IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 22, NO.

3, MARCH 2013

1223

Efcient Improvements on the BDND Filtering Algorithm for the Removal of High-Density Impulse Noise
Iyad F. Jafar, Rami A. AlNamneh, and Khalid A. Darabkh

Abstract Switching median lters are known to outperform standard median lters in the removal of impulse noise due to their capability of ltering candidate noisy pixels and leaving other pixels intact. The boundary discriminative noise detection (BDND) is one powerful example in this class of lters. However, there are some issues related to the ltering step in the BDND algorithm that may degrade its performance. In this paper, we propose two modications to the ltering step of the BDND algorithm to address these issues. Experimental evaluation shows the effectiveness of the proposed modications in producing sharper images than the BDND algorithm. Index Terms Impulse noise, median lter, noise detection, switching median lters.

I. I NTRODUCTION

HANNEL transmission errors and faulty switching of acquisition devices may result in corrupting images with impulse noise. Pixels contaminated with impulse noise are characterized by having relatively low or high intensity values when compared to their neighboring pixels. This could dramatically affect images quality and possibly make them unsuitable for human or machine vision applications. In attempt to restore the original values of noisy pixels, ltering techniques are usually applied with the objective of suppressing the noise while minimizing the distortion introduced to the sharpness of edges and details in the original image. The standard median lter [1], which is a nonlinear order-statistic lter, is one of the most popular lters that is used in the removal of impulse. This fact triggered the development of several algorithms that build on the standard median lter to improve its performance. Examples include, but not limited to, weighted median lters [2], [3], center weighted median lters [4], recursive median lters [5], [6], and adaptive length median lters [7]. Despite the simplicity and effectiveness of these medianbased lters, they still result in unnecessary degradation in

Manuscript received February 8, 2012; revised October 30, 2012; accepted November 7, 2012. Date of publication November 20, 2012; date of current version January 30, 2013. The associate editor coordinating the review of this manuscript and approving it for publication was Dr. Farhan A Baqai. I. F. Jafar is with the Computer Engineering Department, University of Jordan, Amman 1192, Jordan (e-mail: iyad.jafar@ju.edu.jo). R. A. AlNamneh is with Department of Software Engineering at Jordan, University of Science and Technology, Irbid 22110, Jordan (e-mail: ramir11@just.edu.jo). K. A. Darabkh is with the Computer Engineering Department at the University of Jordan, Amman 1192, Jordan (e-mail: k.darabkh@ju.edu.jo). Color versions of one or more of the gures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identier 10.1109/TIP.2012.2228496

image quality [8][10]. This is related to the fact that ltering in such algorithms is applied to all pixels in the image and ignoring the fact that not all pixels are noisy. A simple solution to this problem is to apply these lters selectively such that only noisy pixels undergo the ltering operation. This requires using a detection mechanism to reliably identify candidate noisy pixels in the image. These lters are usually referred to as switching median lters [11][19]. The key issue in switching median lters is the detection of noisy pixels. Thresholding ltering [20] utilizes efcient noise detectors to reduce the misclassication of noise-free pixels. A trilateral lter [21] combined with an impulse detector that detects impulse noise according to the local image statistics was also proposed. In [22], noise detection incorporates edgedirected approach to preserve the details and edges. A fuzzy technique that is capable of detecting and removing impulse noise was proposed in [23]. The technique relies on longrange correlation within different parts of the image. The noise adaptive soft-switching median lter [24] applies an approach that utilizes the fuzzy-set concept to classify pixels into noisefree pixels, isolated noisy pixels, non-isolated noisy pixels, and edge pixels. The lter applies the identity lter, standard median lter, or a new class of fuzzy lter to lter the pixel depending on the class it belongs to. Reference [25] presents an adaptive hierarchal lter that can remove impulse noise while preserving the details of the image by inferring the global structure of the image from a set of pyramid images that are used as prior information in order to apply different lters adaptively. Generally, the proposed algorithms produce satisfactory results; however, they still tend to remove ne details of the image and fail to detect some of the noisy pixels, especially when the noise density is high. In fact, most of the presented algorithms are tested under moderate noise densities (< 50%). A quite interesting switching median lter is the boundary discriminative noise detection (BDND) lter that is proposed in [26]. The BDND lter is proven to operate efciently when compared to other lters, even under high noise densities (up to 90%). Being a switching-based median lter, the BDND algorithm lters the noisy image in two steps. The rst step is essentially a noise detection step which is based on clustering the pixels in the image in a localized window into three groups, namely; lower intensity impulse noise, uncorrupted pixels, and higher intensity impulse noise. The clustering is based on dening two boundaries using the intensity differences in the

