Efficient Multi-Scale Cosine Attention Transformer for Image Super-Resolution
Yuzhen Chen, Gencheng Wang, Rong Chen
Abstract
Transformer has gained increasing attention in the field of computer vision due to its global modeling capability. However, its high computational cost and weak capture of local pixel information restrict its application in realistic scenes. To address these issues, we propose a multi-scale cosine attention Transformer network (MCATN) for single image super-resolution. Specifically, we propose a residual multi-scale Transformer group (RMTG), which consists of a local feature extraction module and a multi-scale Transformer block (MTB). RMTG is designed to capture the local information of features and model long-range dependencies. Within the MTB, the multi-scale cosine attention layer divides the input features into groups and further splits them into blocks of different sizes to enable fast cosine self-attention computation. MTB mitigates information loss during image reconstruction on feature block boundaries and prevents the network from being dominated by specific pixels. Extensive experimental results demonstrate that MCATN outperforms several state-of-the-art lightweight models in terms of accuracy.