FE-FAIR: Feature-Enhanced Fused Attention for Image Super-Resolution

Abstract

Transformers have performed better than traditional convolutional neural networks (CNNs) for image super-resolution (SR) reconstruction in recent years. Currently, shifted window multi-head self-attention based on the swin transformer is a typical method. Specifically, the multi-head self-attention is used to extract local features in each window, and then a shifted window strategy is used to discover information interaction between different windows. However, this information interaction method needs to be more efficient and include some global feature information, which limits the model’s performance to a certain extent. Furthermore, optimizing the utilization of shallow features, which exhibit significant energy reserves and invaluable low-frequency information, is critical for advancing the efficacy of super-resolution techniques. In order to solve the above issues, we propose the feature-enhanced fused attention (FE-FAIR) method for image super-resolution. Specifically, we design the multi-scale feature extraction module (MSFE) as a shallow feature extraction layer to extract rich low-frequency information from different scales. In addition, we propose the fused attention block (FAB), which introduces channel attention in the form of residual connection based on shifted window self-attention, effectively achieving the fusion of global and local features. Simultaneously, we also discuss other methods to enhance the performance of the FE-FAIR method, such as optimizing the loss function, increasing the window size, and using pre-training strategies. Compared with state-of-the-art SR methods, our proposed method demonstrates better performance. For instance, FE-FAIR outperforms SwinIR by over 0.9 dB when evaluated on the Urban100 (×4) dataset.