RDNet: Lightweight Residual and Detail self-attention Network for infrared image super-resolution

Recently, Convolutional Neural Network (CNN) and Transformer have shown great potential for infrared image Super-Resolution (SR). However, most CNN-based SR methods pursue large receptive fields by stacking massive convolutional layers, which results in excessive number of parameters and high computational complexity. On the other hand, most Transformer-based SR methods utilize symmetric window-based self-attention to model long-range dependencies, which tends to introduce sub-similar image patches into the group of similar image patches, thus hindering accurate detail reconstruction. To address the above issues, we propose a lightweight Residual and Detail self-attention Network (RDNet) for infrared image super-resolution. Specifically, we design an Efficient Residual Self-Attention (ERSA) module, which guides feature weighting for the deep self-attention module with the shallow self-attention map to accurately extract favorable global features, thus strengthening the details of infrared images substantially. Then, we devise an Efficient Detail Self-Attention (EDSA) module, which exploits an asymmetric window-based self-attention to alleviate the interference of sub-similar image patches, thereby facilitating accurate detail reconstruction. Extensive experiments validate that the proposed RDNet outperforms state-of-the-art methods in both objective evaluation metrics as well as subjective visual perception with fewer parameters. •Introduces RDNet, an innovative lightweight super-resolution method for infrared images.•Designs ERSA that stimulates the proposed RDNet to highlight important global features.•Presents EDSA with asymmetric windows that motivates our RDNet to strengthen infrared image details.•Demonstrates the superiority of the proposed RDNet compared with state-of-the-art methods.