Abstract
We have already proposed Learning Motion Vector Quantization(LMVQ)that can eliminate the optimal codebook design procedure on Motion Vector Quantization(MVQ). In addition to motion estimation, this LMVQ method is simultaneously to do the codebook learning according to the nature of input image. In this paper, we will show a learning method that refers to the frequency counters of codewords at codebook learning and report the simulation results employing the Flower Garden video data. In performance evaluation of motion compensation, the SNR of this method was better than that of MVQ. In comparison with MPEG-1, although the SNR of LMVQ was about 0.4dB poorer, we could achieve the significant results such as reducing the computer amount of motion compensation to 1/11 and shortening the total processing time to 33%.