10577149/$31.00 2012 IEEE

1224

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 22, NO. 3, MARCH 2013

ordered set of the pixels in the window. The pixel is classied as uncorrupted if it belongs to the middle cluster. Otherwise it is a noisy pixel. This noise detection mechanism showed impressive detection accuracy under different impulse noise models [26]. Once the noise map is determined, the second step is the ltering step, which is supposed to replace the noisy pixel with an estimate of its original value. This step is applied on the identied noisy pixels only. The ltering is essentially a median ltering operation that is applied on the uncorrupted pixels found in the ltering window. The critical parameter that is required to be dened in the ltering step of the BDND algorithm is the size of the ltering window. The size of this window is determined as follows. A window of size 3 3 is used as initial size for the ltering window. If the number of uncorrupted pixels in the window is less than half the window size, then the window is expanded outward by one pixel in all directions. This is repeated until the number of uncorrupted pixels in the window is greater than or equal half the number of pixels in the window or the current window size is less than or equal a maximum window size. The maximum window size of the condition is ignored and the window is expanded if no uncorrupted pixels are found. In this case, window expansion is repeated until one uncorrupted pixel is found. Basically, this step is an adaptation from the ltering process proposed in [24] and is reported to perform well even under high noise densities. The BDND lter is proven to operate effectively under different impulse noise models. However, two main observations can be made about its ltering step. First, expanding the window until the number of uncorrupted pixels is at least half the number of pixels in the window may impose additional blurring in the output image. The impact of this is clearly noticeable under high noise densities. Second, the ltering step relies on computing the median value of the uncorrupted pixels found in the window without any regard to the spatial relationship of these pixels to the noisy pixel, and the deviation of the pixels intensities from the median value. This also affects the quality and the sharpness of the edges in the ltered image. In this paper, we propose two modications to the ltering step in the BDND algorithm in order to improve its performance. The rest of the paper is organized as follows. The proposed modications are explained in Section II. Comprehensive evaluation of the proposed modications is presented in Section III. Finally, the paper is concluded in Section IV. II. P ROPOSED M ODIFICATIONS In this section, we discuss the modications introduced to the ltering step of the BDND algorithm. These modications are based on some observations related to the BDND and the median lter itself as outlined earlier. The rst modication is based on loosening the condition imposed on expanding the ltering window. The second modication is proposed to incorporate the spatial information of uncorrupted pixels in the ltering window and the deviation of their intensities from the median when computing the estimated value of the noisy

20 22 21 21 22 * 22 * 21 * * 22 21 21 *

20 21 22 20 150 20 * * 20 * 20 * * * *

23 18 20 152 151
(a)

22 19 151 150 149 * * 151 * * * * * * 149

21 152 152 153 153 * 152 152 * 153 * 152 152 153 153

* 18 * * 151
(b)

23 * * * *
(c)

Fig. 1. (a) Hypothetical image and two noisy versions. (b) Noisy image I. (c) Noisy image II. Noisy pixels are indicated by .

pixel. These are explained through examples in the following subsections. A. Expansion of Filtering Window The rst observation is related to the way ltering is performed in the BDND algorithm which starts by using a 3 3 ltering window that is centered on the noisy pixel. However, the size of this window is considered insufcient for ltering under two conditions: i) the number of uncorrupted pixels Nu is less than half of the number of pixels in the window Nh , where Nh = 1/2 (WF WF ) and WF is the window width, ii) if the number of uncorrupted pixels is zero. In case any of these conditions is violated for the current window, the window is expanded outward by one pixel in all directions. For the rst condition, expansion is allowed as long as the size of the window is less than or equal to a maximum window size of Wmax Wmax . Such approach in expanding the ltering window could be useful in providing a better estimate for the value of the noisy pixel. However, the strict condition of requiring the number of uncorrupted pixels to be greater than half the number of pixels in the window is easily violated under high noise densities. Thus, with high noise densities the ltering window is expected to be expanded and most likely it will reach the maximum size. The direct impact on increasing the window size is the possible loss of correlation between the pixel values inside the ltering window. This may directly affect the value that replaces the noisy pixel, which may lead to blurring and unnecessary distortion in the ltered image. To demonstrate the idea, consider the 5 5 image shown in Fig. 1(a), which contains an edge along the 45 diagonal that separates between two smooth regions. Suppose that this image is corrupted with 60% impulse noise as shown in Fig. 1(b), with the noisy pixels indicated by (*). These noisy pixels are assumed to be detected reliably by the detection step of the BDND algorithm. When the ltering step of the BDND

