A method for image super-resolution reconstruction based on multi-scale features and attention mechanisms

In the task of image super-resolution reconstruction, this paper proposes an image super-resolution method called MSA-SR, which is based on multi-scale features and attention mechanisms. This method effectively captures the low-frequency and high-frequency features of low-resolution images by separating and extracting multi-scale features in both the time and frequency domains. On this basis, high-frequency guided cross-attention is used to selectively enhance high-frequency features, while wavelet convolution is employed to protectively enhance low-frequency features, achieving clear and natural image super-resolution reconstruction effects. The model was validated on the Urban100 and Manga109 datasets, and its performance metrics of Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) showed certain advantages over other deep learning super-resolution methods. From a quality perception perspective, this model has made significant improvements in texture recovery, color restoration, noise suppression, and naturalness of the image, achieving superior visual effects, which proves the effectiveness and superiority of the model.