Medical images are distorted by salt and pepper noise is largely caused by the acquisition or transmission of digital images in a noisy channel . Salt and pepper noise is one type of impulse noise which can degrade the image visualization, where the noisy pixels can take only the minimum and maximum pixel elements of image. The linear filter like mean filter and interrelated filters are not efficient for removal of salt and pepper noise. Non-linear filtering techniques like standard median filter (SMF) and Adaptive median filter (AMF) are widely used to remove salt and pepper noise . However, the SMF, which restores each pixel with median pixel value in the filtered image, exhibits blurring of reconstructed image . SMF is effective only at low noise densities. Several techniques have been implemented for removal of salt and pepper noise at high noise densities [1-9].
Recent methods like efficient detail-preserving approach (EDPA) and efficient edge-preserving algorithm (EEPA) based on alpha-trimmed mean statistical estimator have been proposed for the removal of salt-and-pepper noise without degrading image fine information’s . Then, adaptive impulse detector with center-weighted median (ACWM) filter efficiently to remove salt-and-pepper noise. These techniques simply perform well when an image is degraded with 50% of noise densities and small window size. On the other hand, Progressive Switching Median Filter(PSMF), decision based algorithm (DBA), modified decision based algorithm (MDBA), are one of the fastest and efficient algorithms of capable of salt-andpepper noise removal at high noise densities as 80% . A most important disadvantage of this algorithm is striping problem at high noise densities. To overcome this drawback, modified decision based unsymmetric trimmed median filter (MDBUTMF) is used to remove salt-and-pepper noise at very high densities as 80-90%. In response to this noise problem, the switching median filter , in which an impulse detector of employed to classify the center pixel of the filtering window is proposed. In the filtering algorithm, at high noise density, the processing pixel is replaced by the mean value of pixel elements within the window . This causes blurring of fine image details or high computational time . Examining all those problems exist in literature, we have proposed a new switching based algorithm, which aids to accurate medical image visualization. This algorithm gives better metric values than the existing median based algorithms [2-6].
2. Proposed Algorithm
The modified fuzzy switching median based filter is a non-causal two-stage filter, one stage for finding the pixels are noise-free or noise pixels and second stage for noise reduction as shown in figure Fig. 1. The first step to detect the two salt-and-pepper noise intensities. When a “corrupted pixel” is discovered, it is switched to the next filtering stage . Otherwise, when a pixel is marked as “uncorrupted pixel” it will be left unchanged and the filtering process is spared to avoid replacing any fine details and textures in the noisy image .
Fig. 1.Proposed medical image enhancing structure
2.1 Switching noise detection process
The filter initially estimates the two salt-and-pepper noise intensities of histogram of noise image shown in Fig. 2, and then the detection stage starts by searching for the two local maximums, Lsalt=255 and Lpepper=0, representing the two salt-and-pepper noise intensities starting from both ends of noisy image histogram . These two salt-and-pepper noise pixels will be used to recognize possible “corrupted pixels” in the image as shown in Fig. 3. A binary noise values B(m, n) will be formed based on noise pixels as shown Fig. 4.
Fig. 2.Histogram of Lung CT image distorted with 50% of Salt-and-pepper noise
Fig. 3.(7x7) Noisy image pixel values
Fig. 4.(7x7) Binary mask value of noisy image
where X(m, n) is the pixel at location (m, n) with intensity X, B(m, n)=1 indicates “uncorrupted pixels” to be obtained from the distorted image and B(m, n)=0 indicates “corrupted pixels”.
2.2 Fuzzy noise cancellation
In this stage after creation of binary noise mask value, “corrupted pixels” marked with B(m, n)=0 is replaced by an estimated adjustment pixel value and then increase the window size W2s+1(m, n) with (2s+1)x(2s+1) of chosen noise image. Then, the number of “uncorrupted pixels” in the filtering window W2s+1(m, n) is calculated using
where S value is choose the window size of image
After counting of uncorrupted pixel, the minimum number of “uncorrupted pixel” is less than one then expands the filtering window by one pixel . This process is continued until the condition G2s+1(m, n) greater than or equal to one . The uncorrupted pixels select the median pixel M(m, n), given by
The corrupted pixels select the median pixel M(m, n), given by
After finding the value of median pixel M(m, n), the local information in a 3 × 3 window is extracted by computing the absolute pixel difference α(m, n), given by
Then the local information is described by the maximum absolute pixel difference in the 3×3 filtering window
In filtering stage, the Fuzzy reasoning is applied to the extracted local information d(m, n). The adopted fuzzy set, which is defined by the Two-sided π-membership function(MF) f(m, n) with four predefined thresholds T1(5), T2(10), T3(245) and T4(250) as shown in Fig. 5 .
Fig. 5.Two-sided π-membership function
The two-sided π-membership function , given by
Where S(d(m,n),T1,T2) is S-shaped open-right MF as shown in Fig.6, given by
Fig. 6.S-Shaped membership function
and Z(d(m, n), T3, T4) is Z-shaped open-left MF as shown in Fig. 7, given by
Fig. 7.Z-Shaped membership function
Where the local information d(m, n) is used as the fuzzy variable and the four thresholds T1, T2, T3 and T4 are set to 5,10,245 and 250 respectively for best performance . The restored image Y(m, n) is acquired by linear combination of weighted processing pixel X(m,n) and median pixel M(m, n) of noisy image .
