Blind image quality prediction with hierarchical feature aggregation

Abstract

Blind image quality assessment (BIQA) aims to evaluate the perceptual quality of image with no pristine image for comparison, which attracts extensive attention and is of wide applications. Research on human visual system (HVS) indicates visual perception is classically modeled as a hierarchical process. Meanwhile, empirical evidence shows that different levels of distortion generate individual degradation on hierarchical features. Although previous BIQA methods exploited different levels of feature, the degradations and correlations on hierarchical features are not detailedly analyzed. In this paper, hierarchical degradation is systematically analyzed by quantitative and qualitative experiments using interpretable hierarchical features. Additionally, as there exist strong relationships, these features cannot be simply concatenated for quality prediction. Inspired by the Bayesian brain theory, correlations among hierarchical features are thoroughly deduced with joint probability. Guided by the mathematical derivation, a new feature aggregation network (FANet) is designed to eliminate correlations for hierarchical features fusion. Finally, a novel end-to-end BIQA framework based on hierarchical feature aggregation is proposed, which eminently analyzes the degradations and correlations on hierarchical features. Experimental results on five benchmark databases show that the proposed method achieves state-of-the-art performance. Furthermore, cross-database evaluations demonstrate the generalization capability and robustness of the proposed method.