A post-processing method for vector quantization to achieve higher PSNR and nearly constant bit rate

Abstract

In conventional vector quantization (VQ) encoding methods, the reachable best PSNR can be realized by a full search (FS) equivalent method when the codebook is fixed. To achieve even higher PSNR and an almost undegraded bit rate, a post-processing method for VQ encoding is proposed. This method first checks how well the matching is for an image block after conventional VQ by its block PSNR/sub b/. Then it partitions a badly matched image block into 2 partial blocks by a certain pattern (a pair of masks) and encodes them again just within a small already encoded nearby area. The overall matching can generally be improved by using 2 smaller partial blocks. Meanwhile, bit rate can be kept nearly constant by using 2 shorter winner pointers instead of 2 winner indexes. At the receiver, the badly matched image block is synthesized in reverse to realize a better reconstruction. In the case of the low-detailed image Splash, image PSNR can be unproved more than 1 dB for the 1024 codebook while the bit rate has almost undegraded.