Deformable Attention-Optimized Swin Transformer for Image Aesthetic Assessment
Year: 2025
Type: conference-paper
Abstract: Image aesthetic evaluation aims to enable computers to simulate human aesthetic thinking and objectively evaluate the aesthetic quality of images automatically. The existing methods often have a rigid convolutional attention mechanism, making it difficult to achieve hierarchical integration of multi-scale representations, and the computational cost is relatively high. This paper proposes a deformable attention fusion model based on Swin Transformer and integrates deformable attention for the first time in aesthetic assessment. A deformable self-attention module is embedded in the hierarchical Transformer, and the key aesthetic regions in the image are adaptively focused by the model through dynamic offset prediction. The multi-task pre-training strategy combining mask reconstruction and contrastive learning enhances the representation ability of the model. Experiments on the AVA dataset show that the performance of the model in this paper is superior to that of the traditional CNN model, and the deformable attention mechanism plays a positive role in extracting image features and evaluating aesthetics.
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7127053141