Mixed Network Based on Feature Modulation and Parameter-Free Attention for Efficient Image Super-Resolution

In recent years, single image super-resolution (SISR) based on deep learning algorithms has received more attention from the research community. There are many ways to solve this task. The SR method based on convolutional neural network (CNN) is mainly effective in restoring high-frequency details, but often ignores the key low-frequency contour information. Transformer-based SR methods, although including global structural details, usually come with a large number of parameters, resulting in high computational overhead. In this paper, we propose a network that combines convolutional feature modulation with parameter-free attention to achieve lightweight image super-resolution. First, a multi-receptive field hybrid convolution module (MRHC) is used, which can capture multi-scale features and expand the receptive field. Secondly, we use feature shift mixer (FSM) to fuse features, and introduce a parameter-free attention mechanism. We use the origin symmetric activation function to directly generate weights, which greatly reduces the amount of calculation and parameters. Finally, we use convolutional channel mixer (CCM). The CCM is used to interact with channel information and achieve overall fusion of spatial features and channel information. Experimental results show that this lightweight model achieves good results for SR tasks.