Method, apparatus and software for color image compression based on non-linear perceptual representations and machine learning

Ref. P200801943 Oficina Española de Patentes y Marcas (jun. 20th 2008)

It is provided a method for color image compression based on the following key steps: (1) expressing the coefficients of local space-frequency representation of achromatic and chromatic opponent channels in perceptually meaningful contrast units, (2) applying divisive normalization non-linearities including relations among the coefficients of the achromatic and chromatic contrast channels, and (3) using machine learning algorithms to select relevant data from the non-linearly normalized channels.

Psychophysical and numerical experiments have been conducted to estimate and validate the parameters of the contrast and divisive normalization transforms. Besides, an implementation of the proposed method has been experimentally compared to a JPEG implementation. In both cases 16×16 DCT blocks were used, and when applicable, equivalent chromatic and frequency dependent parameters were taken for a fair comparison. In these conditions, experimental results on a 25 color image database show that the proposed method gives rise to substantial compression gain over JPEG at the same (RMSE, SSIM and S-CIELab) distortion in the commercially meaningful bit-rate range [1, 2.2] bits/pix. Depending on the distortion measure applied, the compression gain is about 120% for the same RMSE, 45% for the same SSIM, and 170% for the same S-CIELab.

Apparatus for implementing the method and software product are also claimed. 

References: 40