JAFAR et al.: EFFICIENT IMPROVEMENTS ON THE BDND FILTERING ALGORITHM

1225

algorithm is applied on the central pixel with a 3 3 window, then the set of uncorrupted pixels Vu in the window is simply {18, 20, 151}. This implies that Nu is 3, which is less than half the number of pixels in the window (Nh = 4.5). This violates the rst condition that is imposed on the size of the ltering window, since Nu is less than Nh . If Wmax is set to 3, then the ltering window is expanded to 5 5 since WF equals Wmax . Given this new window, the set of uncorrupted pixels Vu is {18, 20, 20, 21, 22, 151, 151, 152, 152, 153}. This means that the number of uncorrupted pixels Nu is 10, which still less than half the window size (Nh = 12.5). Thus, the condition is violated again. However, the current window size is greater than Wmax , so window expansion stops. Of course, the second condition for window expansion is false since the number of uncorrupted pixels is not zero. Consequently, the ltered value of the pixel under consideration Xij is replaced by a new value Yij , which is simply the median of the uncorrupted pixels found in the ltering window using Yi j = median{ X i s , j t |(s , t ) W X i s , j t Vu } where W = {(s , t )|(W F 1)/2 s , t (W F 1)/2}. (2) (1)

For the example image given in Fig. 1(b), this implies that the output value for the center pixel after ltering is 86, which is far away from the original pixel value. Additionally, this implies that the edge position between the two regions is displaced by one pixel. The main reason for such a problem in the BDND algorithm results from the condition imposed on expanding the ltering window which requires the number of uncorrupted pixels to be at least half the window size. However, such requirement is hard to satisfy under high noise densities. Actually, expanding the window in such circumstance may not solve the problem since the number of required uncorrupted pixels increases nonlinearly when the window is expanded. As a result, the maximum window size is usually reached with high noise densities, which in turn results in additional blurring and edge displacement in the ltered image. In order to address this problem, we propose the following modication to the window expansion process in the BDND algorithm. Basically, the condition is modied to take into consideration the estimated noise density P that is determined from the detection step of the algorithm and the total number of pixels NT in the ltering window, such that while the number of uncorrupted pixels Nu is less than 1/2(1P)NT and WF is less than or equal to Wmax , then the ltering window is expanded by one pixel outward in all directions. The term (1-P) basically is the percentage of uncorrupted pixels that are expected to be found in the ltering window. Including this term in the condition makes it adaptive to the noise density. In other words, when the noise density increases, the condition is loosened since the expected number of uncorrupted pixels decreases. This in turn reduces the occasions of window expansion. This is unlike the BDND algorithm that uses a xed threshold of 1/2(WF WF ) regardless of the noise density. As a matter of fact, this is hard to achieve with high noise densities and small windows.

The use of the total number of pixels instead of 1/2(W F W F ) makes it explicit that windows centered at boundary pixels in the image utilize pixels available in the window instead of replicating boundary pixels to match the size of the window. This issue was not explicitly discussed in the BDND algorithm. In fact, the way boundary pixels are treated is crucial and signicantly affects the process of expanding the ltering window. In general, the modication outlined earlier is expected to reduce the cases where the window is expanded unnecessarily, which in turn results in ltering windows containing values that are likely to be more correlated to the original value of the noisy pixel. This is based on the fact that there is a strong correlation between pixels values and their spatial locations in small neighborhoods in images [1]. In other words, if a set of uncorrupted pixels is found in the ltering window, even with small count, then it is more logical to use their values in the ltering operation since they are spatially closer to the central pixel (the noisy pixel) than those uncorrupted pixels found in the expanded ltering window. This would imply a stronger correlation between the pixel values in the small window and the original value of the noisy pixel than the case for the pixel values in the larger window. Once the ltering window is determined based on the condition discussed previously, ltering is performed using (1). However, since expanding the ltering window is now adaptive to the noise density, the occasions of over expanding the window in the modied condition are less likely to occur. In case the ltering window size reaches the maximum allowed value and no uncorrupted pixels are found in the window, then, and similar to BDND, the second condition is invoked and the window is expanded until at least one pixel is found. In the example presented earlier, when the window size is 3 3, the set of uncorrupted pixels Vu is {18, 20, 151} and the noise density is 60%. This implies that Nu is 3, NT is 9, and the percentage of uncorrupted pixels is 40%. Based on these values, the proposed modication on the window expansion condition is false since Nu is greater than 1/2(1P)NT , (3 > 1.8), thus the window is not expanded. Accordingly, the new pixel value Yij is basically the median of Vu , which is 20. Comparing this value with that obtained using BDND, 86, we can see that this case is more consistent with original pixels value. This in turn implies that the edge location in the ltered and the original images is located at the same place. B. Incorporating Spatial and Intensity Information Actually, there is another observation that can be made about the performance of the BDND ltering step and other median-based lters which rely on replacing the noisy pixels with the median of uncorrupted pixels in the ltering window. Consider the noisy image shown in Fig. 1(c), which is basically a different noisy version of the image shown in Figure 1(a) with the noise density being 60%. In this example, if the central pixel is to be replaced, then a 3 3 ltering window would be insufcient for both the BDND algorithm and the modication proposed in the previous subsection since Nu is zero. Thus, in both cases, the window size is expanded

1226

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 22, NO. 3, MARCH 2013

(a)

(b)

(c)

(d)

