7th International Scientific and Expert Conference TEAM 2015 Technique, Education, Agriculture & Management Belgrade, October 15-16, 2015 ADAPTIVE CENTER WEIGHTED MEDIAN FILTER Vedran Novoselac* and Zlatko Paviฤ Mechanical Engineering Faculty in Slavonski Brod, J. J. Strossmayer University of Osijek, Croatia * Corresponding author e-mail: vedran.novoselac@sfsb.hr Abstract In this paper, the weighted median is presented and its propertie distance to center data. This relation is implemented in the image denoising filter. In the paper impulse image noise model is considered. The quality of reconstructed images provided with proposed filter are measured with PSNR metric. Keywords: Weighted median, image processing, impulse noise, PSNR 1. Introduction The weighted medina [1] is a natural extension of the medina. While the median of data vector ๐ = [๐ฆ1 , … , ๐ฆ๐ ]๐ , ๐ฆ๐ ∈ ๐น, minimizes the ๐ฟ1 -distance (or least absolute deviation, LAD), the weighted median of ๐ = [๐ฆ1 , … , ๐ฆ๐ ]๐ with positive corresponding weight vector ๐ = [๐ค1 , … , ๐ค๐ ]๐ , minimizes the weighted ๐ฟ1 -distance adaptive center weighted median filter (ACWMF) is constructed and tested on various experimental images. Proposed algorithm has been developed under the PSNR criteria. 2. Weighted median and its application In this section is given a solution of the weighted median problem [1]. The following theorem proofs minimization problem of the weighted ๐ฟ1 -distances. Theorem 1 Let ๐ฆ(1) ≤ โฏ ≤ ๐ฆ(๐) denote the ordered observation and ๐ค(1) , … , ๐ค(๐) the corresponding positive weights. Then the weighted median of ๐ = [๐ฆ1 , … , ๐ฆ๐ ]๐ is med(๐, ๐) = ๐ฆ(๐+1) , where ๐ l = max {โ: ∑ ๐ค(๐) ๐=1 ๐ฅ (2) ๐=1 Proof. Let ๐น: ๐น → ๐น be a function defined as ๐ ๐น(๐ฅ) = ∑ ๐ค๐ |๐ฆ๐ − ๐ฅ|. ๐ med(๐, ๐) = argmin ∑ ๐ค๐ |๐ฆ๐ − ๐ฅ|. ๐ 1 < ∑ ๐ค๐ }. 2 (3) ๐=1 (1) Notice that on each interval ๐=1 If ๐ค1 = โฏ = ๐ค๐ = 1, the global minimum is denoted by med(๐) and called the median of data vector. In case that the number of observation ๐ is an odd number (๐ = 2๐ − 1), the distance between center weight median and center observation ๐ฆ๐ can be observed. It is shown that distance between center weight median and ๐ฆ๐ decrease as the ๐ฅ increase. The weighted median problem is used in many methods for outlier detection [4] and can be found in various branches of applied research (robotics, signal and image processing [2,3], etc.). In image processing the reconstruction of noise image is a problem which can be solved with different methods. For that purpose the class of stack filter is developed. Standard median filter removes impulse noise and preserve image edges [2]. However, median filter has detail preserving difficulties. Weighted median is useful because of their flexibility. In weighted median filter weights are used to preserve image details and suppress noise as well [3]. For that purpose the new filter represented by thresholds are developed based on relation between center weight median and distance to center observation. In that case an (−∞, ๐ฆ(1) ), [๐ฆ(1) , ๐ฆ(2) ), … , [๐ฆ(๐−1) , ๐ฆ(๐) ), (๐ฆ(1) , ∞), (4) ๐น is a linear function and slopes of those linear function are consecutively ๐โ , โ = 0, … , ๐, where ๐ ๐ ๐0 = − ∑ ๐ค๐ , ๐๐ = ∑ ๐ค๐ , ๐=1 (5) ๐=1 and for โ = 1, … , ๐ − 1 โ ๐ ๐0 = 2 ∑ ๐ค(๐) − ∑ ๐ค๐ = ๐โ−1 + 2๐ค(โ) . ๐=1 (6) ๐=1 Since ๐โ+1 − ๐โ = 2 ๐ค(โ+1) > 0, the sequence (๐โ ) is increasing and ๐0 < ๐1 < โฏ < 0 ≤ ๐๐+1 < โฏ < ๐๐ . (7) It fallows from (7) that ๐น is decreasing on (−∞, ๐ฆ(๐+1) ) and increasing on (๐ฆ(๐+1) , ∞), therefore the minimum of ๐น is attained for med(๐, ๐) = ๐ฆ(๐+1) .