Wavelet-based robust image coding and packetization techniques

In this research we primarily focus on the wavelet-based robust image coding and packetization techniques. The discrete wavelet transform is used throughout this research. One of the major objectives of the research is the development of efficient and robust packetization scheme for the transmission of SPIHT encoded bitstreams over packet erasure channels. We present several approaches on robust coding and packetization of SPIHT bitstreams. The first approach is based on genetic algorithms. Each packet is associated with a decision bit, which is mapped to a gene of a chromosome. In this way, each specific packetization instance determined by a sequence of decision bits is mapped to a chromosome. Next, a steady-state genetic algorithm is applied to search for the optimal chromosome that minimizes the distortion between the original and the reconstructed image. The second approach is based on the dynamic programming technique. We first formulate the optimal packetization problem as a discrete constrained optimization problem of finding the optimal packing vector from its vector space subject to a strict robustness constraint, and then provide a heuristic algorithm to solve the optimization problem using dynamic programming. Finally, a two-step process based on our proposed Packetized Error-Resilient Entropy Code (PEREC) algorithm is presented. In the first step, a multiplexer is used to partition the embedded bitstream into multiple variable-length blocks with block sizes close to the fixed packet size. In the second step, the PEREC algorithm is proposed to reorganize these variable-length blocks into fixed-size packets.; The other important objective in this research is parallel image coding techniques for multi-core DSPs or MIMD processors. Two parallel image coding algorithms are presented. The first algorithm is based on the optimal quadtree segmentation and VBVQ. Coefficients from the same subband are organized as a quadtree and encoded in parallel. The low-low band is coded by DPCM. The other subbands are independently coded by Variable Blocksize Vector Quantization (VBVQ). For a given bit budget, this scheme is capable of finding an optimal quadtree segmentation for each subband so that the cumulative distortion is minimized. The second algorithm is based on a memory-efficient SPIHT algorithm. The wavelet coefficients are partitioned into tree blocks to enable parallel processing. Each tree block is further re-organized into line buffer to facilitate DMA transfer between the on-chip and off-chip memory to ensure on-chip memory access. The three dynamically-allocated temporary lists of the original SPIHT algorithm are replaced with fixed-size circular queues to reduce memory requirement. Finally, the Lagrange multiplier method is applied to find the optimal rate allocation for each tree block.