查看文章

Recent, the single image super-resolution (SISR) methods are primarily based on building more profound and more complex convolutional neural networks (CNN), which leads to colossal computation overhead. At the same time, some people introduce Transformer to low-level visual tasks, which achieves high performance but also with a high computational cost. To address this problem, we propose an attention-based feature fusion super-resolution network (AFFSRN) to alleviate the network complexity and achieve higher performance. The detail capture capability of CNN makes its global modeling capability weak, we propose the Swin Transformer block (STB) instead of convolution operation for global feature modeling. Based on STB, we further propose the self-attention feature distillation block (SFDB) for efficient feature extraction. Furthermore, to increase the depth of the network with a small computational …