The Adaptive Arithmetic Coding program, written in ANSI C, consists of a number of easy to execute modules. When all the background is removed by extracting the anatomical features, a compression ratio of around 7:1 can be achieved ([16]).

Lossy coding by vector quantization
         In the current design of AFLC-VQ, a four level wavelet decomposition (Figure 2) is followed by vector quantization with eight bits for vector indices in the first level and nine bits for the other levels for the lowest compression used. To achieve higher compression, a combination of quantization of the higher two levels and selective retention of subimages of the first two levels are used. This decoder has the identical codebook for regenerating all the vectors and an inverse wavelet transform module to reconstruct the original image. The fidelity of the reconstructed image is evaluated by computing the MSE ( mean square error ) and PSNR ( peak signal to noise ratio ). Perceptual metrics have not been incorporated in the current design.

         For color images, the image is traditionally split into the RGB color planes. The process is then repeated for each color plane and a codebook is generated for each color plane. At the decoder, the reconstructed color planes are synthesized to form the original image.

         Based on the statistical analysis of the wavelet coefficients, we have chosen 12 filter taps of Daubechies wavelet (DAUB12) for wavelet decomposition to ensure minimum variability in the wavelet coefficient distribution ([13], [18]). This also allowed us to vary the control parameters in the clustering algorithm smoothly to achieve a desired bit rate. The original size of the Visible Human color image a_vm1480 is 2048 x 1216 (Figure 1). Vector quantization is performed using the size 2048 x 1024. The reduction in size is done to make each side as a multiple of two as required by the algorithm, however this reduction has been done carefully so as not to exclude any part of the relevant image.

         Figure 2 shows the 4-level wavelet transform of a_vm 1480 image. Figure 3 and Figure 4 show relatively less visual distortion in the reconstructed images introduced by AFLC-VQ when compared to JPEG, and EZW reconstruction at compression ratios, 74:1, and 132:1. Table II shows the mean square error and peak signal-to-noise ratio of the reconstructed images between AFLC-VQ, JPEG and EZW. It is noteworthy, however, that AFLC-VQ at this stage does not include a subsequent entropy coding stage after the quantization stage as in JPEG and EZW. Therefore the CR values in Table II for AFLC-VQ should be even higher when the entropy stage would be included and would represent less MSE than those computed for JPEG and EZW for corresponding CRs.

Selection and Transmission of Compressed Visible Human Images over TCP/IP Using EZW

         An experimental implementation of the necessary software to allow an Internet user to retrieve and view male Visible Human color images using the lossy compression algorithm Embedded Zerotrees Wavelet (EZW) has been done as a collaborative effort between Texas Tech University and the Communications Engineering Branch at the Lister Hill National Center for Biomedical Communications. A user can select a quality factor through a graphical user interface (Figure 5) to help with the slice selection procedure, and download the slices to the clients’ memory media. The user may then decompress the images using the viewer application, display the images to the screen, and save the images in raw data format in the clients’ computer memory. The Embedded Zerotree Wavelet coefficient algorithm was chosen as the compression technique as a comparative benchmark for AFLC-VQ compression because it had been shown to have a better performance compared to the JPEG standard when compressing an RGB images at compression ratios higher than 50:1 ([10]). At these compression ratios the images compressed and reconstructed with the JPEG standard show a mosaic effect introduced by the blocking artifact induced by 8 x 8 blocks used by the algorithm. The EZW algorithm does not show a blocking artifact since it works with the entire image at different resolutions quantizing the coefficients obtained from the wavelet decomposition. However, at high compression EZW tends to introduce more blurring than the AFLC-VQ. Figure 6 shows a schematic diagram for selecting different quality images using a graphical user interface (Figure 5).