Study of correlation vector quantization for image coding

Some adjacent image blocks may have the same address when coding them with vector quantization (VQ), especially in the stationary and uniform regions of an image. This paper proposes a correlation vector quantization scheme to decorrelate the addresses of adjacent coded blocks. Correlation codebook and improved self-organizing feature maps (ISOFM) codebook are used to encode the four adjacent blocks simultaneously in a window. For a typical image Lenna, the computational quantity during the coding process is reduced 50 percent and the bit rate is reduced by about 40% over the memoryless VQ. Since preferential measures are taken to the vectors belonging to the edge classes in the training process of Kohonen's self-organization neural network, the subjective quality of the decoded image is significantly improved.