Fig. 2. Experimental results for different images. (a) Original images. (b) Original images corrupted by 80% impulse noise. (c) Proposed approach. (d) BDND.

to 5 5. In this window, Vu is {20, 21, 21, 22, 23, 149, 152, 152, 153, 153}. This makes the number of uncorrupted pixels 10. For the BDND algorithm, if Wmax is assumed to be 3, then window expansion stops. Similarly, window expansion is stopped using the modied condition discussed previously

since Nu is greater than 1/2(1P)NT , (10 > 5). Now, for both cases, the output value Yij is 86, which is the median of Vu . Note how this value is not consistent with the original pixel value. The reason behind this large deviation of the output value from the expected value is that the median lter is based

JAFAR et al.: EFFICIENT IMPROVEMENTS ON THE BDND FILTERING ALGORITHM

1227

45 Proposed BDND 40

45 Proposed BDND 40

45 Proposed BDND 40

PSNR (dB)

35

PSNR (dB)

35

PSNR (dB) 10 20 30 40 50 60 70 Noise Density % 80 90

35

30

30

30 25 25

20

10

20

30

40 50 60 70 Noise Density %

80

90

20

25

10

20

30

40 50 60 70 Noise Density %

80

90

(a)
50 45 40 PSNR (dB) PSNR (dB) Proposed BDND 35 40

(b)
45 Proposed BDND 40

(c)
Proposed BDND

PSNR (dB) 10 20 30 40 50 60 70 Noise Density % 80 90

35

35 30

30

30

25 25 20 20 25

10

20

30

40 50 60 70 Noise Density %

80

90

20

10

20

30

40 50 60 70 Noise Density %

80

90

(d)
Fig. 3.

(e)

(f)

PSNR values for different test images. (a) Lena . (b) Boat. (c) Peppers. (d) Cameraman. (e) Baboon. (f) Toys.

on ranking the values of the uncorrupted pixels found in the ltering window without any regard to the deviation of the pixels values from the median value and the spatial relation of the uncorrupted pixels to the noisy pixel. In order to address this issue, the output value Yij , which is essentially the value that is in the middle of sorted version of Vu , has to be either decreased or increased by considering the relative difference between the pixel values and the median value, and by exploiting the spatial relationship between pixels in the ltering window. Actually, this problem might be serious in case the number of uncorrupted pixels is even as illustrated in the previous example. To deal with this issue, we propose the following modication to the computation of the estimated pixel value in the ltering window. As outlined previously, the values of pixels in small regions in images tend to be correlated to some degree. In other words, pixels that are close to each other tend to have close values, especially in smooth regions. Accordingly, the proposed adjustment of Yij incorporates the spatial relation between the noisy pixel and the uncorrupted pixels in the window and the relation between the values in Vu and the Yij . Basically, the adjustment corresponds to adding a term to Yij , which is sum of deviations of the values in Vu from Yij weighted by a factor that is inversely proportional to the distance of the uncorrupted pixels to the noisy pixel under consideration, i.e. the center of the window. In other words, the adjusted output value Zij is given by Z i j = Yi j + 1 D
Nu k =1

where d(k) is the spatial distance between the pixels of the kth value in Vu and the noisy pixel at location (i,j) d (k ) = | S (k ) i | + | T (k ) j | (4)

with S(k) and T(k) are the row and column indices of the pixel, and
Nu

D=
m =1

1 . d (m )

(5)

Vu (k ) Yi j d (k )

(3)

Note how the adjustment term could be positive or negative. Let Vu1 be the set of values in Vu that are less than Yij while Vu2 be the set of values that are greater than Yij . In case the total difference of values in Vu2 from Yij is greater than the total difference of the values in Vu1 from Yij and they are spatially closer to the noisy pixel, then the adjustment term will be positive. This implies that Yij is increased to be more coherent with the values in Vu2 since they are spatially closer to the noisy pixel. On the other hand, if the difference of the values in Vu1 from Yij is greater than that of the values in Vu2 and their spatial coordinates are closer to the noisy pixel, then the adjustment term will be negative. This implies that Yij is decreased to be more coherent with the values in Vu1 . So, the median value Yij is adjusted based on the intensity differences as well as the spatial relationship between pixels in the ltering window. For the noisy image given in Fig. 1(c), the median set of uncorrupted pixels Vu is {20, 21, 21, 22, 23, 149, 152, 152, 153, 153} with a median value of 86. Accordingly, the sets Vu1 and Vu2 are {20, 21, 21, 22, 23} and {149, 152, 152, 153, 153}, respectively. The corresponding distances for

1228
6 7

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 22, NO. 3, MARCH 2013

6

9.5 9 8.5 8 7.5 7 6.5 6

x 10

1.3 1.2 1.1

x 10

x 10

6.5 TEN 1 0.9 Original Proposed BDND 10 20 30 40 50 60 70 Noise Density % 80 90 0.8 0.7 Original Proposed BDND 10 20 30 40 50 60 70 Noise Density % 80 90 5.5 TEN 6

