It is well known that one can always obtain better performance by coding source data in the form of vectors instead of scalar. Vector Quantization (VQ) is a coding scheme for mapping a sequence of continuous or discrete vectors into a digital sequence suitable for communication over or storage in a digital channel. There are mainly three problems of VQ systems for image data compression. The first is large computation and memory costs, intrinsic in VQ systems. Secondly, VQ suffers from block effects in reconstructed images, which always occur in block-wised coding at low rates. In addition, VQ systems exhibit some degradations in performance due to the statistical difference between the image used for codebook design and encoding images.
In this thesis, we review various VQ methods proposed so far in order to meet the above problems. The main obstacle limiting the performances of low-rate VQ systems is the small codebook size not sufficient to cover the full dynamic range of encoding vectors.
To meet the obstacle, and LPVQ (Linear Predictive VQ), originally employed for speech coding, is tried for image data compression. In the encoding process, a current vector is backward-predicted using neighboring pels around block boundaries and the predicted vector is subtracted from the current vector, and then the resulting vector is coded by a VQ.
Computer simulation results show that LPVQ exhibits better performance than previous VQ methods such as IVQ (Interpolative VQ) and MSVQ (Mean-Separated VQ), which use forward-prediction methods. Especially the block effects in the reconstructed images by LPVQ can be significantly reduced.