Some Experiments in Adaptive and Predictive Hadamard Transform Coding of Pictures

In a recent paper 1 we considered Hadamard transform coding of still pictures using a small three-dimensional block (a 2 X 2 X 2 array of picture elements). There we described the design of optimum quantizers for the Hadamard transform coefficients based on psychovisual criteria in the transform domain. Starting with subjective tests to evaluate the visibility of quantization noise, we then developed a design procedure to minimize the "mean-square subjective distortion" (MSSD) due to quantization noise. We compared the performance of the resulting quantizers with the widely used Max-type 2 quantizers (i.e., quantizers which minimize the mean-square quantization error) and demonstrated our quantizers to be better in terms of picture quality and entropy of the quantizer output, for a given number of levels. The present paper, which consists of three parts, extends the previous 1531 work by considering techniques for adaptive and predictive coding of the transform coefficients, based on both statistical and psychovisual criteria. In the first part, we develop prediction algorithms for predictive coding of the coefficients. Although the small block size that we use ensures that the quantization noise can be placed in those parts of the picture where it is least visible, thereby permitting coarser quantization and thus achieving a higher coding efficiency, it does not exploit the statistical correlation between adjacent blocks. To overcome this, we consider predictive coding for the coefficients.