Lightweight multi-scale global attention enhancement network for image super-resolution

The Transformer-based depth model has achieved impressive results in the field of image super-resolution (SR). However, these algorithms still face a series of complex problems: redundant attention operations lead to low resource utilization, and the sliding window mechanism limits the ability to capture multi-scale feature information. To address these issues, this paper proposes a lightweight multi-scale global attention enhancement network (LMGAE-Net). Specifically, to overcome the window limitations in Transformer models, we introduce a multi-scale global attack block (MGAB), which significantly enhances the model’s ability to capture long-range information by grouping input features and calculating self-attention with varying window sizes. In addition, we propose a multi-group shift fusion block (MSFB), which divides features into equal groups and shifts them in different spatial directions. While maintaining the parameter quantity equivalent to 1×1 convolution, it expands the receptive field, improves the learning and fusion effect of local features, and further enhances the network’s ability to recover image details. Extensive experiments demonstrate that LMGAE-Net outperforms state-of-the-art lightweight SR methods by a large margin.