Optimized Filtering With Binary Descriptor for Blind Image Quality Assessment

Blind Image Quality Assessment (BIQA) techniques evaluate the perceptual quality of a distorted image without access to its distortion-free version. In this paper, a novel BIQA measure is proposed in which interest points drawn by visually attractive regions in a grayscale image are characterized using a binary descriptor. Then, a regression technique maps the feature space to subjective opinion scores to provide the quality prediction. In this method, an additional image filtering step prior the feature extraction is used. The filtering is obtained as a solution to a problem of finding a correlation between the image quality and the keypoint detection results. The attention of the human visual system (HVS) in presence of distortions, mimicked by an interest point detector, is enhanced using the proposed quality-aware image filtering. A keypoint descriptor which mimics retinal photoreceptors in the HVS is also applied to the filtered images. In this paper, it is shown that the proposed BIQA method provides a highly competitive performance to the state-of-the-art measures on popular large-scale IQA benchmarks. It is also demonstrated that a simple application of the developed quality-aware filtering can improve the results of BIQA measures.