An Image Super-Resolution Network Using Multiple Attention Mechanisms

Single image super-resolution (SR) aims to reconstruct high-resolution (HR) images from low-resolution (LR) images. Existing SR algorithms often lose information when dealing with complex textures and details, and the model's feature weighting in different regions is unreasonable, leading to poor reconstruction effects. Recently, networks based on attention mechanisms have shown excellent performance. Attention mechanisms can enhance the model's use of critical input information and reduce the emphasis on non-critical information. In this study, we improve the traditional transformer framework by incorporating channel and spatial attention mechanisms in the deep feature extraction stage to capture the channel and spatial relationships of each feature map, enhancing the model's high-dimensional information extraction capability. We also utilize a pixel attention mechanism to improve the up-sampling module, allowing the model to retain more detail during up-sampling. The validation on benchmark datasets demonstrates that our method outperforms other models.