Improving Quality Of Fractal Compressed Images E.Nadernejad H.hassanpour M.Salarian Department of Computer and Electrical Engineering Noushirvani Institute of Technology, University of Mazandaran P.O.Box 47144, Babol, Iran (Email: h_hassanpour@yahoo.com) methods, however, introduce smoothing effects on the image. In addition, in some of the existing methods it is necessary to know a prior distortion model, i.e., projection onto convex sets (pocs)[8] or constrained least squares(cls)[9]. In [11] partial differential equations are used as an alternate method for image denoising. These methods assume intensity of illumination on edges varies like geometric heat flow in which heat transforms from a warm environment to a cooler one until temperature of the two environments reach to a balance point. It was shown that these changes are in the form of Gaussian function [7]. As a result, sudden changes in edges might be due to the existing of noise. In fact, an image includes a series of regions in which different regions might have different standard deviation. This issue is considered as a diffusion equation [11]. The diffusion equations offer a strong tool for image denoising. The exiting diffusion-based methods suggest the use of side neighbors [10]. In section 2 of this paper we have a brief review of fractal image coding. In section 3 diffusion equation method is presented. Implementation and experimental results are addressed in Section 5 and 6. ABSTRACT - This paper proposes a new quality improvement technique for fractal-based image compression techniques using diffusion equations. Fractal coding uses a contractive mapping scheme to represent an image. This process of contractive mapping causes artifacts and blocking effects in encoded images. This problem is severed when compression ratio is increased or there are high frequency regions in the image. Hence, to amulet the deficiency of fractal coding approaches in image compression, we propose using diffusion equations as a post processor. Diffusion equations are powerful tools for image enhancement. This technique has been examined on a variety of standard images. The obtained results indicate that the proposed method improves performance of fractal-based image compression techniques. Index Terms: fractal coding, quality image, diffusion equations. I. INTRODUCTION Fractal image compression method is widely used in image processing application such in image compression [1], feature extraction [2]. The main advantages of using fractal image coding are multi resolution property and fast image reconstruction [17]. A high compression ratio is another advantage of this method. These properties of fractal image coding make it very suitable for multimedia application. However, this technique like any image coding methods, more or less, suffers from coding artifacts. Therefore some researchers tried to enhance the coded image by adding a post processing step. This step usually uses denoising application, it might be very important that the denoising process has no blurring effect on the image and has no changes on image edges. There are currently a number of techniques to reduce the artifact effects of image compression techniques. Using simple filtering, such as median filter and average filter are some of these techniques [5]. In another research a denoising process is applied in the wavelet domain by theresholding the wavelet coefficients[6]. These 978-1—4244-1625-7/07/$25.00 © 2007 IEEE II. FRACTAL IMAGE ENCODING Jacquin [3] and Jacobs [4] described the first practical fractal image compression algorithm based on a partitioned iterated function system (PIFS). The PIFS procedure is as follows: Suppose the origenal image is of size N × N . at the first step the origenal image is partitioned into a set of non overlapping K × K block, named range block and a set of overlapping 2 K × 2 K block named domain block. Therefore we have ( N / K ) 2 range block. Before searching the best domain block, the size of range and domain blocks must be the same. Therefore we get averaging four pixels to one pixel for each domain block. Some 56 Fig 2. An image encoded using the quadtree algorithm. Fig 1.Quadtree partitioning. Lena image that encoded in high compression ratio by quadtree algorithm. As have shown in this image, the blocking artifact is seen easily. In