Ensemble Strategy for Underwater Image Quality Assessment and Dataset Construction

Learning-based Underwater Image Enhancement (UIE) methods have made significant progress. Limited by the manual label selection process, the limited quantity and outdated label quality of UIE datasets have severely hindered the development of UIE society. The urgent demand for more and better paired training samples motivates us to propose Ensemble-Select (E-Select), a strategy that can serve as an alternative to fully manual annotation and can enable continuous expansion of dataset size and optimization of label quality. However, expanding size will encounter new images, optimizing labels will encounter new algorithms and the proposed strategy is required to maintain strong generalization in both scenarios, which overwhelms many IQA methods. This work improves generalization in two ways. First, three primary influencing factors and their interrelationships in quality assessment are systematically analyzed. Specifically, we first explore the interactive relationship between content and distortion perception, and further investigate the guiding value of aesthetic-aware features in image quality perception. Then, the proposed Distortion-Content Interaction Module (DCIM) enables the network to focus on perceptually important distortion features guided by content. Second, we investigate a multi-perspective quality evaluation framework based on the ensemble learning paradigm. Building upon the availability of numerous outstanding IQA works, we initially demonstrate their distinct excel regions and evaluation biases. Subsequently, we explore the ensemble of their results through the proposed Aesthetic-Guided Quality Regression module (AGQR), which generates dynamic quality regression layers and derives image-specific quality perception rules based on aesthetic features. We then construct the first expandable, updatable UIE dataset with the help of E-Select. We collect over 50k real underwater image pairs with optimal labels, covering diverse scenes and varied degradation characteristics. Unlike other datasets, our dataset can consistently expand the number of paired samples and maintain optimal labeling without requiring extensive human labor. Experiments show the facilitating effect of the newly constructed dataset on UIE and the SOTA performance of E-Select. Codes and datasets are available at URL.