A Modified Image Coder using HVS Characteristics

Transcription

1 A Modified Image Coder using HVS Characteristics Mrs Shikha Tripathi, Prof R.C. Jain Birla Institute Of Technology & Science, Pilani, Rajasthan Abstract: Embedded Zero Tree (EZW) coding, introduced by J.M Shapiro, is a very effective and computationally simple technique for image compression. Here we offer a technique which uses EZW and some simple blocks thereby increasing the compression ratio four times with a little degradation in Peak Signal to Noise ratio (PSNR) The proposed technique can be implemented without any change in the existing image coders by adding a Down Sampling Filter (DSF) and a Contrast Sensitivity Function (CSF) mask at the transmitter and a Up Sampling Filter (USF) at the receiver end. Besides improving the compression ratio the new method is about three times faster than the existing coding techniques. Keywords: Wavelets, HVS, CSF mask, Image compression 1.INTRODUCTION Wavelets, which are known to the scientific community since 1910, have found applications in many diverse fields such as data compression, denoising, edge detection, image fusion, etc. The most widespread application of Wavelet transform has been for data compression [1]. Wavelet coding, on account of many merits of it, is being used for new JPEG 2000 coding standards. A breakthrough in lossy wavelet coding occurred in 1993 when Shapiro introduced Embedded Zero-tree Wavelet (EZW) coding [2] in which he exploited the multiresolution nature of the wavelets. EZW marked the beginning of modern wavelet coding, since improved wavelet coders proposed subsequent to EZW are based on fundamental concepts from EZW coding [3]. This algorithm provided improved performance at low bit rates compared to the DCT based compression scheme. In this paper, a modified wavelet based coding technique for image compression is proposed. In this scheme, EZW followed by Huffman encoder has been used to generate the bit stream to be transmitted. This technique can be applied to any of the existing encoders like EZW, Set Partitioning In Hierarchical Trees (SPIHT), Low complexity and Low memory Entropy Coder (LLEC), Space frequency quantizer etc [2, 4, 5, ]. Two aspects of images have been explored in this technique. The pixels in most images have very high degree of two dimensional (horizontal and vertical) correlation. Hence the image is down sampled by a factor of 2 both horizontally and vertically and sent to the encoder for further processing. This technique improves compression ratio four times than that obtained by using standard technique [2, 4, 5, ] with a little degradation (2-4dB) in the image quality. The other aspect is the Human Visual System (HVS) characteristics that exploit eye s inability to discriminate between different brightness levels. A Contrast Sensitivity Function (CSF) mask, further explained in section 2.1, is used in addition to the Down Sampling Filter (DSF) to give further compression. 2. MODIFIED IMAGE CODER The block diagram of the proposed system is shown in Fig.1. Down Sampling Filter (DSF) is a block where the alternate pixels of the image data are dropped both horizontally and vertically. Up Sampling Filter (USF) is a block where the skipped rows and columns are replaced using interpolation techniques. The CSF mask is developed using HVS characteristics as explained later in 2.1. The other blocks in the diagram are standard blocks.

