Blind image quality assessment via aesthetic dataset construction and hierarchical feature fusion

Blind image quality assessment (BIQA) has significantly progressed due to rapid advancements in deep learning techniques. However, the objective of the BIQA problem remains ambiguous and is typically approached from two perspectives: the technical perspective, which evaluates the perception of distortions; and the aesthetic perspective, which focuses on content preference and recommendations. Most existing studies predominantly focus on the technical perspective, with relatively few studies addressing the aesthetic perspective. To address this problem, this paper proposes the Aesthetic-Technical Aggregation Quality Assessment (ATAQA) framework by leveraging aesthetic dataset construction and hierarchical feature fusion. Specifically, to enhance aesthetic expression, we first design the Pre-trained Aesthetic-Technical Aggregation (PATA) strategy, whose capabilities for aesthetic feature learning are improved by the Image Aesthetic Quality Dataset (IAQD). Further, we design the Dense Feature Aggregation (DFA) module, that integrates the transformer features hierarchically into the quality-aware feature representation, enabling the model to utilize visual information from low to high levels. Extensive results on several IQA datasets demonstrate that ATAQA significantly outperforms current state-of-the-art (SOTA) methods. Our code will be available after the paper is accepted.