Quality-Aware Domain Adaptation for Underwater Image Enhancement Quality Assessment

As the demand for high-quality underwater images continues to grow in the field of underwater research， underwater image enhancement （UIE） algorithms have been widely applied. To evaluate the quality of these enhanced underwater images， researchers have proposed several underwater image enhancement quality assessment （UIEQA） algorithms. However， UIEQA algorithms trained on known underwater scenes often face challenges when applied in unknown underwater scenes. Additionally， existing UIEQA algorithms typically rely on large amounts of annotated data， which are often difficult and resource-intensive to obtain. To address these issues， this paper proposes a quality-aware domain adaptation-based underwater image enhancement quality assessment （QaDA-UIEQA） algorithm. The proposed method includes quality assessment and quality-aware domain adaptation modules. First， the quality assessment module performs supervised quality assessment training on the source domain data to ensure the accuracy of the main task. Second， the quality-aware domain adaptation module， guided by textual information， used a cross-attention （CA） module to extract important quality characteristic information from visual feature information. Then， domain adaptation techniques were used to narrow the gap in quality characteristics between the source and target domains， thus enabling models trained on known underwater scenes to effectively generalize to unknown underwater scenes. Experimental results on the SAUD+ dataset showed that the proposed method achieves optimal results on four key performance metrics， compared with 13 other existing methods. Among them， the SRCC improved by 8.5%， compared with the second-best model. Additionally， ablation studies demonstrated that our proposed multimodal approach significantly enhances model performance. The proposed method not only exhibited excellent performance in UIEQA but also surpassed other comparison methods in terms of prediction accuracy and generalization capability in a group maximum differentiation competition. Therefore， QaDA-UIEQA has stronger generalization and robustness， and it can maintain efficient and stable performance in complex real-world applications.