Multi-scale perceptual image super-resolution reconstruction based on coordinate attention mechanism

In this paper, we propose a multi-scale perceptual image super-resolution reconstruction network based on coordinate attention mechanism by modifying the SRGAN algorithm. Firstly, we introduce a multi-scale perceptual block to extract feature information at different scales. It can enable our model to focus on the important feature area and improve the visual perception of the generated image. Secondly, we add spectral norm normalization layer to the discriminator network to constrain the weight of the discriminator during the training, to enhance the stability of the network during training. Finally, we replace the original MSE loss function with the Charbonnier loss, which is more robust to outliers, to optimize and train the proposed algorithm. Furthermore, the regularization loss is added to suppress image noise and artifacts. Experimental results show that the generated images using our proposed algorithm have better visual perception and richer high-frequency details.