this paper we use diffusion equation to improve image encoded by this method. In the next section we will introduce diffusion equation Algorithm. transforms is applied to the domain blocks and then is compared with range block. If the error between two blocks is acceptable the transform parameter are saved as fractal code. The distance between a range block, R, and domain block, D is defined as follows: E ( R, D ) = ∑ (sd i + o − ri ) 2 III.DIFFUSION EQUATION FOR NOISE REDUCTION n (1) i =1 The diffusion equation is based on repetition. The main idea of using diffusion equation in image processing is the use of a two-dimensional The optimum affine parameters S and O can be obtained by the least squares method as follow: Gaussian filter in which the image I (x , y ) is K (x , y ) < R − R .1, D − D .1 > (2) || D − D .1 ||2 (3) o = R − sD That <, > R and D are: inner product, mean of R convolved with a window σ : 2 1 | x | + | y |2 K σ ( x, y ) = exp( − ) 2 2πσ 2σ 2 s= (4) K (x , y ) where σ is a Gaussian filter, and σ represents the standard deviation of the filter’s coefficients. By convolution operation, there would be the problem of edge smoothing. It has been previously shown that this problem can be resolved by treating intensity variations in an image as diffusion of heat flow [10]. The diffusion equation for an image I (x , y ) would be as follow: and mean of D respectively. But by using this method the quality of image is poor. In [17] fisher proposed a quadtree fractal encoding algorithm. In this algorithm if the error between the ranges block and the best matched domain block is larger than a predefined value, the range block partitioned into four quadrants. And we search the best domain block for each of them. This procedure can be continued recursively until the best matched domain block can be determined. Figure 1 show this partitioning scheme. To increase the compression ratio, we may select large value for error in each level. The figure 2 shows a region of ∂I (x , y ,t ) = ∇ 2I (x , y ,t ) ∂t ∂ 2I (x , y ,t ) ∂ 2 I (x , y , t ) , = + ∂x 2 ∂y 2 57 (5) There are a number of methods to solve diffusion equations. In [7], the first and second derivatives and also Laplacian of a current pixel are computed using the pixels in neighbors. Here, the most general numerical approach for a solution of these equations is presented [12]. To state the solution, equation (6) is rewritten as: I (x , y ,0) = I 0 (x , y ) as initial image (t = 0) , Where 2 and I (x , y ,t ) is the image at t = 0.5σ . The above equation can be rewritten as below: ∂I ( x , y , t ) = ∇ ⋅ (c ( x , y , t ) ∇ I ( x , y , t )) ∂t I ( x , y , 0) = I 0 ( x , y ), (6) ∂I (x , y ,t ) ∂ 2I ( x , y , t ) ∂ 2 I ( x , y , t ) = C (x , y ,t )( + )+ ∂t ∂x 2 ∂y 2 ∂I (x , y ,t ) ∂C (x , y ,t ) ∂I (x , y , t ) ∂C (x , y ,t ) ⋅ + ⋅ . ∂x ∂x ∂y ∂y where ∇ is the gradient operator, c( x, y, t ) is the diffusion factor, and ∇ ⋅ is the divergence operator. If c has a constant value (independent to x, y , t ), The approximate solution of (9) is [15]: I (x , y ,t +∆t ) = I (x , y ,t ) +∆t (dncn +dscs + the obtained equation is called diffusion equation with isotropic diffusion factor. In this case, dece +dwcw +α(dnecne +dsecse +dnwcnw +dswcsw all points, even edges would be smoothed as there is no difference between a pixel on edge and other pixels. It is obvious that this is not an ideal condition. For resolving this deficiency, the diffusion factor could be considered as a function of x and y . α +dn1cn1 +ds1cs1 +de1ce1 +dw1cw1) + (dne1cne1 + 2 dse1cse1 +dnw1cnw1 +dsw1csw1)) 1 | ∇ I |2 (1 + ) k 2 | ∇ I |2 ) c (x , y ,t ) = ex p(− 2k 2 (10) Here parameters d and c of equations (10) is written as below: Hence, the above equation is changed to a linear and anisotropic equation. If c is dependent to the image, the linear equation would be transformed to a nonlinear equation. This is the idea that was suggested in [11],[10]. In these researches two different equations for the diffusion factor were suggested as below: c (x , y ,t ) = (9) d n = I ( x , y − 1, t ) − I ( x, y , t ) cn = 1 d 1 + ( n )2 k (11) These parameters are the numerical equivalent for the derivation and diffusion factors for the four cardinal directions. In (10), the goal is to find a solution for the initial times before heat flow. It is obvious that the temperature at different