MEDIAN FILTERING TECHNIQUES FOR VECTOR VALUED SIGNALS Constantin Vertan, Marius Malciu, Vasile Buzuloiu, Viorel Popescu Applied Electronics Department, Faculty of Electronics and Telecommunications, Bucuresti “Politehnica” University PO Box 35-57, Bucuresti 35, Romania E-mail: vertan@alpha.imag.pub.ro ABSTRACT Vector (multichannel) signal processing has arisen in the last decades as a major area of interest. The classical single-channel (scalar valued) processing techniques cannot deal with the interchannel dependencies, yielding to the development of new methods. This paper investigates two approaches to the median filtering of vector valued signals: filtering based on the reduced ordering in the space of signal samples and filtering with channel decorrelation. A novel filter is proposed that uses channel independentization by the central limit theorem. This filter performs better than classical filters in terms of mean square error measures and visual perception. 1. INTRODUCTION In the last years, the area of vector valued (multichannel, multispectral, multicomponent) signal (and particularly: image) processing has dramatically increased. The leading edge of development and interest is in the domain of remote sensing (dealing with 3 to 18 channels images), but the classical color images still remain the preferred research domain. Typically, a color image is represented in each pixel by a three component vector. The vector components are in general (or at least in the acquisition stage) the amounts of pure red, green and blue that compose the local color. These vector valued signals cannot be reduced to a stack of separately processed scalar components, due to the inherent (although not obvious) correlation between these components. This is why standard (scalar) processing techniques are not very appropriate. In the following of the paper, Section 2 discusses some median filters based on a reduced ordering in the signal samples space and Section 3 presents some possible ways of using the channel decorrelation in median filtering (including the novel proposed NDMMF - Normalized Decorrelated Marginal Median Filter). Finally, Section 4 summarizes some experimental results and Section 5 presents conclusions and further developments. 2. REDUCED ORDERING BASED MEDIAN FILTERING The median filtering [1] is a nonlinear, local order based processing technique, providing good results in image enhancement (impulsive noise removal with detail preservation). The main processing task behind this filter is the ascending ordering of the selected pixel values. Although natural and immediate in the scalar case (gray scale images), the ordering becomes a problem in a vector space. A classical paper in the domain of vector valued data ordering is [2]. In [2] several ordering criteria are discussed and classified as marginal, conditional, partial and reduced (or aggregate) ordering. The marginal ordering means the scalar ordering of each component (recalling the model of stack of separately processed components) and we shall refer to the MMF (Marginal Median Filter) as the basic vector filter. The main disadvantage of this ordering is the fact that the ordered values can be significantly different from the original data to be ordered. The conditional ordering implies to order the data according to a hierarchical order of the components; the lexicographic ordering is the best example of such a criterion; this ordering is a complete one (from the mathematical point of view) but the space topology is not preserved. The results of applying this ordering for the RGB space are not good. The partial ordering is based on the convex hull like set decomposition of the data samples; the inner-most set will be a kind of median. The implementation problems are related to the time consuming convex hull determination and to the inaccuracy due to the very sparse nature of the cluster of data points. (We recall that the data points are issued from an analysis window, typically a 3 by 3 window containing only 9 samples). Finally, the reduced ordering seems more appropriate. The idea is to order some scalars, each scalar being computed from the components of the vector from the data set; the same ordering is assumed for the vectors. A general case is the ordering with respect to a fixed reference vector α and a generalized distance determined by a positive semidefined matrix A. In this case, the scalars are the distances from each data vector to the T reference vector d ( x i , α ) = ( xi − α ) ⋅ A ⋅ ( x i − α ) . The classical VMF (Vector Median Filter) [3] is introduced by means of the aggregate (cumulative) distance (the sum of distances from a vector to all other vectors): N D( xi ) = λ ( xi − x j ) T ⋅ A ⋅ ( xi − x j ) . j =1 The median is the vector with the minimum aggregate distance. In [4] the same principle was used for