No reference image quality assessment based on statistical distribution of local Sub-Image-Similarity

Abstract

The research on no reference image quality assessment (NR IQA) is the most attractive one in the area of image quality perception. In this paper, we propose to use the statistical distribution of local Sub-Image-Similarity (SIS) measures for NR IQA model design. Here the mean and the difference properties among the local SIS measurements in different directions are synthesized into five quality labels to depict the perceptual quality property of deteriorated images. The proposed NR IQA model is developed based on the statistical distribution of quality labels over whole image, via a SVM regression. Experiments show that the proposed model performs best according to the predictive accuracy when compared to the published NR IQA models, and works stably with different parameter selections and cross database evaluations.