TEN

Original Proposed BDND

10

20

30

40 50 60 70 Noise Density %

80

90

(a)
x 10 10 1.8 9.5 9 TEN TEN 8.5 8 7.5 7 6.5 10 Original Proposed BDND 20 30 40 50 60 70 Noise Density % 80 90 1.6
6

(b)
x 10
7

(c)
x 10 9 8.5 8
6

1.4 1.2 1

TEN Original Proposed BDND

7.5 7 6.5 6 Original Proposed BDND 10 20 30 40 50 60 70 Noise Density % 80 90

10

20

30

40 50 60 70 Noise Density %

80

90

(d)
Fig. 4.

(e)

(f)

TEN values for different test images. (a) Lena . (b) Boat. (c) Peppers. (d) Cameraman. (e) Baboon. (f) Toys. TABLE I S UGGESTED F ILTERING W INDOW S IZE BASED ON N OISE D ENSITY P Noise Density 0% < P 20% 20% < P 40% P > 40% Wmax Wmax 33 55 77

the pixels in these two sets from the center pixel are {3, 2, 3, 3, 2} and {3, 3, 4, 2, 3}, while the deviation from the median value are {66, 65, 65, 64, 63} and {63, 66, 66, 67, 67}, respectively. Note the closer distances and lower deviations of pixels values in Vu1 when compared to those in Vu2 . This implies that the adjusted median Zij value would be pulled toward the intensities of the upper region in the image according to (3). Actually, the adjusted median value using (3) is 81 while it is 86 for the case of the BDND algorithm. Such adjustment is expected to reduce the distortion introduced into the ltered image as discussed in Section III. III. E XPERIMENTAL R ESULTS In this section, we present some of the results obtained by the proposed modications and compare them with those of the original ltering step of the BDND algorithm [26]. The results discuss the effectiveness of the proposed modications over the BDND algorithm for both monochrome and color images. The comparison and discussions are based on visual evaluation and different objective measures. Also, we investigate and discuss the effect of the new modication on the size of ltering window under different noise densities. In all experiments, it is assumed that noisy pixels are identied ideally by the detection step of the BDND algorithm. This is essential in order to put aside any issues related to the detection since our focus is on the performance improvements achieved by the modications introduced to the ltering step. In other words, the input to the original and proposed ltering steps is basically the noise map that is obtained when the noise is randomly generated. This completely neutralizes any effect

of the detection step, such as its dependence on the noise model [26], on the evaluation of the ltering step. However, when noisy images are displayed, the salt-and-pepper noise model is assumed. Regarding the maximum size of the ltering window, we use the suggested values in [26] which are shown in Table I. A. Performance Metrics To aid the subjective evaluation of the results, we utilize different performance metrics that are commonly used in this regard. The rst metric is the Peak-Signal-to-Noise-Ratio (PSNR) which is dened by PSNR = 10 log10 2552
1 MN M N i =1 j =1

(Zij Xij

)2

(6)

where M and N are the number of rows and columns in the image, respectively. PSNR values are usually used to measure the similarity between two images with higher values indicating higher similarity or lower distortion. To evaluate

JAFAR et al.: EFFICIENT IMPROVEMENTS ON THE BDND FILTERING ALGORITHM

1229

1 0.95 0.9 0.85 FOM FOM 0.8 0.75 0.7 0.65 10 20 30 40 50 60 70 Noise Density % 80 90 Proposed BDND

1 0.95 0.9 0.85 0.8 0.75 0.7 0.65 10 20 30 40 50 60 70 Noise Density % 80 90 FOM Proposed BDND

1 0.95 0.9 0.85 0.8 0.75 0.7 Proposed BDND

10

20

30

40 50 60 70 Noise Density %

80

90

(a)
1 Proposed BDND 0.95
1 0.95 0.9 0.85

(b)
1 Proposed BDND 0.95 0.9 FOM 0.85 0.8 0.75 0.7 0.65

(c)
Proposed BDND

FOM

FOM

0.9

0.8 0.75

0.85

0.7 0.65

0.8

10

20

30

40 50 60 70 Noise Density %

80

90

10

20

30

40 50 60 70 Noise Density %

80

90

10

20

30

40 50 60 70 Noise Density %

80

90

(d)
Fig. 5.

