Local-global aggregation transformer for enhanced image super-resolution

Abstract

Recent advancements in Transformer-based methods have significantly improved image super-resolution (SR) tasks, outperforming traditional CNN-based approaches. However, most existing Transformer-based methods focus predominantly on global dependencies while neglecting local information, thereby limiting their effectiveness. To address this, we propose the Local-Global Aggregation Transformer (LGAT), which combines convolution-based attention and window-based self-attention to leverage both global statistics and strong local fitting capabilities. Additionally, we introduce the Spatial Frequency Fusion Block (SFFB) to model long-range dependencies and enhance feature expression. Furthermore, we propose a novel Spatial-Gate Multi-Layer Perception (SGMLP) to mine additional non-linear spatial information and reduce redundancy. Extensive experiments on benchmark datasets demonstrate that LGAT achieves impressive performance, outperforming state-of-the-art SR methods both objectively and subjectively. Our contributions include the development of LGAT, which utilizes both local and global features for better reconstruction, the introduction of SGMLP and SFFB, and the demonstration of LGAT's effectiveness through extensive experiments.