. Step 3: Shift the J value to I register. Step 4: Shift the value in temporary register to J register. Step 5: If I is not less than J, then go to step 1 C. Switching Stage The switching stage as shown in figure 5 consists of mainly a threshold detector and multiplexer. where, TV- threshold value. CV- current pixel intensity value. Fig 3. Logic Diagram of Salt Noise detector f ( 1) z(0) z(1) z(2) z(3) z(4) z(5) z(6) z(7) (ii) Noise Detector Let us consider an input A=A 7, A6 A 0 Z = A 7 A 6 A 5 A 4 A 3 A 2 A 1 A 0 ; f (0) = not(z); If f (0) = 1 then A is corrupted pixel (Pepper Noise detected); Else A is uncorrupted Pixel; The figure 4 shows the logic diagram and boolean expression of pepper noise detector. It detects whether all the bits of a pixel intensity value are zero (0) or not. MV-median value of noise free neighborhood. The threshold detector checks whether the input value, that is the absolute difference between processing pixel intensity value and the median of noise free neighborhood in 3 x 3 sliding window lies in the range of 27 to 30 or not. If it lies in that range, then the output of threshold detector is one, otherwise it is zero. The figure 6 shows the logic diagram and boolean expression of the threshold detector. The inputs to the threshold detector are 8 bits of the absolute difference T(8), T(7), T(0) and their inverted values T(8), T(7), T(0). It detects whether the absolute difference lies within the threshold range [27-30] with a high level output, which is given to the selection line of the multiplexer. If the output of threshold detector is logic 1, then the multiplexer selects the actual pixel intensity value X ij otherwise it selects the median value already obtained. T(8)' T(7)' T(6)' T(5) T(4) T(3) T(2)' F f Fig 4. Logic Diagram of Pepper Noise Detector ' 0 z(0) z(1) z(2) z(3) z(4) z(5) z(6) z(7) T(3) T(1) T(2) Fig 6. Logic Diagram of Threshold Detector

4 T F T ( 8) T (7) T (6) T (5) T (4) (3) T (2) T (3) T (1)) T (2) IV. ILLUSTRATION The proposed denoising algorithm is illustrated by considering a 3 x 3 window for three different cases as given below and the simulation result of VHDL implementation for the same three cases is also shown in figure. 7. Case (i): The intensity value of the center pixel in a 3 x 3 sliding window lies between 0 and 255 (ie.uncorrupted) and few remaining pixels are with both salt and pepper noisy intensity value. The noise free pixel intensity values (other than 0 and 255 ) in the filtering mask are collected in an array U. Since the absolute difference between center pixel and median value of the array U is lesser than 30, X ij is retained with the same intensity value (67) U = (43, 56, 75, 78) Median value of U is 66. (X ij median) is 67-66=1 and 1 < 30 Case (ii): In this case center pixel is having the noisy intensity value 0 (ie.corrupted) and 60% of remaining pixels are also with both salt and pepper noisy intensity value. The noise free pixel intensity values (other than 0 and 255 ) in the filtering mask are collected in an array U. The absolute difference between center pixel and the median value is greater than 30 and X ij is replaced with the median value (153) U = (148, 153, 161) Median value of U is 153. (X ij median) is = 153 and 153 > 30 Case (iii): In this case center pixel is having the noisy intensity value 0 and all other remaining pixels are also with both salt and pepper noisy intensity value. Since all the elements in the filtering mask are 0 and 255 (ie.corrupted), the median value of the 3 x 3 filtering mask 0 is used to replace X ij Median value of ( 0, 0, 0, 0, 0, 255, 255, 255) is 0. A. Simulation Results IV. RESULTS AND DISCUSSION In this section the extensive experiments are conducted on a variety of standard gray scale test images like Darkhair, Living room, Satellite and Mandril of size 512 x 512 with gray level intensity of 8 bits/pixel for noise level varying from 10% to 90% to evaluate the performance of the proposed algorithm. The visual results and quantitative results of proposed filter are compared with existing algorithms, namely SMF, ACWMF, TMF, PWMAD, ARWMF, REBF, MDBUTMF and NAWNF. The restoration performance and processing time for proposed filter and existing filters are analyzed under the following subsections (i) and (ii). Based on the experimental results, it is observed that to attain better visual quality and PSNR value, the threshold value for noise detection is found as 30. (i). Quantitative and visual results Comparisons The quantitative results in terms of peak signal-to-noise ratio (PSNR), Mean Absolute Error (MAE), and Image Enhancement Factor (IEF) are presented in table I-XII to illustrate the performance of the proposed filtering algorithm. The experiments are conducted using three test images, namely Darkhair, Satellite and Mandril that contain different characteristics like more smooth region and more edge detail. The SMF, ACWMF and PWMAD perform well only for very low noise density. The PSNR value obtained by TMF algorithm is slightly better than the above said algorithm due to its noise detection capability. The ARWMF uses adaptive window size to remove higher density noise and weight calculation for weighted median filtering is done iteratively till the least mean square error is obtained. Hence the better PSNR value obtained and minimum mean absolute error are obtained than SMF, ACWMF, PWAMD and TMF. The quantitative and qualitative performance of robust estimation based filter is equally good to proposed algorithm due to the following reasons. The first one is that, REBF uses adaptive window size based on noise density and the second is due to the calculation of influence function based on the local estimate of image standard deviation which is also used to find the estimated value of the corrupted pixel. The MDBUTMF works equally better to proposed filter for noise density up to 70%, but the performance is degraded at higher noise level. The recently proposed NAWMF gives better PSNR value and less MAE value than proposed filtering algorithm, since the window size is enlarged till the minimum and maximum intensity values of two successive windows are equal which leads to better noise detection and filtering mechanism. But it takes longer CPU time to run and needs complex hardware architecture for real time implementation. The MAE and IEF displayed in tables II, III, V, VI, VIII & IX show that the proposed filtering algorithm produces better IEF and less MAE than other existing algorithms for test images with various characteristic. The visual result of proposed algorithm and other existing algorithms for all the four test images, namely, Dark hair, Living room, Satellite and Mandril corrupted with 30%, 60% and 90% are presented in figures 8, 9 and 10 which show that the proposed filter performs better in qualitative aspect also.

