Dual-attention pyramid transformer network for No-Reference Image Quality Assessment

No-Reference Image Quality Assessment (NR-IQA) is a fundamental and important task in the field of computer vision. Most NR-IQA methods have limitation in making desirable NR-IQA performance due to the lack of sufficiently rich features. To address this problem, we propose a dual-attention pyramid Transformer network for NR-IQA. In the proposed method, a feature extraction module is firstly used to extract multi-scale features which contain rich distortion and semantic information. Then, a pyramid Transformer network with channel and spatial attentions is designed to learn multi-scale global features from spatial and channel aspects. The combination of pyramid structure and dual attentions enables our network to focus on features in different regions of the image and learn richer and more comprehensive global features. This in turn improves the quality score prediction performance. Finally, the score prediction module predicts the quality scores in different stages of the pyramid Transformer network by channel adaptive prediction branches and determines the final quality score by aggregating these quality scores. Extensive experiments performed on four widely used public databases show that our proposed method is superior to some state-of-the-art NR-IQA methods in perceiving image quality.