Research On Improved Methods And Image Super-Resolution Of Generative Adversarial Networks

Generative adversarial networks(GAN)has been widely studied and gotten great attention by many scholars,soon after its emergence in 2014.As more deep learning network technologies and structures are proposed,the generative adversarial networks has also achieved many achievements in theories and applications.Despite these,key problems such as convergence condition,unstable training and unsatisfied image quality in GAN are still developmental issues and full of challenges,and how to better apply the adversarial training method to pixel-level image processing tasks,such as image super-resolution,is also a problem worthy of further study.In this thesis,key issues of GAN such as gradient disappearance,mode collapse and its application in image super-resolution are studied.Several improved methods for both the hybrid model and the standard model are proposed,which improve training stability and generate high quality images.By introducing self-attention,dense connection and adversarial training strategies,the proposed super-resolution model generates super-resolution images with global consistency and texture sharpness.The research results of this thesis are summarized as follows.For the hybrid method of boundary equilibrium generative adversarial networks(BEGAN),a simple and effective method is proposed to reduce noise-like areas and improve the diversity and smoothness of the generated images.Augmenting the discriminator’s training criterion with a denoising loss and adding the batch normalization to the network architecture,the generated images are smoother and with richer characteristics.The effects of different techniques on the BEGAN framework are evaluated,and the generalization ability of the generator is proved by the space consistency experiments.For the standard GAN,two simple but effective and easy-to-implement methods,DRRGAN and GAN-RL,are proposed to improve the key problems of training instability,generator gradient disappearance and mode missing.Taking the feature output of the discriminator as the input of the generator to denoise the real data and the reconstruction of the latent code,DRRGAN allows the entire network to learn useful information about the real data distribution,thus improving the performance of GAN and producing higher quality images.Using the features learned from the discriminator to reconstruct real data through the generator,GAN-RL encourages the discriminator to learn more informative features and the generator to produce all modes of real data.GAN-RL can provide stable gradient signals for both the generator and the discriminator,is robustness to hyperparameter changing and can be easily combined with the other methods to improve their performance.In order to produce realistic and sharpness super-resolution images,adversarial training strategy is introduced to overcome the shortages of over-smoothness and blurry texture caused by pixel loss alone.To extract fully features from low-resolution images and reduce the dependence of receptive field on depth,the proposed DSSA_GAN adopts self-attention and dense connection between residual sets for its generator.The generator is pre-trained by pixel loss and then fine-tuned by adversarial train,which encourages the discriminator to concentrate on texture differences,resulting in more realistic super-resolution images.All the methods proposed in this thesis are experimentally verified on different datasets.The experimental results show that the improved method for the BEGAN improves the training stability,and the improved methods for standard GAN,DRRGAN and GAN-RL,have the advantages of fast training,high efficiency and less computational load.The effectiveness of the proposed super-resolution model DSSA_GAN is verified by evaluating the quality and quantity on several benchmark datasets.The research results of this thesis provide new ideas and insights for improving the performance and theoretical research of GAN,and they also provide instruction for image super-resolution based on adversarial training.