RSC-WSRGAN super-resolution reconstruction based on improved generative adversarial network.

Traditional generative adversarial network models have made significant progress in generating high-quality images. However, there are still problems such as smooth edges of reconstructed images, distortion of details, and color shifts in generated images. In this regard, this paper proposes a new improved generative adversarial network image super-resolution reconstruction RSC-WSRGAN model. This model redesigns the residual block in the super-resolution using a generative adversarial network generator network and removes the batch normalization (BN) layer in the residual block, introduce RFAConv, replace the original rectified linear unit activation function with the smooth maximum unit activation function, and add the convolution attention module CBAM to build the RSC module. The model's focus on features is enhanced, improving the detail and clarity of the reconstructed image. In addition, Wasserstein distance is used to replace the JS divergence that measures the data distribution in the generative adversarial network to optimize the network training process. It also solves the problem of gradient disappearance due to the removal of the BN layer, making the network training process more stable. At the same time, the loss function is improved to better quantify the error between the reconstructed image and the real image, the network model is optimized, and the quality of the generated image is improved. Finally, the improved RSC-WSRGAN image super-resolution reconstruction model was used to conduct reconstruction experiments on the Div2k data set. The results showed that compared with the original model value, the PSNR was improved by 0.534 dB, the SSIM value was improved by 0.038, and the LPIPS value was optimized by 0.018. Image reconstruction experiments were conducted with other mainstream models on public general data sets such as Set5, Set14, and BSD100. The results show that the model proposed in this article further strengthens the edge contour of the reconstructed image, and the color does not suffer from distortion or offset. The overall visual quality is better, the look and feel has been significantly improved, and the realism of the reconstructed image has been improved. [ABSTRACT FROM AUTHOR]