3. Simulation Results and Discussion
In this section, the performance of the proposed fuzzy based median filter was tested with medical images with noise levels are varied from 10% to 90% and compared with other existing salt-and–pepper noise removal filters. The restoration performances are quantitatively measured by several different kinds of metric value such as peak signal to noise ratio (PSNR), mean square error (MSE), RMSE, image enhancement factor (IMF), mean absolute error (MAE) and elapsed time .
The PSNR is most commonly used as a quality measurement for lossy compressed images. The PSNR is the ratio of the maximal power of original image and the noise power of distorted image. It is represented in the logarithmic domain because the powers of signals are usually in a wide dynamic range . Its formula is given by
MSE is the most widely used image quality assessment based on error sensitivity . It is computed by taking the average of squared intensity differences in every pixel of a reference image and a distorted image:
where I(m, n) is Original image, Y(m, n) is Denoised image X (m, n)-Noise image and (M x N)-size of the image.
To obtain more correct assessments, the structural simailarity (SSIM) metric which represents perceptual image quality based on the structural information . SSIM is an objective image quality metric and is superior to traditional quantitative measures such as MSE and PSNR .
A general form of SSIM is
Where α>0, β>0 and γ >0 are parameters used to adjust the relative importance of the three components and m, n are image patches.
Where l(m,n) is luminance comparison
c(m, n) is contrast comparison
and s(m,n) is structural comparison
Where C1, C2, C3 are constants and μm, μn,σm, σn, σmn are local sample means and standard deviations of m and n.
The MSE, RMSE, MAE, PSNR, IEF, elapsed time in seconds and SSIM value are calculated for the proposed algorithm by varying the noise density from 10% to 90% and comparisons of performance values of various existing median based filters for lung CT scan image are tabulated in Tables 1-7.
From the Table 1, shows that the MSE value of proposed algorithm is very low compared to other existing noise removal methods even noisy densities varied from 10% to 90%. The RMSE value in Table 2 shows that at high noise levels, RMSE values of existing noise removal filters are very high as compared to the proposed method. From Table 3, we illustrate that the MAE value of proposed method is very low compared to other existing noise removal method.
Table 1.MSE of Lung CT scan image for various filters at different noise densities
Table 2.RMSE of Lung CT scan image for various filters at different noise densities
Table 3.MAE of Lung CT scan image for various filters at different noise densities
The PSNR, IEF and SSIM values of proposed method in Table 4, Table 5 and Table 7 shows that at higher noise levels the PSNR, IEF and SSIM values of existing methods are very low as compared to the proposed method. From all the performance values of proposed method, it is clearly observed that the performance of the proposed algorithm is moderately significant than the existing algorithms at both low and high noise densities. The performance analysis against noise densities for lung CT scan image is also done with help of plot shown in Figs. 8-13.
Table 4.PSNR of Lung CT scan image for various filters at different noise densities
Table 5.IEF of Lung CT scan image for various filters at different noise densities
Table 6.Elapsed time of Lung CT scan image for various filters at different noise densities
Table 7.SSIM of Lung CT scan image for various filters at different noise densities
Fig. 8.Comparison of MSE value of different algorithms for Lung CT scan image
Fig. 9.Comparison of RMSE value of different algorithms for Lung CT scan image
Fig. 10.Comparison of MAE value of different algorithms for Lung CT scan image
Fig. 11.Comparison of PSNR value of different algorithms for Lung CT scan image
Fig. 12.Comparison of IEF value of different algorithms for Lung CT scan image
Fig. 13.Comparison of SSIM value of different algorithms for Lung CT scan image
The difference in detail preservation performance of the proposed method may be better observed by looking at the output images of the different methods . For this purpose, the Lung CT scan image shown in Fig. 14(a) of corrupted by 80% noise and noise image shown in Fig. 14(b) is restored by using different median based filters. Fig. 14(c)-14(h) shows the output images of the different methods. It is observed from this figure that performance of the SMF and AMF are very close to each other. Some noise mark are easily visible in the output images of these two filters. The output images of the DBA, PSMF and UTDBMF are almost indistinguishable from each other and they are significantly better than those of the SMF and AMF. They show very good noise removal performance but considerably blur the small details of the image. It is observed that the proposed algorithm yields much better preservation performance .
Fig. 14.(a) Original Lung CT scan image; (b) Corrupted image with 70% of salt-and-pepper noise; (c) SMF image; (d) AMF image; (e) DBA image; (f) PSMF image; (g) UTDBMF image and (h) Proposed filtered image.
The PSNR value of proposed filter is high compared to existing methods as shown in Fig. 10.
Fig. 15 show the reconstructed results of MRI scan brain corrupted with 80% of noise of various filters . From the restored results the proposed filter produce excellent quality image for high noise density. From the Table 6 and Fig. 16, we observe that the proposed filter provides better trade off among elapsed time.
Fig.15(a) Original MRI scan brain image; (b) 80% noise image; (c) Proposed filtered image; (d) Original CT scan brain image; (e) 80% noise image (f) Proposed filtered image
Fig. 16.Comparison of elapsed time value of different algorithms for Lung CT scan image
In this paper, a proposed fuzzy based median filtered algorithm has been implemented for enhancing the different medical images in terms of removal of salt-and-pepper noise from highly corrupted noisy image. The extracted local information for optimal performance is achieved by two-sided π-MF compare with other fuzzy membership function. Results show that the proposed algorithm improves the metric values in comparison with other existing algorithms in terms of higher PSNR, IEF and SSIM for noise densities from 10% to 90%. In addition, the moderately elapsed time and easy implementation for image denoising methods.