Super-Resolution Reconstruction Algorithm Based on Improved ESRGAN

In response to the current image super-resolution reconstruction algorithm still suffers from insufficient detail processing such as texture and artifacts, this paper proposes an image super-resolution algorithm based on an improved enhanced super-resolution generative adversarial network. The Enhanced Super-Resolution Generative Adversarial Networks (ESRGAN) model is improved in terms of both generative network and loss function. First, the shift convolution is introduced in the generative network to effectively extract the image's local structure information while keeping the complexity comparable to that of 1×1 convolution; the multi-scale feature extraction module is introduced in the densely connected module, and then a channel attention mechanism is added after this module to adjust the feature response values of different channels, thus improving the deep information extraction module in the generative network model. Secondly, the loss function of the network is improved by adding a local perceptual loss function to further improve the detail retention capability of the network. Finally, the improved reconstruction module is used to complete the image super-resolution reconstruction. The 2-fold and 4-fold super-resolution reconstructions are tested on datasets such as Set5, and the proposed algorithm is compared with methods such as Bicubic. The results show that the reconstructed images obtained by the proposed algorithm have sharper edges and can provide relatively more details, which greatly improves the visual effect of the images.