Lightweight image super-resolution based on stepwise feedback mechanism and multi-feature maps fusion

In recent years, deep learning has made remarkable breakthroughs in single-image super-resolution (SISR). However, the improvements often come with the increased network size, which is impractical for resource-constrained mobile devices. To alleviate this problem, an SISR method based on stepwise feedback training and multi-feature maps fusion (SFTMFM) is proposed in this paper, with fewer parameters amidst improved performance. Specifically, to better balance the performance and model parameters, a symmetrical CNN (SCNN) based on parameter sharing is constructed. In addition, to make up the deficiency of CNN module, the Swin Transformer layer (STL) is adopted to extract similar features over long distances. Lastly, to further improve the reconstruction ability of the model, a stepwise feedback training strategy is designed, which combines the cross-feature maps attention module as a feedback mechanism with the multi-feature maps fusion module to gradually reconstruct the model with higher-quality images. Under × 2 upscaling, our method achieves the PSNR(dB) of 38.10, 33.69, 32.25, 32.33, and 39.00 for SET5, SET14, BSD100, Urban100, and Managa109 datasets. Compared with the state-of-the-art lightweight SISR methods, our method shows better reconstruction performance and less computational cost.