Adaptive wavelet compression of imagery.

Abstract

The discrete wavelet transform has emerged as a powerful tool for the lossy compression of imagery. In this work, two adaptive still image wavelet coders are presented, each based on a variant of trellis coded quantization (TCQ). TCQ is an effective technique for quantizing memoryless sources with moderate complexity. Entropy-constrained trellis coded quantization (ECTCQ) and a newly developed quantizer, universal trellis coded quantization (UTCQ) form the basis of each system. UTCQ offers several advantages over ECTCQ. It requires no computationally expensive training algorithm, no stored codebooks, and performs "on-the-fly" training on a subset of the reconstruction levels. It is shown that the performance of eight state UTCQ is within ≈0.1 dB of eight state ECTCQ for most encoding rates. Unlike most systems present in the literature, these coders are complete end-to-end image compression/decompression systems. Adaptive arithmetic coding is used to generate output bitstreams. Several wavelet subblock classification schemes are investigated to improve the PSNR performance of the coders. The 3-map classification technique was seen to improve SNR performance by ≈0.8 dB for the Lenna image above 0.5 bpp. Perceptual coding improvements were developed for the UTCQ system as well. The image coders proved to be quite robust, performing well on a wide array of imagery. The quality of the encoded imagery from these systems is found to be very competitive with other coders in the literature and significantly better than JPEG-based coding.