Perceptually and statistically decorrelated features for image representation: application to transform coding

Transform coding consists of a scalar quantization of the features of an image representation. These features should be independent enough to justify the scalar approach. The coefficients of the commonly used DCT representation still show some dependence that may reduce its efficiency. In this work, a perceptually inspired nonlinear transform is used to map the DCT into a new representation that largely reduces the statistical and perceptual relations between the coefficients thus improving the compression performance.