A comparison of the visual effects of two transform domain encoding approaches

TLDR: The results of this experiment suggest that image appearance may be improved by designing transform coefficient quantization rules to approximate the effects of additive noise rather than to omit low energy image components, as dictated by conventional rate-distortion theory.

Abstract: Rate-distortion theory using the mean squared error criterion is often used to design digital image coding rules. The resulting distortion is, in theory, statistically equivalent to omitting components of the image from transmission. We compare a rate-distortion simulation using the discrete cosine transform to a method which is statistically equivalent to adding uncorrelated random noise to the image. This latter method is based on a PN (pseudo-noise) transform, which is generated from a Hadamard matrix whose core consists of the cyclic shifts of a binary maximum length linear shift register sequence. Visual comparisons of the two approaches are made at the same mean squared error. In all cases, the images encoded using the PN transform method showed superior definition of detail and less geometrical distortion at transform block boundaries than the images encoded using the discrete cosine method. The results of this experiment suggest that image appearance may be improved by designing transform coefficient quantization rules to approximate the effects of additive noise rather than to omit low energy image components, as dictated by conventional rate-distortion theory.© (1977) COPYRIGHT SPIE--The International Society for Optical Engineering. Downloading of the abstract is permitted for personal use only.