2 In the first part of this technique the image is down sampled by a factor of two, each horizontally and vertically. This is done by skipping alternate rows and columns. Using a DSF a N x N image results in a N\2 x N\2 image. The distortion in the reproduced image is inversely proportional to row-to-row and column-to-column correlation. The Up Sampling Filter used in the receiver chain is a median filter because a median filter generally gives better PSNR (Peak Signal to Noise Ratio) compared to other filters []. The performance of USF is image dependent. Median filter is more suitable for high frequency images. For image with both alternate rows and columns skipped the median filter used is as shown in Fig.2. A1. A2. A3 X1 X2 X3.. B1. B2. B3 X1 = (A1 + B1) / 2; X2 = median (A1, A2, B1, B2, X1, X3); X3 = median (A1, A2, A3, B1, B2, B3); Fig.2. Median filter. Another up sampler where the skipped rows and columns are replaced by duplicating the previous rows and columns was also tried out. The PSNR in this case slightly improved by a factor of 0.1 db when the correlation between pixels was high. In the second part of this technique the HVS criteria is exploited to improve compression. HVS research aims at developing mathematical models regarding human vision. Models have already been developed to characterize human eye sensitivity to brightness and color [8]. The Contrast Sensitivity Function (CSF) describes our eye s sensitivity to spatial frequencies. A CSF model for luminance (grayscale) images, originally proposed by Mannos et al [8], is given as: [ (0.114 f ) ] H ( f ) = 2.( f ) e...(1) where spatial frequency is f = ( f x + f y ) f has units of cycles / degree of visual angle subtended on the eye. f x and f y are the spatial frequencies in the horizontal and vertical directions respectively. The spatial frequency is normalized with the following relation: f(cycles/degree) = f n (cycles/pixel).f. s (pixels/degree). (2) where f n is the normalized spatial frequency with units of cycles/pixels. f s is set to 4 pixels/degree as the observers have been supposed to be sitting at an approximate viewing distance of 2 feet. This value for f s is set for a viewing distance of four times the image height (the image height is approximately inches). In Equation (2), since f has a range between 0 and f s /2, f n has a range between 0 and 0.5. Fig.3. depicts the CSF curve. This characterizes luminance sensitivity of the HVS as a function of normalized spatial frequency. The HVS is most sensitive to frequencies between 0.03 and 0.23 and less sensitive to very low and very high frequencies CSF Masking CSF masking is used to refer to the method of weighting the wavelet coefficients relative to their perceptual importance. The design problem is how to transform the CSF curve in Fig.3. into a mask of weights that will multiply the wavelet coefficients. The first step in creating a DWT CSF mask is to perform a wavelet decomposition of the CSF curve. Next the mask weights from this decomposition are obtained as follows: 1) Take the 1-D DWT of the CSF curve in Fig.3. A 2-level Daubechies wavelet decomposition is carried out with the subspaces W 1 and W 2 as shown in Fig.4. W 1 is the subspace obtained after first level 1.1

3 decomposition of the CSF curve and W 2 is the subspace obtained when 1-D wavelet transform is applied to the approximation coefficients of the first level. 2) Label the peak of W 1 subspace as p 1 and the peak of W 2 subspace as p 2. 3) In case of images when a 2-D wavelet transform is applied the first level of decomposition gives LL, HL, LH and HH subbands. The weights for LH, HL, HH subbands are given by p 1, the peak of the first level of decomposition. 4) The LL subband is further decomposed to give LLLL, HLLL, LHLL and HHLL subbands. The weights for these subbands are given by p 2, the peak of the second level of decomposition of the CSF curve. The number of decompositions can be increased keeping the number same in both CSF curve and EZW quantizer. The number of weights obtained is equal to the number of decompositions. The wavelet transformed and CSF weighted image is then sent to EZW encoder and other blocks for further processing. 3. RESULTS AND DISCUSSION The following results were obtained with monochrome, 8 bpp and 512 x 512 images. Table.1 gives PSNR Vs bpp data of Lena and Barbara comparing the values EZW encoder and the one with additional DSF and median filter as USF. Fig 5. gives PSNR Vs bpp data of different images using DSF and USF. Very high compression ratios (about four times) were achieved at the compressed image quality with PSNR of about db, which is an acceptable range for many images. Hence for extremely low bit rates this method outperforms the conventional methods, which do not use these additional blocks. For baboon, goldhill, and cameraman though the PSNR is low, due to high frequency contents present in these images, perceptually the images look good. Table.2 gives PSNR and bpp data of CSF mask in addition to the existing blocks. In technique (1) a 5-level decomposed CSF mask is applied to the wavelet transformed (5-level) image and the image is reconstructed. The image is then down sampled by DSF and up sampled on the receiver side by median filter. Comparing the results obtained in table.1 (a), where the two proposed blocks are not present, with that in table.3 technique (1) it is observed that for 2.4 db PSNR the compression ratio is 3.8 times improved. In technique (2) the image is first down sampled then the CSF mask and EZW quantization is applied for a 2-level decomposed image keeping the same USF. It is observed that the PSNR now improves by 2 db, compared to previous technique for passes in the EZW algorithm, with slight reduction in the compression ratio. The PSNR is further improved by about 0.1 db using USF where the skipped rows and columns are obtained by duplicating the previous rows and columns. By suitably designing the USF depending on the image type the PSNR can be improved. In table.1 (e) where DSF, CSF mask and median filter as USF is used for Barbara it is observed that both PSNR and compression ratios are improved compared to (d) where only Sampling filters are used. 4. CONCLUSION This technique is considered to be very well suited for applications like Progressive Image Transmission (PIT) where, in the initial passes, more than the quality, bandwidth and storage are the main considerations. The compression ratio increases by a factor of 4 with 2-3 db decrease in PSNR. The down sampling technique gives good results if the image has low frequency components (Lena) and the HVS technique performs better for high frequency images (baboon, Barbara). Better results can be obtained by increasing the number of wavelet decomposition levels. REFERENCES [1] M.R. Raghuveer and S.B. Ajit, Wavelet transforms, introduction to theory and applications, Addison Wesley, 2000.

