Geometric coding for error resilient image compression
Andy C. Hung · 1995
Digital image and video compression is crucial for the seamless integration of visual data into multimedia systems. It reduces the high volume of visual data and increases the available transmission bandwidth to others. The traditional compression standards such as JPEG and MPEG focus on compression over networks and for storage where there is some guarantee on the reliable reception of information. This research focuses on compression techniques for the emerging mobile computing environment, where transmitted bits are often corrupted or lost due to channel noise. Since mobile channels experience noise that fluctuates rapidly and differs between receivers, the compression technique must be inherently resilient to potential data corruption. In a nutshell, compression is achieved through exploiting the smooth character of natural images and eliminating information that isn't noticeable. The typical way to accomplish this is through transforms or subband filters, which separate the slowly varying image data into different frequency bands. The data in each frequency band has often been modeled with Laplacian-like probability densities. Geometrically, in high dimensions, these Laplacian densities have equiprobable surfaces which are shaped as multidimensional pyramids. This thesis develops methods of error resilient image compression which exploit the geometric structure of multidimensional Laplacian probability densities, yielding elegant compression solutions that improve image compression performance and increase robustness to channel bit errors and fades. The first part of this thesis method applies geometric rotations to subband filtered image data to align the probability contours with the axes. This increases the performance of fixed rate scalar quantization up to 4 dB on typical image data and preserves image details and contours in the presence of severe bit errors and fades. The second half of this thesis focuses on Pyramid Vector Quantization (PVQ), which is a method to encode the points on the probability pyramid surfaces. We introduce new PVQ encoding methods that reduce the effects of channel error at high rates by 3 dB over previous techniques. Our PVQ subband compression consistently outperforms the standard image compression algorithm JPEG with custom generated Huffman tables, and exhibits high robustness to data corruption. The techniques in this thesis facilitate efficient, reliable wireless transmission of high quality images and video.