5 F ig 7. Simulation result of VHDL implementation of the proposed algorithm for the cases (i-iii). TABLE I PSNR COMPARISION OF VAIROUS FILTERS FOR DIFFERENT NOISE DENSITY DARKHAIR IMAGE TABLE II MAE COMPARISION OF VAIROUS FILTERS FOR DIFFERENT NOISE DENSITY DARKHAIR IMAGE

6 TABLE III IEF COMPARISION OF VAIROUS FILTERS FOR DIFFERENT NOISE DENSITY Noise % DARKHAIR IMAGE TABLE IV PSNR COMPARISION OF VAIROUS FILTERS FOR DIFFERENT NOISE DENSITY SATELITE IMAGE TABLE V MAE COMPARISION OF VAIROUS FILTERS FOR DIFFERENT NOISE DENSITY SATELITE IMAGE TABLE VI IEF COMPARISION OF VAIROUS FILTERS FOR DIFFERENT NOISE DENSITY SATELITE IMAGE

7 TABLE VII PSNR COMPARISION OF VAIROUS FILTERS FOR DIFFERENT NOISE DENSITY MANDRIL IMAGE TABLE VIII MAE COMPARISON OF VAIROUS FILTERS FOR DIFFERENT NOISE DENSITY MANDRIL IMAGE TABLE IX IEF COMPARISON OF VAIROUS FILTERS FOR DIFFERENT NOISE DENSITY MANDRIL IMAGE (ii). Computation time Comparison The performance of the proposed algorithm is also evaluated quantitatively in terms of CPU time and compared with other existing algorithm for all four test images, namely Lena, Living room, Satellite and Mandril by varying noise density from 10% to 90%. The proposed filter and other existing algorithms are simulated in MATLAB7.1 on a PC equipped with 2GHZ operating speed 2GB RAM and demonstrated in the tables X, XI & XII. The run time for SMF, PWMAD and TMF is lesser than the proposed algorithm, since there is no or inefficient detection mechanism involved in those filtering algorithms. Also, their visual results and other performance metrics are not as good as proposed algorithm. It is seen from the table that, the adaptive window algorithms ACWMF, NAWMF and REBF take more computation time than the proposed fixed window algorithm, since the number of pixels to be processed is more, especially when the window size is very large at higher noise density. Similarly, the ARWMF is also executed with much more CPU time than all the other algorithms, due to the number of iterations invoked in weight calculation is more. It is also inferred from the table that, though CPU takes lesser run time to execute MDBUTMF for low noise density, execution time is increased for noise ratio greater than 70% which is more than the proposed algorithm. Hence it is clear from the above analysis that, the proposed algorithm shows better performance quantitatively without compromising quality in visual results.