(e)

(f)

FOM values for different test images. (a) Lena . (b) Boat. (c) Peppers. (d) Cameraman. (e) Baboon. (f) Toys.

the distortion in color images, we use the CIELAB color space as it provides more accurate quantitative measurements of perceptual error between two color images [27]. Thus, the original and processed color images are converted to CIELAB color space and then the difference of colors is computed using
E ab =

TABLE II C OUNTS OF THE F ILTERING W INDOW S IZE FOR D IFFERENT N OISE D ENSITY P P = 10% 33 55 77 99 11 11 13 13 25 994 108 0 0 0 0 25 994 108 0 0 0 0 P = 50% 112 164 17 655 1109 36 0 0 47 784 22 642 12 973 47 565 0 0 P = 90% BDND 157 0 0 235 557 68 1 134 075 49 237 31 909 20 493 68 1

W F W F Proposed BDND Proposed BDND Proposed

( L )2 + ( a )2 + ( b )2 .

(7)

The remaining metrics are related to the edge quality. The Tenegrad (TEN) measure [28] is used to evaluate the strength of the edges in the ltered image. The Tenegrad is a wellknown benchmark image sharpness measure which is based on gradient magnitude maximization and is considered one of the most robust and functionality accurate image quality measures. It is computed from the gradient at each pixel in the image where the partial derivatives are obtained by a highpass lter such as the Sobel operator. For the whole image, the TEN is computed using TEN =
i j

i j ,

i j >

(8)

where nd and nt are the number of edge pixels detected in the original and output images, respectively, dk is the distance from the kth edge pixel in the output image to the closest possible edge pixel in the original image. The term alpha is usually set to 1/9. FOM value is unity for ideal edge match and it decreases as the miss-localization of edges increases. B. Size of Filtering Window One of the proposed modications introduced in this paper is related to the condition used to dene the size of the ltering window. Unlike the condition dened in the BDND algorithm, the modied condition is made adaptive to the noise density. This has the effect of reducing the occasions in which the window is expanded unnecessarily. Table II lists the counts of the ltering window size used in the BDND and proposed modication for the image Lena

where ij is the gradient magnitude and is a threshold that is used to eliminate low edge responses which usually correspond to low levels of noise in the image. In our experiments, is chosen to be the mean value of the gradient magnitude. The gure of merit (FOM) metric proposed by Pratt [29] is also used to quantify the distortion introduced into edges in terms of edge localization. The FOM is dened by FOM = 1 max{n d , n t }
nd k =1

1 2 1 + dk

(9)

1230

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 22, NO. 3, MARCH 2013

(a)
Fig. 6.

(b)

(c)

(d)

Filtering results for color images. (a) Original. (b) Original corrupted with 80% impulse noise. (c) Proposed. (d) BDND.

(shown in Fig. 2) when it is corrupted by 10%, 50%, and 90% impulse noise. Note that with 10% noise density the counts for both ltering schemes are the same. However, as the noise density increases, the proposed condition has lower counts for larger windows. This is clear under 90% noise density, where the average window size over the entire image is 3.74 for the proposed modication and 5.99 for the BDND algorithm. As mentioned before, this is because the BDND condition is easily violated with high noise densities, which in turn forces the window to expand. For windows greater than (Wmax + 1 Wmax + 1), these are used in the two approaches when the number of uncorrupted pixels in the window is zero. This would happen in high noise densities since noise blotches could present due to the irregularity in the noise distribution in the image. In this case, the behavior of BDND and the proposed condition is the same. Remember here that the condition for expansion is applied as long as the window size is less than or equal to Wmax . C. Monochrome Images For the purpose of evaluating the proposed modications on monochrome images, several tests were conducted. In this section we present six examples, which are: Lena, Boat, Peppers, Cameraman, Baboon, and Toys. Fig. 2 shows

the original test images and the noisy version of each that is corrupted by 80% impulse noise (more results for other noise densities are available on the following website http://driyad.ucoz.net/index/ieee_results/0-39). The ltering results are also shown in Fig. 2. Inspecting these results reveals the power of the two approaches in restoring the original images efciently. However, careful investigation of the results proves the capability of the proposed approach in producing higher quality images when compared to the BDND algorithm. This is supported by the PSNR values that are shown in Fig. 3 under noise densities ranging from 10% to 90%. From these values, we can see how the performance of the two approaches is close to each other under low noise densities. This is due to the fact that the size of the ltering window in both approaches is almost the same in this case as discussed in the previous subsection. As a matter of fact, higher PSNR values for the proposed modications under low densities are strongly associated with the second modication explained in Section II. Regarding the strength and localization of edges, it is evident in Fig. 3 how the images produced by the proposed modications are visually sharper and more distinctive than those obtained using the BDND algorithm. Fig. 4 shows the TEN values for the original and processed images. The proposed modications produced higher values which imply

JAFAR et al.: EFFICIENT IMPROVEMENTS ON THE BDND FILTERING ALGORITHM

1231

3 2.5 2

x 10

8 Proposed BDND 7 6 5
Eab

x 10

7 Proposed BDND 6 5
Eab

x 10

Proposed BDND

* Eab

4 3 2 1 0

1.5 1 0.5 0

4 3 2 1

10

20

30

40 50 60 70 Noise Density %

80

90

10

20

30

40 50 60 70 Noise Density %

80

90

10

20

30

40 50 60 70 Noise Density %

80

90

(a)
Fig. 7.

