A self organizing map approach to image quality.

A block-wise vector quantization of images was carried out by using different types of self organizing maps. The quality of the coded images was estimated by human observers with the result that good image quality assumes a balance of two types of image blocks, namely flat regions and edges. On the basis of these types of blocks a segmentation error model was derived which enables to estimate the quality of a distorted image in comparison with an original. In this model the segmentation of errors was done block-wise by a classification into those types found as important for image quality. An error metric was formed by a non-linear combination of segmented components over space and classes which for a large set of images and distortions shows better conformity with the quality estimates of human observers than standard technical measures.