angular distances. In [5] a linear combination of distances to the (local) marginal median, marginal mean and center located vector for noise removal from color images was successfully used (the AHM - Adaptive Hybrid Median). If A=I (identity matrix), the Euclidean distance is obtained. The idea of the MVMF (Modified Vector Median Filter) approach is to use a weighted Euclidean distance. The weighted Euclidean distance is defined by a nonidentity matrix A, in the simpler case a diagonal matrix. Since in the analysis window the components (red, green, blue for the particular case of color images) are not equally distributed or proportional, it seems natural to use some weights that emphasize the local relative importance of each component. There are several choices for the assignment of weights: a component is considered important according to its range (the difference between the local maximum and minimum), its normalized range (range to average ratio), its variance or contrast scaling (variance to average ratio, as defined in [6]). Another possible choice is the use of the covariance matrix for A, yielding to the Mahalanobis distance, wellknown in the pattern recognition problems. As the experimental results have shown (see Section 4), the best results were obtained using the weighting by the range. This seems normally, since some difference between the components of two vectors could be important if the range is small or relatively neglectable if the range is big. 3. DECORRELATION BASED MEDIAN FILTERING As suggested by [7], a component wise processing of multichannel images could be successfully performed if the components are decorrelated first. The best choice for this decorrelation procedure is the Karhunen Loeve transform, but other (faster) transforms could be applied, if allowed by the statistical model of the image. The DMMF (Decorrelated Marginal Median Filter) is composed from a decorrelation block (direct integral transform), a scalar median filter for each decorrelated component and a recorrelation block (inverse integral transform). Since the Karhunen Loeve transform (and the Discrete Cosine transform DCT) concentrates most of the spectrum energy (relevant feature) in a few coefficients, it is natural to assume that the median filtering is not necessary to be performed on every component, but just on the most significant one. This approach is used in the TDMMF (Truncated Decorrelated Marginal Median Filter), reducing the computational amount. The component to be processed is chosen as the most important decorrelated component with respect to a given importance criterion (range, normalized range, variance, contrast scaling, “dccoefficient”). We may notice that the truncated filtering approach can also be applied in the initial RGB space (with the same local importance criteria). As the results show, a significant improvement can be obtained only for small noise amounts. Further improvements are obtained by the better NDMMF (Normalized Decorrelated Marginal Median Filter). The idea of this filter is that normal random variables, KL transformed, become independent (so decorrelated). The point is to obtain from the original RGB components distribution a normal distribution. It seems two methods are at hand for this task: the histogram specification technique [6] and the central limit theorem. The histogram specification leads to a Gaussian shaped distribution that includes “gaps” and is not invertible (after the median filtering, another histogram specification must be performed to obtain an approximation of the original distribution). According to the central limit theorem, the sum of a few independent, identical distributed random variables has approximately a normal distribution. The sum can be introduced by a local linear invertible transform. This transform replaces each pixel value within the analysis window by the sum of the other 8 values in the window; the result of a median filtering of these sums is assigned to the central value and the inverse transform is performed. Using this technique, some averaging properties are introduced in the median filter; not surprisingly, this filter performs very good in the RGB space too (without any decorrelation). H W λλ MCRE = 4. EXPERIMENTAL RESULTS The tests were conducted on two classical 256 by 256 true color (24 bits per pixel) images, “fish” and “waterfall”. Both were corrupted with impulsive (salt & pepper) noise with 1% to 4% noise affected pixels in each component, with no correlation in between, Gaussian (normal) zero-mean, variance 25 additive noise and a mixture of Gaussian and 2% impulsive noise. The filtering efficiency for noise removal was measured by the standard [4] NMSE (Normalized Mean Square Error) and MCRE (Mean Chromaticity Error), defined by: H W λλ NMSE = f (i , j ) − f ( i , j ) i =1 j =1 ⋅100 and H W λλ 2 f (i , j ) 2 i =1 j =1 i = 1 j =1 2 f ( i, j ) f ( i, j ) − f (i , j ) f (i , j ) ⋅ 100 . HW In the definitions above, H and W are the image dimensions 2 1 in pixels, is the L norm and is the L norm, f(i,j) and f (i , j ) are the original and filtered values of the pixel at location (i,j). A subjective quality mark was accorded by visual inspection of the results. Tables 1 and 2 and Figures 1 to 4 summarize the results of some typical filters for the two test images: original noisy images (1), Marginal Median Filter (2), Vector Median Filter (3), Modified Vector Median Filter with weighted Euclidean distance by local range (4), Decorrelated Marginal Median Filter by KL transform (5) and DCT (6), Truncated Decorrelated Marginal Median Filter by KL transform and median filtering of the component with biggest local range (7), Normalized Decorrelated Marginal Median Filter by KL transform (8), DCT (9) and Normalized Marginal Median Filter in the RGB space (10). Table 1: Median filtering of “waterfall” image Noise 1% 4% Mixt. Gauss Noise 1% 4% Mixt. Gauss Filter type 1 MCRE NMSE 1.049 15.365 4.072 31.049 19.330 29.495 18.604 21.396 Filter type 2 MCRE NMSE 1.880 9.316 2.146 9.697 8.013 13.476 8.048 13.422 Filter type 3 MCRE NMSE 2.136 10.941 2.476 12.401 10.627 17.049 10.311 16.036 Filter type 4 MCRE NMSE 2.056 10.636 2.264 11.272 10.352 16.791 10.307 16.102 Filter type 5 MCRE NMSE 1.750 9.429 1.887 10.447 6.762 13.780 7.758 13.639 Filter type 6 MCRE NMSE 1.758 9.646 1.859 10.556 7.647 13.666 7.697 13.302 Filter type 7 MCRE NMSE 1.302 15.234 2.908 27.311 15.582 31.785 15.559 27.902 Filter type 8 MCRE NMSE 1.754 8.521 2.541 10.180 7.802 13.276 7.700 12.703 Filter type 9 MCRE NMSE 1.781 8.773 2.600 10.327 7.800 13.133 7.657 12.472 Filter type 10 MCRE NMSE 1.831 8.434 2.790 9.503 8.168 12.976 8.021 12.606 Filter type 3 MCRE NMSE 6.734 21.003 7.395 22.898 19.341 25.908 18.568 24.543 Filter type 4 MCRE NMSE 6.567 20.678 7.007 21.969 19.060 25.593 18.527 24.495 Filter type 5 MCRE NMSE 5.621 19.861 5.955 20.963 11.932 24.162 12.015 23.793 Table 2: Median filtering of “fish” image Noise 1% 4% Mixt. Gauss Filter type 1 MCRE NMSE 1.409 18.975 4.996 36.112 23.462 32.143 22.556 21.833 Filter type 2 MCRE NMSE 5.991 19.184 6.749 19.563 16.412 21.922 16.561 21.717 Noise 1% 4% Mixt. Gauss Filter type 6 MCRE NMSE 5.635 20.288 6.022 21.481 11.270 22.891 11.188 22.618 Filter type 7 MCRE NMSE 5.749 33.682 7.381 42.107 21.182 52.292 21.330 48.563 Filter type 8 MCRE NMSE 5.533 17.602 7.182 19.317 12.178 22.178 11.819 21.382 By analyzing the results, we can notice the improvement obtained by the weighted distance technique (MVMF) and the decorrelation based techniques (DMMF). The NDMMF shows the best results in terms of NMSE and visual accuracy of the image. The computational amount is almost the same for all the filters; just the implementation of the KL transform is more annoying, but the DCT can be successfully used instead. Filter type 9 MCRE NMSE 5.541 18.034 7.246 19.849 11.773 21.152 11.264 20.614 Filter type 10 MCRE NMSE 5.856 17.036 7.935 17.966 16.261 20.188 15.856 19.699 The presented color median filtering techniques give encouraging results. With the same computational amount as the VMF, or even less, the DMMF and the NDMMF obtain significant better enhancement of the images, both visual and measured. Fig. 3: MVMF filtering of Fig. 1 with range weighting, MCRE=2.264%, NMSE=11.272% Fig. 1: 4% impulsive noise degraded “waterfall”, MCRE=4.072%, NMSE=31.079% Further developments may consider the processing in some other color representation space that either emphasize the color characteristics (like HSI - Hue Saturation Intensity) or maximizes the relative intercolor distances, enabling better color discrimination (the CIE 1964 U*V*W * representation [6]). 6. REFERENCES Fig. 2: NDMMF filtering of Fig. 1 using DCT, MCRE=2.600%, NMSE=10.327% 5. CONCLUSIONS [1] I. Pitas, A.N. Venetsanopoulos: Nonlinear Digital Filters - Principles and Applications, Norwell MA, Kluwer Academic Publ. 1990 [2] V. Barnett: “The ordering of multivariate data”, J. R. Stat. Soc. A, vol. 139, part 3, pp. 318-343, 1976 [3] J. Astola, P. Haavisto, Y. Neuvo: “Vector median filter”, Proc. IEEE, vol. 78, pp. 678-689, Apr. 1990 [4] P.E. Trahanias, A.N. Venetsanopoulos: “Vector Directional Filters - a new class of multichannel image processing filters”, IEEE Trans. on Image Processing, vol. 2, no. 4, pp. 528-534, Oct. 1993 [5] K. Tang, J. Astola, Y. Neuvo: “Nonlinear Multivariate Image Filtering Techniques”, IEEE Trans. on Image Processing, vol. 4, no. 6, pp. 788-798, June 1995 [6] A.K. Jain: Fundamentals of Digital Image Processing, Englewood Cliffs NJ, Prentice Hall Intl., 1989 [7] B.R. Hunt: “Karhunen-Loeve multispectral image restauration, part I: theory”, IEEE Trans. on Accoust., Speech and Signal Processing, vol. ASSP-32, no. 3, pp. 592599, June 1984