(b)
values under different noise densities for color images. (a) Family. (b) House. (c) Plane. E ab

(c)

stronger edges. This is consistent with the perceived sharpness for the ltered images in Fig. 3. Note how the TEN values for the proposed modications are closer to the TEN values of the original images and while the TEN values of the BDND algorithm drops dramatically as the noise density increases. Moreover, Fig. 5 shows that the proposed modications resulted in higher FOM values compared to the BDND algorithm. This reects their capability of preserving edge locations better than the original BDND ltering algorithm.

IV. C ONCLUSION Switching median lters are effective in removing impulse noise as they are applied to candidate noisy pixel only. The BDND lter is one of the popular switching median lters as it is proven to overtake other lters especially at high noise densities. However, the ltering step imposes a strict condition on the size of the ltering window that does not take into account the noise density. Additionally, it does not consider the spatial relationship and deviation of the pixels intensities in the ltering window from the central pixel and the median value of the window. In this paper, we proposed two modications to the BDND ltering step to alleviate the effect of these problems on the quality of the ltered image. The modications basically loosen the condition imposed on expanding the ltering window and incorporate the spatial information of the pixels in the ltering process. Experimental evaluation showed the effectiveness of these modications on improving the performance of the BDND algorithm. R EFERENCES
[1] R. C. Gonzalez and R. E. Woods, Digital Image Processing. Upper Saddle River, NJ: Prentice Hall, 2002. [2] D. R. K. Brownrigg, The weighted median lter, ACM Commun., vol. 27, no. 8, pp. 807818, Aug. 1984. [3] R. Yang and L. Yin, Optimal weighted median ltering under structural constrains, IEEE Trans. Signal Process., vol. 43, no. 3, pp. 591604, Mar. 1995. [4] S. J. Ko and Y. H. Lee, Center weighted median lters and their applications to image enhancement, IEEE Trans. Circuits Syst., vol. 38, no. 9, pp. 984993, Sep. 1991. [5] G. Qiu, An improved recursive median ltering scheme for image processing, IEEE Trans. Image Process., vol. 5, no. 4, pp. 646648, Apr. 1996. [6] M. P. McLoughlin and G. R. Arce, Deterministic properties of the recursive separable median lter, IEEE Trans. Acoust., Speech, Signal Process., vol. 35, no. 1, pp. 98106, Jan. 1987. [7] H. M. Lin and A. N. Willson, Median lters with adaptive length, IEEE Trans. Circuits Syst., vol. 35, no. 6, pp. 675690, Jun. 1988. [8] A. C. Bovik, T. S. Huang, and D. C. Munson, Edge-sensitive image restoration using order-constrained least squares methods, IEEE Trans. Acoust., Speech, Signal Process., vol. 33, no. 10, pp. 12531263, Oct. 1985. [9] T. A. Nodes and N. C. Gallagher, The output distribution of median type lters, IEEE Trans. Commun., vol. 32, no. 5, pp. 532541, May 1984. [10] A. C. Bovik, Streaking in median ltered images, IEEE Trans. Acoust., Speech, Signal Process., vol. 35, no. 10, pp. 493503, Oct. 1985.

D. Color Images For the purpose of evaluating the performance of the proposed modications on color images, P% of pixels in each color channel are assumed to be contaminated with noise. As mentioned at the beginning of this section, the ideal noise map of each channel is used to identify noisy pixels, however, when the images are displayed, salt-and-pepper noise model is used. This is essential to study the performance of the ltering step only. Filtering of color images is done using the scalar approach instead of the vector approach. That is, each RGB channel is ltered separately. It was shown in [26] that scalar approach is more efcient that the vector approach since the window size in the vector approach will result in excessive window expansion since uncorrupted pixels are searched in the three color channels for any ltering window. The results for three color images that are corrupted by 80% impulse noise are presented here, namely: Family, House, and Plane. The original and noisy images as well as the ltering results for the BDND and the proposed modications are shown in Fig. 6 (color results are available on the following website http://driyad.ucoz.net/index/ieee_results/0-39). Comparing the ltering results of the two approached reveals the effectiveness of the proposed approach in producing higher quality and sharper images. For further investigation, Fig. 7 for the two approaches and when illustrates the values of E ab the original images are corrupted by noise densities ranging from 10% to 90%. From these numbers, it is clear how the proposed modications result in lower distortion in the colors when compared to the BDND numbers.

