Encoding of still pictures with successive clustering and predictive vector quantization

Abstract

AbstractThis paper describes the predictive quantization of still pictures using the newly developed clustering with faster convergence than L.B.G. algorithm. In the proposed scheme, the input vector is separated into the average density and the deviation from the average (average density separated vector). Then the latter is quantized predictively using two code books of category 1 and category 2. Examining the utilization frequency of the codeword, the coding efficiency is improved by suppressing the redundancy. The code book of category 1 is designed using 256 codewords and successive clustering. Letting [] be the codeword used in the quantization of the immediately preceding input vector, the code book of category 2 is composed of the set J([]) with high probability to be used in the next quantization. It is predicted that the input data following the codeword [] is quantized by an element of J([]), and encoding is performed by the index attached to that element. The number of elements in J([]) is 4 to 32 including the control code, which is less than that of the code book 1, realizing an improvement in the coding efficiency. If the codeword to be encoded is not found in J([]), the code book 1 is used. The codewords in code book of category 1 are arranged in the order of their variances, which reduces the retrieval time in the code book to one‐ninth to one‐sixth the total retrieval time. Using the proposed scheme and seven still pictures with 256 × 240 pixels and 256 gray scales as the training sequence, eleven pictures including the forementioned seven were encoded. The result revealed that 0.8 to 0.5 bits/pixel is required and the signal‐to‐noise ratio is 30 to 36 dB, indicating the excellent performance of the proposed scheme.