โ Vedran Novoselac and Zlatko Paviฤ 1 7th International Scientific and Expert Conference TEAM 2015 Technique, Education, Agriculture & Management Belgrade, October 15-16, 2015 The next theorem shows relation between the center weight vector ๐(๐ฅ) = [1, … , ๐ฅ, … ,1]๐ and |๐ฆ๐ − med(๐(๐ฅ), ๐)|. Theorem 2 Let ๐ = [๐ฆ1 , … , ๐ฆ๐ ]๐ , ๐ = 2๐ − 1, be data vector with weight vector ๐(๐ฅ) = [1, … , ๐ฅ, … ,1]๐ . Then ๐น(๐ฅ) = |๐ฆ๐ − med(๐(๐ฅ), ๐)|, (8) is monotonically decreasing on ๐ท๐น = [1, ∞). Proof. Let 1 ≤ ๐ฅ1 ≤ ๐ฅ2 . If ๐ฆ๐ = ๐ฆ(๐) the function ๐น is constant an the proof in this case is trivial. Suppose that ๐′ correspond to position of ๐ฆ๐ in ordered observation and ๐ ′ ≤ ๐. In this case from Theorem 1 it is easy to see that med(๐, ๐) = ๐ฆ(๐+1) where ๐ + 1 ∈ [๐ ′ , ๐]. So it is sufficiently to indicate that ๐1 ≥ ๐2 where med(๐(๐ฅ1 ), ๐) = ๐ฆ(๐1 +1) , (9) and med(๐(๐ฅ2 ), ๐) = ๐ฆ(๐2 +1) . Situation when center weights are ๐ฅ1 , ๐ฅ2 ≥ 1 (10) 1 2 (or ๐ฅ2 ≥ ) indicate that ๐1 + 1 = ๐2 + 1 = ๐′, and 2 statement of theorem is proven. neighborhoods ๐ Observed nontrivial case is when ๐ฅ1 , ๐ฅ2 ≤ , i.e. 2 ๐1 + 1, ๐2 + 1 > ๐′. In that situation, according to Theorem 1, it can be conclude that ๐ − ๐ฅ1 − 1 (11) ๐1 = max {โ: โ < }, 2 and ๐ − ๐ฅ2 − 1 (12) ๐2 = max {โ: โ < }, 2 what directly implies statement of theorem because ๐ฅ1 ≤ ๐ฅ2 . Situation when ๐ ′ > ๐ is considered also when ๐ ๐ฅ1 , ๐ฅ2 ≤ , otherwise the proof is trivial. Considered 2 that case, it can be conclude that ๐1 + 1, ๐2 + 1 < ๐′, and it is sufficiently to indicate that ๐1 ≤ ๐2 . According to Theorem 1 it can be conclude that ๐1 = max {โ: โ < ๐ + ๐ฅ1 − 1 }, 2 (13) ๐2 = max {โ: โ < ๐ + ๐ฅ2 − 1 }. 2 (14) and From (13) and (14) the theorem is proven, i.e. ๐1 ≤ ๐2 , because ๐ฅ1 ≤ ๐ฅ2 .โ 3. Adaptive center weighted median filter X = [๐ฅ๐๐ ](๐,๐)∈๐บ โฏ โฑ โฏ ๐ฅ0,๐−1 โฎ ]. ๐ฅ๐−1,๐−1 (15) In such a system, the indices (๐, ๐) ∈ ๐บ of the image matrix correspond to (๐, ๐)-th image intensity ๐ฅ๐๐ . The term grey level is often to refer to the intensity of monochrome image. In that situation the matrix elements (image pixels) are integers in the range [0, … ,255]. Digital images are often corrupted by impulse noise during the transmission through communication channels. It appears as black or withe impulses on the image. It can be modeled as follows: ๐๐๐ , ๐ฆ๐๐= { ๐ฅ๐๐ , with probability ๐, with probability 1 − ๐, (16) where X = [๐ฅ๐๐ ](๐,๐)∈๐บ denote original image, Y = [๐ฆ๐๐ ](๐,๐)∈๐บ noisy image, and ๐ noise ratio. For a impulse noise ๐๐๐ is probability distribution with corresponding probability density function ๐๐ , ๐(๐๐๐ ) = { ๐๐ , 0, for ๐๐๐ = ๐, for ๐๐๐ = ๐ , otherwise, (17) where ๐๐ + ๐๐ = 1 and ๐๐ , ๐๐ ≥ 0. Most usually observed situation is when ๐ = 0, ๐ = 255, and ๐๐ = ๐๐ = 0,5. In image processing, filters are constructed to process every image element ๐ฆ๐๐ , (๐, ๐) ∈ ๐บ. In that way reconstructed image ๐ ∗ = [๐ฅ ∗ ๐๐ ](๐,๐)∈๐บ is constructed. Filter process different neighborhoods of ๐ฆ๐๐ where 3 × 3 filtering window are most commonly used for impulse noise. In that case filtering window is defined as ๐๐๐ = {๐ฆโ๐ : |โ − ๐| ≤ ๐ & |๐ − ๐| ≤ ๐}, (18) where ๐ denotes size window which is in our case is ๐ = 3. Filtering window ๐๐๐ can be presented as a vector ๐๐ = [๐ฆ1 , … , ๐ฆ๐ ]๐ , ๐ = ๐2 , where ๐ = ๐ โ ๐ + ๐ correspond to (๐, ๐)-th position. In this paper adaptive center weighted median filter (ACWMF) is proposed. Ideally the filtering should be applied only to the noisy pixels, and noise-free pixels should be kept unchanged. So, in filter is implemented a noise detector. We proposed an scheme by successfully combining center weighted median filters [3] and compares them with the observed pixel value. The output of proposed ACWMF is obtained by The digital