4 [2] J.Shapiro, Embedded image coding using zero-trees of wavelet coefficients, IEEE Trans. Signal Processing, vol. 41,pp , Dec [3] Bryan E. Usevitch, A Tutorial on modern lossy wavelet image compression: Foundations of JPEG 2000, IEEE Signal Processing Mag., vol 18, pp 22-35, Sept [4] Amir Said and William Pearlman, A new, fast and efficient image codec based on Set Partitioning In Hierarchical Trees, IEEE Trans. On Circuits and Systems for Video Technology, vol., pp , June 199. [5] Debin Zhao, Y.K Chan and Wen Gao, Low complexity and low memory Entropy coder for image compression, IEEE Trans. On Circuits and Signals for Video Tech. Vol.11, pp , Oct [] Ziang Xiong, K.Ramachandran and M.T.Orchard, Space Frequency Quantization for wavelet image coding, IEEE Trans. On Image Processing, vol., pp. -93, May 199. [] Farid Ghani, E.Khan and S.Hami, A modified JPEG Image Compression Technique, IETE Journal of Research,vol.4, pp , Oct [8] J.L.Mannos and D.J.Sakrison, The effect of visual fidelity criterion on the encoding of images, IEEE Trans. Inform. Theory, vol. IT- 20, pp , July 194. Image Input DSF DWT CSF mask EZW Encoder Huffman Encoder Compressed data Monitor Image display USF EZW Decoder Huffman Decoder Fig 1. Block diagram of coder and decoder W 2 a1 d1 a2 d2 a3 d3 a4 d4 a5 d W Fig.3. Luminance CSF curve Fig.4. 1-D DWT of CSF curve

5 Appendix-A: Result Obtained Table 1 No. Of Passes (a) Lena using EZW No. Of Passes (b) Lena using DSF and median filter as USF (c) Barbara with EZW (d) Barbara with DSF & (e) Barbara with DSF, CSF median filter as USF & median filter as USF Sl. Technique Passes No. 1 Mask + EZW (2 level decomposition) 2 Mask + DSF + EZW + USF (median) DSF + Mask + EZW + USF (median) DSF + Mask + EZW + USF (average) Table 2: CSF mask with EZW encoder for various techniques P S N R lena --girl -. baboon x goldhill * cameraman Fig 5. PSNR plots for different images of different sizes using DSF and USF: Girl (512x512), Lena (512x512), Baboon (512x512) Goldhill (25x25), Cameraman (25x25) Bits/Pixel Note: The images for the results obtained, are available with the authors and could not be included due to space constraint.