8 TABLE X RUN TIME COMPARISION OF VAIROUS FILTERS FOR DIFFERENT NOISE DENSITY DARKHAIR IMAGE SMF TMF ACWMF PWMAD ARWMF RE BF MDBUTMF NAWMF PROPOSED TABLE XI RUN TIME RUN TIME COMPARISION OF VAIROUS FILTERS FOR DIFFERENT NOISE DENSITY SATELITE IMAGE SMF TMF ACWMF PWMAD ARWMF RE BF MDBUTMF NAWMF PROPOSED TABLE XII RUN TIME COMPARISION OF VAIROUS FILTERS FOR DIFFERENT NOISE DENSITY MANDRIL IMAGE Nois e% SMF TMF ACWMF PWMAD ARWMF RE BF MDBUTMF NAWMF PROPOSED B. Hardware implementation results The VLSI architecture of proposed trimmed median filtering algorithm is described with VHDL. The Xilinx Fig 11. Experimental Setup 10.1 Modelsim 6.2. The developed algorithm is tested with 256 x is used to produce a gate level net list and synthesized using 256, 8-bits/pixel gray scale Lena image for 60% salt and pepper noise density. The intensity values in rows of noisy image are stored as column vectors in a file and processed in proposed architecture and restored intensity values are stored in another file. Finally MATLAB tool is used to convert the estimated value of all pixel intensities available as column vectors in a file into an image as per the experimental setup shown in figure 11. The restoration result of 60% noise corrupted Lena image is also shown in figure 12. The logic element comparisons were made in ultra Quartus II [23]. The Quartus II power play analyzer tool is used to measure the power Consumption and delay analysis

9 IAENG International Journal of Computer Science, 43:1, IJCS_43_1_06 is made in Quartus II classic timing analyzer tool. The table XIII shows the simulation results of the proposed architecture for a 3x3 window. It can be seen from the table XIII that, the architecture of REBF and NAWMF need more logic element which increases the power consumption and hence more PDP, due to the requirement of more comparator and multiplier units. Also at higher noise density the logic elements required for processing the current pixel and calculating the estimated value of the corrupted pixel have increased much due to larger window size. Since the ARWMF is executed in an iterative manner, the hardware implementation is much complicated. It is also inferred from the table that, the VLSI architecture for proposed filtering algorithm achieves better PDP than other existing filtering algorithm except MDBUTMF. The reason is that, there is no threshold comparison step in MDBUTMF as used in the proposed filtering algorithm to identify the presence of noise. Since the efficient noise detection mechanism is involved to achieve better PSNR value, the architecture of the proposed filter needs additional circuitry which increases the required logic elements which in turn increases the PDP. Though logic elements required and PDP obtained for the proposed filter are almost equivalent to the other fixed window algorithms like SMF and MDBUTMF, the proposed algorithm gives better quantitative and qualitative results than the same. TABLE XIII COMPARISON OF LOGIC ELEMENTS, POWER, DELAY AND PDP FOR A SINGLE 3X3 WINDOW PROCESSING TOTAL PARAMETER/ LOGIC DELAY PDP POWER ALGORITHM ELELMENTS (ns) (pj) (mw) SMF [3] Decision Based Algorithm [10] Parallel Sorting [17] MDBUTMF [11] REBF [10] NAWMF [9] Proposed (a). Original (b). Noisy (c). Restored Fig 12. Simulation result of VLSI architecture of proposed algorithm for 60% noise corrupted Lena image. (a) (b) (c) (d) Fig 8. Restoration results of (a). Darkhair, (b). Livingroom, (c). Satellite, (d). Mandril image using various filters namely SMF, TMF, ACWMF, PWMAD, ARWMF, REBF,MDBUTMF, NAWMF and proposed algorithm for 30% salt and pepper noise density.