1232

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 22, NO. 3, MARCH 2013

[11] T. Chen, K. K. Ma, and L. H. Chen, Tri-state median lter for image de-noising, IEEE Trans. Image Process., vol. 8, no. 12, pp. 18341838, Dec. 1999. [12] T. Chen and H. R. Wu, Impulse noise removal by multi-state median ltering, in Proc. Int. Conf. Acoust., Speech, Signal Process., Jun. 2000, pp. 21832186. [13] S. Zhang and M. A. Karim, A new impulse detector for switching median lters, IEEE Signal Process. Lett., vol. 9, no. 11, pp. 360363, Nov. 2002. [14] Z. Wang and D. Zhang, Progressive switching median lter for the removal of impulse noise from highly corrupted images, IEEE Trans. Circuits Syst. II, Analog Digit. Signal Process., vol. 46, no. 1, pp. 7880, Jan. 1999. [15] T. Sun and Y. Neuvo, Detail-preserving median based lters in image processing, Pattern Recognit. Lett., vol. 15, no. 4, pp. 341347, 1994. [16] D. Florencio and R. Schafer, Decision-based median lters using local statistics, in Proc. Int. Symp. Vis. Commun. Image Process. Conf., 1994, pp. 264268. [17] P. S. Windyga, Fast impulsive noise removal, IEEE Trans. Image Process., vol. 10, no. 1, pp. 173179, Jan. 2001. [18] S. Zhang and M. A. Karim, A new impulse detector for switching median lters, IEEE Signal Process. Lett., vol. 9, no. 11, pp. 360363, Nov. 2002. [19] I. Aizenberg and C. Butakoff, Effective impulse detectors based on rank-order criteria, IEEE Signal Process. Lett., vol. 11, no. 3, pp. 363366, Mar. 2004. [20] I. Aizenberg, C. Butakoff, and D. Paliy, Impulsive noise removal using threshold Boolean ltering based on the impulse detecting functions, IEEE Signal Process. Lett., vol. 12, no. 1, pp. 6366, Jan. 2005. [21] K. Kondo, M. Haseyama, and H. Kitajima, An accurate noise detector for image restoration, in Proc. Int. Conf. Image Process., Sep. 2002, pp. 321324. [22] R. Garnett, T. Huegerich, C. Chui, and W. He, A universal noise removal algorithm with an impulse detector, IEEE Trans. Image Process., vol. 14, no. 11, pp. 17471754, Nov. 2005. [23] D. Zhang and Z. Wang, Impulse noise detection and removal using fuzzy techniques, Electron. Lett., vol. 33, no. 5, pp. 378379, Feb. 1997. [24] H. L. Eng and K. K. Ma, Noise adaptive soft-switching median lter, IEEE Trans. Image Process., vol. 10, no. 2, pp. 242251, Feb. 2001. [25] T.-N. Lin and K.-J. Cahn, Adaptive-hierarchical ltering for noise removal, Displays, vol. 29, pp. 209213, Aug. 2007. [26] P. E. Ng and K. K. Ma, A switching median lter with boundary discriminative noise detection for extremely corrupted images, IEEE Trans. Image Process., vol. 15, no. 6, pp. 15061516, Jun. 2006. [27] M. Tkalcic and J. F. Tasic, Color spaces: Perceptual, historical and application background, in Proc. IEEE EUROCON Conf., Sep. 2003, pp. 304308. [28] A. Buerkle, F. Schmoeckel, M. Kiefer, B. P. Amavasai, F. Caparrelli, A. N. Selvan, and J. R. Travis, Vision-based closed-loop control of mobile microrobots for micro handling tasks, Microrobot. Microassem. III, vol. 1, pp. 187198, Oct. 2001. [29] W. K. Pratt, Digital Image Processing. New York: Wiley, 2001.

Iyad F. Jafar received the B.Sc. degree in electrical engineering from the University of Jordan, Amman, Jordan, in 2001, the M.Sc. degree in electrical engineering from the Illinois Institute of Technology, Illinoise, in 2004, and the Ph.D. degree in computer engineering from Wayne State University, Michigan, in 2008. He is currently an Assistant Professor with the Department of Computer Engineering, University of Jordan. His current research interests include signal and image processing, pattern recognition, and multimedia and computer networks.

Rami A. AlNamneh received the B.S. degree in computer engineering from the Jordan University of Science and Technology, Amman, Jordan, in 2000, the M.S. and Ph.D. of Philosophy degrees from the University of Alabama in Huntsville, Huntsville, in 2003 and 2006, respectively. He is an Assistant Professor with the Department of Software Engineering, Jordan University of Science and Technology. His current research interests include parallel programming, signal processing, and wireless communications.

Khalid A. Darabkh received the Ph.D. degree in computer engineering from the University of Alabama, Huntsville, Alabama, in 2007. He is currently an Associate Professor with the Computer Engineering Department, University of Jordan, Amman, Jordan. His current research interests include wireless and mobile communications, queuing theory, trafc management, multimedia systems and networking, congestion control architectures and resource allocation, parallel computing, and pattern recognition. Dr. Darabkh is a member of many honor societies, such as Phi Kappa Phi, Tau Beta Pi, Eta Kappa Nu, and Sigma Xi.