Research On Image Super-resolution Reconstruction Based On RDB-WGAN

With the wide application of deep learning in image field,the methods on image super-resolution reconstruction have gradually been transformed from traditional algorithms to algorithms based on deep learning by some scholars.The purpose of image super-resolution reconstruction is to make the fuzzy image clear through some techniques and algorithms.The model named SRGAN that using Generative Adversarial Network(GAN)on super-resolution reconstruction of images is studied in this paper,which can reconstruct more realistic images through the learning method of confrontation,To address the problem of instability of the GAN when training and to improve the quality of reconstructed images further,some improved methods are proposed in this paper,and the superiority of the method is verified by some experiments.The research work and innovations of this paper are mainly reflected in the following aspects:(1)The basic working principles of the popular super-resolution reconstruction methods which are widely used at present are introduced,and the advantages and disadvantages among these algorithms are analyzed through some comparative experiments.The experiments show that the super-resolution reconstruction method based on GAN can generate images with clearer details and better visual effects than other methods.(2)On the one hand,removing all Batch Normalization layers of generative network to improve the efficiency of learning is proposed in this paper;on the other hand,combined with the idea of the dense network and residual network,residual dense block(RDB)is proposed in this paper instead of the original residual block to extract the features of multiple levels on low-resolution images,the method can improve the quality and perception of the reconstructed picture.(3)The traditional generative adversarial network is unstable when training and the samples generated by GAN are lack of diversity,In order to solve these problems of the original GAN,it's proposed to replace the JS divergence with Wasserstein distance to measure the difference between the reconstructed high-resolution image and the real image in this paper,which enhances the robustness of the model effectively.(4)Aiming at the problem that the evaluation index of traditional image superresolution reconstruction is inconsistent with the visual effect observed by eyes,Laplacian gradient is proposed in this paper as a new evaluation index to calculate the sharpness of the image,which improves the reliability of the experiment.The improved model is named RDB-WGAN in this paper.Experimental results show that compared with other existing methods of super-resolution reconstruction,the proposed method has the better visual effect and the higher sharpness on human faces images and medical images.