Srgan-Sa: Super-Resolution Generative Adversarial Network with Integrated Attention Mechanisms

Abstract

With the development of the times, this paper proposes a dynamic spatial attention enhancement SRGAN-SA model to better solve the above problems, such as the loss of high-frequency details in traditional image super-resolution reconstruction (SISR) methods, and the insufficient attention of SRGAN global feature processing mechanism to key areas. Based on the deep learning module, the model designs a learnable 7 * 7 convolution to generate dynamic attention maps. It’s embedded in the residual module of SRResNet generator to realize adaptive focus and feature reweighting of regions. The model adopts the confrontation training paradigm, combines VGG perception loss and confrontation loss optimization generator, and uses depth convolution discriminator to enhance texture identification. Experimental results show that compared with SRGAN, the PSNR of this method on Set5 test set is increased by 1.747dB, reaching 31.6381, and the SSIM is increased by 0.0377, reaching 0.8863. The visualization results show that the improved model can effectively suppress artifacts. The ablation experiment further verified that the contribution rate of the dual pool strategy and the dynamic convolution design to the performance improvement reached 0.82 dB and 0.65 dB respectively. It provides a new solution for the deep integration of attention mechanism and confrontation training.