Image Quality Assessment Based on Mutual Information in Pixel Domain

The natural scene statistics (NSS) model is widely used in image quality assessment algorithms, the NSS based features in frequency domain provide a good approximation to image structure, but not to the image content. To get a metric which is effectively to both structural distortion and content distortion, a new image quality assessment framework in image pixel domain based on mutual information is proposed. First, a non-overlapping segmentation set is acquired to establish the relation with image pixels. Second, the saliency and specific information are measured to catch the image content changes, and entanglement to the image structure change. Finally, the differences of image content and structural information are used to measure image quality. The experimental results show that the proposed framework has good consistency with subjective perception values.