Jeffrey Brian Erbrecht
Context Adaptive Space Quantization for Image Coding
JPEG, one of the most widely used lossy image coders in the world, has a significant limitation: only one quantization table can be used per colour channel. In this seminar, we describe a new kind of quantization method called context adaptive space quantization (CASQ), as well as a new image coding scheme based on CASQ. This image coding scheme partitions an image into homogeneous and non-homogeneous regions, or contexts; subsequent quantization is then conditioned on these contexts, achieving bit rate reduction while simultaneously preserving more information from the perceptually important parts of the image. We formulate CASQ in both hard-decision quantization (HDQ) and soft-decision quantization (SDQ) settings. For HDQ with JPEG-style Huffman coding, CASQ yields an average gain of 1.78~dB PSNR performance compared to the standard JPEG quantization table, and 0.23~dB compared to state-of-the-art HDQ methods. Using adaptive runlength coding instead of Huffman, these gains are even higher. Significant gains are also observed in the SDQ setting. All of this comes at a very low increase in computational complexity.