Skip to main content

Level Reassignment: A Technique for Bit-Rate Reduction

01 January 1978

New Image

It is well known that the statistics of picture signals are nonstationary and that the required fidelity of reproduction demanded by the human eye varies from picture element (pel) to picture element. Consequently, for efficient digital representation of pictures, it is desirable to adapt coding strategies to those local properties of the picture signal which determine the visual sensitivity to quantization noise. In this study, we make use of the spatial masking properties of the human observer to adapt the quantization strategies for encoding the picture signal. We define spatial masking as the reduction in the ability of a person to visually discriminate amplitude errors which occur at or in the neighborhood of significant spatial changes in the luminance. To this end, we borrow, from our earlier work, measures of luminance spatial activity in both the transform 1 - 2 and the pel 3 domain, and the relationships (called the visibility functions) between the amplitude accuracy and 61 these measures of spatial detail. We use these visibility functions to change dynamically the input-output mapping of a single quantizer to reduce the bit rates. This is done by reassigning the input of the quantizer to a different representative level than normal in such a way as to reduce the entropy of the quantized output, while keeping the visibility of quantization noise below a certain threshold. We demonstrate three different algorithms for level reassignment and evaluate their potential by measuring the entropy of the quantized o u t p u t for a given picture quality.