ROI-based image coding using multiresolution neural networks

In this paper is presented a ROI-based multiresolution coding scheme, whose main importance is that it achieves both high compression ratios and good reconstruction of the images. It uses optimal, in the mean square error sense, analysis and synthesis filters in the most significant areas (Regions of Interest) while conventional ones are used in the rest of the image. A linear autoassociative neural network architecture is proposed to compute the filters for optimal reconstruction of the images based on low resolution approximations of these. The characteristics of optimal filters are examined in ‘head and shoulder’ videoconferencing images.