Learning to integrate local and global features for a blind image quality measure

In this paper, we present a new algorithm for blind/no-reference image quality assessment (BIQA/NR-IQA). Most existing measures are “opinion-aware”, demanding human opinion scored images to map image features to them. The task of obtaining human scores of images is, however, commonly thought to be uneconomical, and thus we focus on “opinion free” (OF) quality metrics in this research. By integrating local and global features, this paper develops a learning-based BIQA approach with three steps by combining local and global features together. In the first step of extracting local features, we use the quality aware clustering with the centroid of each quality level trained by K-means, while we in the second step compute the global features based on the natural scene statistics. Finally, the third step uses the SVR to train a regression module from the above-mentioned local and global features to derive the overall image quality score. Experimental results on LIVE, TID2008, CSIQ, and TID2013 databases validate the effectiveness of our proposed metric (a general framework) as compared to popular no-, reduced- and full-reference IQA approaches.