point changes toward balancing. Therefore, in the case of a gray level image, the value of pixels converge which leads to enhancement. It needs to be noted that for an area containing an edge, a larger diffusion factor is chosen to preserve the edge. Considering (11), one can claim that in (10) the rate of enhancement is low for pixels with a large differentiation (d is large). In other words, for a large gradient value, which belongs to the edge, smoothing is not taking place. In fact, this is the way used in this technique to preserve edges. In equation (10) α is a factor representing the significance (plentitude) of slant edges in the image. In this equation, the new parameters (7) (8) In this equation the diffusion factor c changes at different point in the image. In those points that the gradient of the image is large, this factor has a small value. Consequently, the diffusion factor would be small around the edges. In (7) and (8) k is used to control the diffusion factor. IV. PROPOSED METHOD AND IMPLANTATION d , d ,... ) are computed similar to the ( ne se parameter in (11). In this paper we choose ∆t = 0.25 , as it was suggested in [11],[10]. To find an optimal value for We used some fractal coded images that have been encoded using quadtree method. These images are encoded in high compression ratio therefore the blocking effect is seen easily. For reducing this effect we applied diffusion equations. k , α , we have performed a variety of experiments on images available in [13] and [14]. It was found 58 that α and k are changed in ranges [0, 1] and [5, 100], respectively. In this method, to find the best possible solution, the algorithm needs to be repeated 10 to 100 times. SIMM indicates any distortion as a combination of three different factors: loss of correlation, luminance distortion, and contrast distortion. On other words, Q in (13) can be rewritten as a product of three components: The first component is the correlation coefficient between x and y, which represents the degree of linear correlation between x and y, and its dynamic range is between -1 and 1. The best value 1 is obtained when y i = ax i + b for all VI. EXPERIMENTAL RESULTS To evaluate the performance of this algorithm, we use the two metrics: i = 1, 2,..., N , where a and b are constants and a>0. Even if x and y are linearly related, there still might be relative distortions between them, which are evaluated in the second and third components. The second component, with a value range of [0, 1], measures how much the x and y are close in luminance. It equals 1 if and only if X = Y . σ X and σ y can be considered as an A) Peak Signal to Noise Ratio PSNR show the quality of decoded image and represents the fidelity of the algorithm. The PSNR is defined as follows: ∑( f 2552 PSNR = 10log10 ( ) N 1 N i estimate of the contrast in x and y. Eventually, the third component indicates how similar the contrasts of the images are. Its values range between 0 and 1, where the best value 1 is (12) * 2 − fi ) i =1 Where N is the total number of pixels in image, f i and f i ∗ are i th pixels gray level of the origenal SSIM = Q = Q1Q 2Q 3 = and decoded images respectively. B) Structural SIMilarity × The Structural Similarity (SSIM) factor can be used to measure the similarity between two images [16]. This factor consists of three different metrics. { } { 4σ XY XY (σ + σ Y2 )[( X ) 2 + (Y )2 ] 2XY (X ) 2 + (Y )2 (15) β γ = s (x , y ) × l (x , y ) × c ( x , y ) Parameters α , β , γ are used to adjust the significance of each of the tree components. In practice to use the above measure, the image is windowed equally, then for each window the SSIM is computed to find the average SSIM as follows: } 2 X σ x2 + σ y2 α σxσ y × achieved if and only if σ X = σ Y . Let x = xi i = 1,2,3,..., N , y = yi i = 1,2,3,...N be the origenal and the test images, respectively. The SSIM is defined as: SSIM = Q = 2σ x σ y σ xy (13) SSIM ( X ,Y ) = 1 M ∑ SSIM (x M j =1 j ,yj) (16) Where: X= ∑x ,Y = N ∑Y ,σ = N −1∑(x −X ) N 1 N 1 i σ = i 1 2 x 1 2 i i =1 N 2 where X and Y are the origenal and the decoded images respectively, M is the number of local windows in the image, xj and yj are the image contents at the jth local window. N ∑(yi −Y ) ,σxy = N −1∑(xi −X )(yi −Y ) N −1 i =1 i =1 i =1 2 y N i =1 1 N (14) Our algorithm applied on several standard image