Research On Image Super-resolution Reconstruction Based On Generative Adversarial Network

Image resolution mainly depends on the performance of the hardware equipment for image acquisition.The hardware performance improvement is complex,difficult and costly.Image super-resolution reconstruction algorithm has attracted the attention of image engineering researchers.In recent years,the image super-resolution reconstruction based on Generative Adversarial Network has indeed improved the shortcomings of traditional methods,but the objective evaluation of reconstruction results has not been improved,even lower than the traditional methods;due to the difficulty of the generated countermeasure network training,it is impossible to judge the training process through the corresponding loss value,how to make the training balance between the generator and the discriminator stable,and how to improve the reconstruction goal is the current need Problems solved.In this paper,the super-resolution reconstruction under the generation countermeasure network is the research goal.1.Analyze the method of super-resolution reconstruction in theory,focus on the realization of generating countermeasure network in super-resolution reconstruction.In view of the current problems of generating countermeasure network,further study its network structure,loss function,etc.,put forward the improvement method,set up a variety of methods for experimental comparison and analysis;2.Based on the deep convolution generation countermeasure network,an improved method is proposed,which combines the residual dense block,improves the network structure,further strengthens the advantages of convolution neural network in extracting image features,multi-layer fusion of information flow between deep networks,realizes the layered feature multiplexing,and recovers the high-resolution image by 4 times of transposition convolution amplification.The simulation results show that the PSNR and SSIM of the reconstructed image are improved by 0.91 db and 0.043 respectively on the basis of srgan;3.Combined with the Wasserstein distance proposed in Wasserstein general advanced network,an improved reconstruction method is proposed by introducing it into srgan model.The training process can be judged by the unique loss,and add the optimized residual block,and replace partial activation function with Selu.This method can improve the performance of network training,accelerate convergence.The simulation results show that the PSNR and SSIM of the reconstructed image are improved by 1.414 db and 0.055 respectively on the basis of srgan.