Reduced-reference image quality assessment based on distortion families of local perceived sharpness

Previous research on reduced reference (RR) image quality assessment (IQA) suggested that appropriate RR features should provide efficient summaries of reference images and be sensitive to a variety of image distortions. The multi-scale local sharpness maps are effective RR features because they can capture smooth, edge, and textured areas of the reference image, and they are affected differently by different distortion types. Motivated by this observation, in this paper, we propose an efficient RR IQA algorithm using local sharpness. Our method, called S4RR, employs four sharpness maps (two FISH maps and two local standard deviation maps) to assess image quality via two main stages. The first stage soft-classifies the distorted image into eight distortion families based on an analysis of the different scatter-plot shapes of the sharpness map values of distorted image vs. reference image. The second st age performs distortion-family-specific quality assessment based on measuring the local sharpness variations between reference and distorted images by using seven types of local statistics and six distance measures. Finally, the soft-classification probabilities computed from the first stage are combined with the distortion-family-specific quality scores to yield a class-weighted average, which serves as the final S4RR quality index. Experiment results tested on various databases show that with less than 5% RR information, the proposed S4RR algorithm achieves better/competitive performance as compared to other state-of-the-art FR/RR IQA algorithms.