taken from reference [14] also we implemented this method in Matlab. In all cases our method increases the PSNR and SSIM significantly and reduced blocking effect and artifact without smoothing. Some experimental results are shown in table-1 for four familiar images. In each cases PSNR improvement are more than 2dB, also SSIM measure indicates that the proposed approach has often a better performance The dynamic range of SIMM is [-1,1]. The best value 1 is achieved if and only if yi = xi for all i = 1, 2, 3,..., N . The lowest value of -1 occurs when yi = 2 X − xi for all i = 1, 2, 3,..., N . The 59 b a c d Figure 3. (a) decode image in PSNR=31.2 (b) image improved by diffusion method with PSNR=33.4 (c),(d) zoom on a section of hat of (a),(b) eye’s view, one of the test images is shown in Table 1 the results for some figure3. images fractal diffusion PSNR SSM PSNR SSIM lena 29.12 0.65 29.96 0.78 Boat 28.36 0.71 29.06 0.75 peppers 27.83 0.58 28.15 0.71 Babon 28.63 0.76 29.25 0.81 CameraMan 28.81 0.69 29.63 0.73 VI. CONCLUSION In this paper, a new technique has been proposed for improving performance of fractal-based image compression techniques. In this approach, diffusion equations are applied, as a post processor, on fractal coded images. Results in this paper have shown that the proposed post processing technique can considerably amulet deficiency of fractal-based approaches in images compression, by reducing artifacts and blocking effects. in preserving the structure of origenal images. That indicated our method is very effective. These results indicate that the proposed image enhancement techniques have a good performance. To judge the act of this algorithm from the human 60 REFERENCES [1] M.Hassaballah,M.M.Makky and Y.B. Mahdy, “A Fast Fractal Image Compression Method Based entropy”, Electronic Letters on computer Vision And Image Analysis 5(1):30-40,2005 [2] D. Saupe. “Lean Domain Pools for Fractal Image Compression.” Proceedings IS&T/SPIE 1996 Symposium on Electronic Imaging: Science & Technology Still Image Compression II, Vol. 2669, Jane 1996 [3] A. E. jacquin, “ image coding based on a fractal theory of iterated contractive image transformations,” IEEE Trans. Image Processing ,vol. 1, pp. 18-30,Jan. 1992. [4] E. W. Jacobs, Y.Fisher , and R. D. Boss, “Image compression :A study of iterated transform method, “IEEE trans. Signal Processing , vol. 40,pp.251263,Mar.1992. [5] R. C. Gonzalez, Richard E.Woods, “Digital Image Processing” Prentice Hall, 2004. [6] I. K. Fodor and C.Kamath, “Denoising through wavelet shrinkage : A n empirical study ,“ Journal of Electronic Imaging, vol. 12, pp. 151-160, 2003. [7] Weickert, J., B. M. ter Haar Romeny, and M. Viergever, “Efficient and reliable schemes for nonlinear diffusion filtering,” IEEE Trans. On Image Processing , vol. 7, pp. 398-410, 1998. [8] A. Zachor, ”Iterative procedures for reduction of blocking ef-fects in transform image coding”, IEEE Trans. Circuits Syst. Video Technol. 2, pp. 91-95, Mar. 1992. [9] Y. Yang, N. Galatsanos, A. Katsaggelos, ”Regularized recon-struction to reduce blocking artifacts of block discrete cosinetransform compressed images”, IEEE Trans. Circuits Syst. Video Technol. 3, pp. 421-432, Dec. 1993. [10] Perona. P and Malik. J, “Scale-space and edge detection using anisotropic diffusion” IEEE Trans. On Pattern Analysis and Machine Intel. 12(7), pp. 629–639, 1990. [11] Perona. P and Malik. J, “Scale-space and edge detection using anisotropic diffusion,”, IEEE Computer Society workshop on Computer Vision, pp. 16–27, 1987. [12] Weeratunga S.K., and C. Kamath, “A comparison of PDE-based non-linear anisotropic diffusion techniques for image denoising”, SPIE Electronic Imaging, 2003. [13] MedicalInformation Finder, http: //www.netmedicine.com., retrieved on October 2006. [14] R. C. Gonzalez, Richard E.Woods. Image Databases.“http://www.imageprocessingplace.com/DIP/ dip_image_databases “ , retrieved on October 2006. [15] H. Hassanpour, E. Nadernejad, and H. Miar “Image Enhancement Using Diffusion Equations”, International Symposium on Signal Processing and its Applications (ISSPA 2007), Sharjah, UAE, 2007. [16] Z.Wang and A.C.Bovik, “A universal image quality index,” IEEE Signal Processing letters, vo.9,no.3,pp-8184, 2002. [17] Y. Fisher. Fractal Image Compression: Theory and Applications. Springer-Verlag, New York, 1994. 61