Image quality assessment with mean squared error in a log based perceptual response domain

Up to now, there existing a lot of models that predict subjective quality of the contents of natural images which have undergone some unknown distortion procedures. These models, no matter fall in to the bottom-up mechanism or belong to the top-down functional modelling, fail to provide an easy-applied and reliable solution. The complex computation procedure prevents them from being widely used in related image processing areas such as image enhancement, image reconstruction and video coding. In the present work, we start from a two stage nonlinear perception model, which transforms the input image into a decorrelated one and then further reduces the redundancy between neighboring pixels by another nonlinear normalization procedure which transforms the previous output into a perceptual response domain. The final quality prediction is computed as the Euclid distance of the reference image and the distorted one in this response domain, this will make the new model be readily applied in other areas.