Self-information weighting for image quality assessment

Several recently proposed image quality assessment (IQA) methods involve a two-stage structure: local distortion measurement followed by pooling. Based on the hypothesis that more weights should be assigned to the image components that contain more information, this paper explored the potential of a Shannon Self-Information based pooling strategy, where self-information measures the “surprisal” of seeing a local image patch in the context of its surround. We combined the self-information based pooling strategy with the multi-scale structural similarity (MS-SSIM) index, yielding a self-information weighted SSIM (SI-SSIM) approach. Extensive evaluations based on six publicly available databases show that the proposed SI-SSIM approach achieves superior or comparable performance as compared with a number of competitive IQA algorithms.