Dual-quality map based no reference image quality assessment using deformable convolution

Abstract

Traditional feature-based no-reference image quality assessment (NR-IQA) technologies rely heavily on the correlation between the handcrafted features and distortions, and often fail to predict perceptual quality because of the diversity of image distortion types. End-to-end deep learning-based metrics are driven mainly by human-labeled data, which are prohibitively labor-expensive to collect. An NR-IQA method that imitates the hierarchical and inferential perceptual behavior of the human vision system (HVS) is proposed to address the challenges of distortion diversity and lack of labeled data. In the proposed method, first, a dual-quality map is generated for use as an intermediate target, which contains multiscale features that are highly related to image quality. To capture human eye-sensitive distortion information in distorted images, a dual attention mechanism is utilized in the quality map generation network. Then, a regression network is designed for mapping the intermediate target to a quality score. The deformable convolution which has adaptive receptive field is introduced to extract multiple scale features from the dual-quality map. Experiments on eight benchmark databases show the superiority of the proposed method over other state-of-the-art NR-IQA metrics.