A block partitioning vector quantization method through post-processing to achieve higher PSNR

Abstract

In conventional vector quantization (VQ) method, the reachable best PSNR is implemented by full search or its equivalent method when the codebook is fixed. To achieve an even higher PSNR by fixed codebook, a post-processing VQ method is proposed . This method firstly checks how well the VQ matching is for a whole image and locates some bad matching image blocks definitely. Then it partitions a bad matching image block into two partial blocks further by a certain pattern and encodes them once again using the same fixed codebook but two different codes partially. It makes total distortion for current bad matching image block smaller. At the receiving end, the current bad matching image block is synthesized by two codes partially according to the adopted partitioning pattern so as to realize a better representation. For 10 gray-level standard images of size 512/spl times/512 with different detail, a 0.3 dB 1.2 dB PSNR improvement can be reached individually by just partitioning 5% worst matching image blocks at the cost of less than 1 bit degradation of equivalent code length. Moreover this improvement is basically for the edge portion of an image. Therefore, subject assessment can be made much better at the same time.