Principal features of human vision in the context of image quality models

Essential aspects of the visual processing of monochrome still pictures in the human observer are: Adaptivity to scene illumination, decomposition into scale and orientation selective features, and pattern specific masking. In the model suggested here these aspects are realized in a psychophysically and physiologically plausible fashion. Adaptivity is provided by local automatic gain control. Applying this gain control in parallel to a number of levels of resolution leads to a ROG (Ratio of Gaussian)-filter pyramid at the input stage of the system. Each ROG-filter output is further decomposed by orientation selective filters. Pattern specific masking is described by saturating nonlinearities acting at each pixel of the filtered outputs. The final quality measure is derived as the inverse of a vector norm difference between the feature vectors of original and disturbed image. The model is applied to a variety of distorted and coded images.