Hierarchical multi-scale cross-attention learning for blind image quality assessment

Abstract

Blind image quality assessment (BIQA) remains a challenging task in computer vision due to the absence of pristine reference images. This paper proposes a novel Hierarchical Multi-Scale Cross-Attention Network (HMCANet) that effectively captures both local distortion patterns and global semantic information for quality prediction. The proposed architecture integrates a Swin Transformer backbone with a novel Cross-Scale Attention Module (CSAM) to aggregate multi-level features hierarchically. Additionally, a Distortion-Aware Feature Enhancement (DAFE) block is introduced to amplify quality-relevant representations while suppressing irrelevant noise. A composite loss function combining mean squared error and rank-aware loss is designed to improve prediction accuracy and monotonicity. Extensive experiments are conducted on four benchmark datasets, including LIVE, CSIQ, TID2013, and KADID-10k. The experimental results demonstrate that HMCANet achieves state-of-the-art performance with SRCC values of 0.971, 0.954, 0.923, and 0.917 on the respective datasets. Furthermore, ablation studies are performed to validate the effectiveness of each proposed component. The cross-dataset evaluation results indicate that the proposed method exhibits superior generalization capability compared to existing approaches.