10 IAENG International Journal of Computer Science, 43:1, IJCS_43_1_06 (a) (b) (c) (d) Fig 9. Restoration results of (a). Darkhair, (b). Livingroom, (c). Satellite, (d). Mandril image using various filters namely SMF, TMF, ACWMF, PWMAD, ARWMF, REBF,MDBUTMF, NAWMF and proposed algorithm for 60% salt and pepper noise density. (a) (b) (c) (d) Fig 10. Restoration results of (a). Darkhair, (b). Livingroom, (c). Satellite, (d). Mandril image using various filters namely SMF, TMF, ACWMF, PWMAD, ARWMF, REBF,MDBUTMF, NAWMF and proposed algorithm for 90% salt and pepper noise density

11 VI. CONCLUSION A new switching based trimmed median filter using 3 x 3 filtering window for effective removal of high density salt and pepper noise and its VLSI architecture is proposed in this work. The image degradation caused from undetected noisy pixels is prevented due to the better noise detection capability of the proposed algorithm. The experimental results in terms of qualitative and quantitative metrics of proposed algorithm and other state of art technique are compared and better performance of proposed filter is demonstrated. In addition to that, VLSI architecture of the proposed filtering algorithm is also implemented and performance in terms of logic element, delay and power delay product is compared with other existing algorithms which clearly show the simplicity of the porposed architecture. [17] Dhanasekaran D. and Boopathy Bagan K., High Speed Pipelined Architecture for Adaptive Median Filter, European Journal of Scientific Research ISSN X Vol.29 No.4, pp , REFERENCES [1] I. Pitas and A. N. Venetsanopoulos, Order statistics in digital image processing, Proc. IEEE, vol. 80, no. 12, pp , Dec [2] D. R. K. Brownrigg, The Weighted Median Filter, Commun.Assoc.Comput.Machin, vol. 22, pp , [3] Sung Jea ko Chen and Yong Hoom Lee, Center-Weighted Median Filters and Their Applications to Image Enhancement, IEEE Transaction on Circuits and Systems, vol. 38, no. 9, [4] Tao Chen and Hong Ren Wu, Adaptive Impulse Detection Using Center-Weighted Median Filters, IEEE Signal Processing Letters, vol. 8, no. 1, [5] Tao. Chen, Kai-Kuang Ma, and Li-Hui Chen., Tri-state Median Filter for Image Denoising, IEEE Trans. Image Process., vol. 8, no. 12, pp , [6] V. Senk V. Crnojevic and Trpovski. Advanced Impulse Detection Based on pixel wise MAD, IEEE Signal Process. Lett, 11(7): , [7] K. S. Srinivasan and D. Ebenezer, A new fast and efficient decision based algorithm for removal of high density impulse noise, IEEE Signal Process. Lett., vol. 14, no. 3, pp , Mar [8] V. R. Vijaykumar, S. Manikandan, D. Ebenezer P.T. Vanathi and P. K. Kanagasabapathi, High Density Impulse Noise Removal in Colour Images using Median controlled Adaptive Recursive Weighted Median Filter, IAENG International. Journal. of Computer science, Vol. 34, No. 1, pp , [9] V. R. Vijaykumar,, P.T. Vanathi, P. K. Kanagasabapath and D. Ebenezer, High Density Impulse Noise Removal using Robust Estimation Based Filter, IAENG International. Journal. of Computer science, Vol. 35, No. 3, pp , [10] K. Aiswarya, V. Jayaraj, and D. Ebenezer, A New And Efficient Algorithm For The Removal Of High Density Salt And Pepper Noise In Images And Videos, in Second Int. Conf. Computer modeling and Simulation,, pp , [11] S. Esakkirajan, T. Veerakumar, Adabala N. Subramanyam, and C. H. PremChand 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 , 2011 [12] Peixuan Zhang and Fang Li A New Adaptive Weighted Median Filter for Removing Salt and Pepper Noise, IEEE Signal Process. Lett., vol. 21, no. 10, pp , [13] Swenson R.L. and Dimond K.R., A Hardware FPGA Implementation of a 2-D Median Filter Using a Novel Rank Adjustment Technique, in 7 th IEEE.Int.conf.proc.on Image Processing and it s Applications, vol. 1, pp , [14] Ravi Teja V.V., Ray K.C., Chakrabarti I., Dhar A.S., High Throughput VLSI Architecture for One Dimensional Median Filter, IEEE-International Conference on Signal processing, Communications and Networking (ICSCN), Madras Institute of Technology, Anna University Chennai India, pp , [15] Kuo-Liang Chung, Yih-Kai Lin A generalized pipelined median filter network, Elsevier signal Processing 63, pp , [16] Vasanth K., Nirmal raj S., Karthik S., and Preetha Mol P., FPGA Implementation of Optimized Sorting Network Algorithm For Median Filters, IEEE.Int.conf.proc.on Emerging trends in Robotics and communication Technologies, pp , 2010