Markov Random Field-Structured Direct Sum Residual Vector Quantization for Classification

Multistage RVQs with optimal direct sum decoder codebooks have been successfully designed and implemented for data compression. The same design concept has yielded good results in the application of image-content classification and has also provided an effective platform to perform image driven data mining (IDDM). To make it computationally feasible, the current design methods entail encoder codebook designed in a sequential but suboptimal manner. Based on the sub-optimal codebook design approach, the sequential search path is greedy based on a stage wise nearest-neighborhood strategy instead of a direct sum nearest-neighborhood requirement. Markov random field (MRF) provides a suitable framework to exploit the structure of multistage residual vector quantizers with optimal direct-sum direct sum decoder codebooks combined with sequential-search encoders to achieve optimized classification in the maximum aposteriori sense (MAP).