No-reference image quality assessment via bidirectional feature fusion and regional distortion extraction

Image quality assessment is a core technology in image processing, primarily used to evaluate the quality and distortion level of images by analyzing their characteristics. It serves as an important metric for other visual tasks. While significant progress has been made in evaluating synthetically distorted images, assessing authentically distorted images remains a challenge. Unlike synthetic distortions, which typically have uniform distribution, authentic distortions are more complex, with unevenly distributed distortion regions and varying distortion types and levels. Moreover, authentic distortions are influenced by both high-level semantic features and low-level visual features. To address these challenges, we propose a no-reference image quality assessment algorithm based on bidirectional feature fusion and regional distortion extraction. Human-perceived visual quality is influenced by both low-level visual features and high-level semantic features, a relationship that is particularly evident in authentically distorted images. To capture this, the proposed method employs a bidirectional feature fusion structure, which integrates both high-level and low-level visual information through top-down and bottom-up pathways, respectively, thereby combining multi-layered features. Synthetic distortions are usually globally distributed, while authentic distortions typically manifest regionally, which complicates their evaluation. To address this, the proposed method uses window attention to extract local distortion features and overlapping cross-window attention to strengthen the interconnections between local distortions, yielding regionally distributed distortion features. Finally, the paper trains a model using contrastive learning to extract distortion features for various distortion types and levels. A mixture-of-experts cross-attention module is introduced to fuse distortion and image features. Experimental results on two synthetic distortion datasets and three authentic distortion datasets demonstrate that the proposed method achieves competitive performance.