COMPRESSION OF COLOUR IMAGES, BASED ON NON-LINEAR PERCEPTUAL REPRESENTATIONS AND MACHINE LEARNING

摘要

The invention relates to the compression of colour images, based on non-linear perceptual representations and machine learning. More specifically, the invention relates to a method for the compression of colour images, comprising the following steps: (1) expressing the coefficients of a local space/frequency representation of the opposing chromatic and achromatic channels in units of perceptual contrast; (2) applying a non-linear transform with divisive normalisation to each channel, such that they include relationships between the coefficients of the achromatic and chromatic contrast channels; and (3) using statistical learning algorithms to select the relevant data from the non-linearly normalised channels. Psychophysical and numerical experiments have been conducted in order to estimate and validate the parameters of the divisive normalisation and contrast transforms. In addition, the above method has been compared experimentally with a JPEG implementation. In both cases, 16x16 DCT blocks were used and, in each case, comparable values were used for the frequency-dependent and chromatic transform parameters. Under these conditions, the experimental results provide an average compression gain of 50% in relation to JPEG for different distortion measurements (RMSE, SSIM and S-CIELab) in the realistic commercial range of [1, 2.2] bits/pix. The invention also relates to the apparatus and code used to implement said method.