关键词: 图像质量评价, 人类视觉系统, 视觉关注, 显著目标检测, 注意力机制

Abstract: Image data in actual business scenarios usually presents the characteristics of rich content and complex distortion performance, which is a great challenge to the generalization of objective image quality assessment algorithms. In order to solve this problem, a no-reference image quality assessment algorithm was proposed, which was mainly composed of three parts: feature extraction network, cross-attention fusion network and adaptive prediction network. Firstly, the global view, local patch and saliency view of the sample were input into the feature extraction network, and the global distortion, local distortion and saliency features were extracted by Swim Transformer. Then, the cascaded Transformer encoder was used to fuse the global distortion features and local distortion features, and the potential correlation patterns of the two were explored. Inspired by the human visual attention mechanism, the feature fusion network tried to use saliency features to stimulate the attention module, so that it could exert additional attention to the visual salient region, so as to improve the semantic parsing ability of the algorithm. Finally, the prediction score was predicted by the dynamically constructed MLP regression network. The results show that compared with the DSMix algorithm, the proposed algorithm improves the Spearman Rank Correlation Coefficient (SRCC) on the TID2013 dataset by 4 percentage points, and shows excellent generalization ability and interpretability, which can effectively deal with the complex distortion situation in the business scenario, and can make adaptive predictions according to the individual characteristics of the sample.

Key words: Keywords: Image Quality Assessment(IQA), Human Visual System(HVS), visual attention, Salient Object Detection(SOD), attention mechanism