image can be represented ๐ × ๐ matrix of the form: 2 ๐ฅ0,0 =[ โฎ ๐ฅ๐−1,0 Adaptive Center Weighted Median Filter 7th International Scientific and Expert Conference TEAM 2015 Technique, Education, Agriculture & Management Belgrade, October 15-16, 2015 ๐ฆ๐๐ , ๐ฅ ∗ ๐๐ = {med(๐(๐ฅ1 ), ๐๐ ), med(๐(๐ฅ2 ), ๐๐ ), ๐ ≥ ๐1 , ๐ < ๐2 , ๐2 ≤ ๐ < ๐1 . (19) In proposed method ๐1 and ๐2 are defined as ๐1 = |๐ฆ๐๐ − med(๐(๐ฅ1 ), ๐๐ )|, and ๐2 = |๐ฆ๐๐ − med(๐(๐ฅ2 ), ๐๐ )| respectively, where 1 ≤ ๐ฅ1 ≤ ๐ฅ2 . The definition of (19) follows the property that ๐1 ≥ ๐2 , showed by Theorem 2. Note that threshold ๐ affects the performance of impulse detection. As a results, impulse noise can be removed while uncorrupted pixels remain unchanged in order to preserve the image details. Consequently, the trade off between suppressing noise and preserving detail is well balanced. In the following section we study influence of center weights ๐ฅ1 , ๐ฅ2 , and threshold ๐ on the filtering performance using variety of the test images. 4. Experimental results range [30,60]. In figure 2 are presented PSNR filtering results of median filter (MF) [2], and ACWMF (๐ = 40) for Lena and Mandrill test images for a noise ratio ๐ ∈ [0,05, 0,5]. (a) (b) Figure 2. Filtering results: (a) Lena, (b) Mandrill Referring to figure 3, one can see that the noise suppression and detail preservation are satisfactorily compromised by using our proposed method (d) as compared to using standard median filter (c). The quality measure of proposed ACWMF has been experimented with center weights ๐ฅ1 = 1 and ๐ฅ2 = 2. The measures are provided via PSNR (Peak Signal-to-Noise Ratio) metric defined as PSNR = 20 log10 255 √MSE , (20) where MSE (Mean Squared Error) is defined as MSE = ๐−1 ∗ 2 ∑๐−1 ๐=0 ∑๐=0 (๐ฅ๐๐ − ๐ฅ ๐๐ ) ๐×๐ (a) (b) (c) (d) (21) . PSNR quality measure is appropriate for insight of impulse reducing because of its robust properties to outliers [4]. The PSNR via threshold ๐ is graphically presented in figure 1 for 256 × 256 test images Lena and Mandrill corrupted with impulse noise ๐ = 0,2. . Figure 3. Results of filtering for: (a) original, (b) noisy image (๐ = 0,2), (c) MF, (d) ACWMF 5. Conclusion Figure 1. Results of PSNR via threshold ๐ It can be seen that the PSNR performance is significantly improved by using threshold ๐ in the In this work a new median filter ACWMF is introduced. By incorporating the weighted median into an impulse noise detection framework is formed for effectively reducing impulse noise while preserving image details. Given a specified threshold ๐ the output of our proposed ACWMF may correspond to one of three possible states, namely the origin pixel value (i.e., the pixel is noise-free), or one of the center weighted output. Vedran Novoselac and Zlatko Paviฤ 3 7th International Scientific and Expert Conference TEAM 2015 Technique, Education, Agriculture & Management Belgrade, October 15-16, 2015 The proposed methodology remains applicable to adjust center weights ๐ฅ1 , ๐ฅ2 , and threshold ๐ according to different noise ratio ๐. In addition, the proposed filter present a quite stable performance over a wide variety of image. [2] [3] 6. References [1] 4 C. Gurwitz, “Weighted median algorithms for ๐ฟ1 approximation”, BIT Numerical Mathematics 12(2002), 524-560. [4] V. Novoselac, S. Rimac-Drlje, “Svojstva I primjena aritmetiฤke sredine”, Osjeฤki matematiฤki list, 14(2014), 1;51-67. V. Novoselac, B. Zovko-Cihlar, “Image Noise Removal by Vector Median Filter”, Proceedings ELMAR-2012, Croatian Society Electronics in Marine-ELMAR, 2012, 57-62. P. J. Rousseeuw, A. M. Leroy, “Robust Regression and Outlier Detection”, Whiley, New York, 2003. Adaptive Center Weighted Median Filter