Font Size: a A A

Residual Generative Adversarial Network Algorithm Research

Posted on:2022-06-14Degree:MasterType:Thesis
Country:ChinaCandidate:D J TuFull Text:PDF
GTID:2518306575483134Subject:Computer technology
Abstract/Summary:PDF Full Text Request
Generative adversarial network is a currently popular unsupervised generation network,which generates images by game learning of the generative network and the discriminative network.The training of a GAN is a minimal-maximal game process where the goal is for the generative network to fully capture the distribution patterns of the real data to generate samples.However,it is not easy to generate the data distribution of real samples in network learning.At the same time,there are some problems such as unstable training,gradient disappearance,poor image quality and so on.Therefore,a residual generative adversarial network(Re-GAN)is proposed based on the residual network(Res Net)and group normalization(GN).First,the residual module is introduced into the generative network of GAN to prevent the gradient disappearance and enhance the stability of the training.Besides,the residual module optimizes the feature transmission between neural network layers and enhances the diversity and quality of the generated image.Second,Re-GAN adopts the group standardized GN to adapt different batches learning,reduce the difficulty of standardization caused by the lack of training samples and stabilize the training process of the network,when the number of samples is enough,GN can make the calculated results well match the sample distribution.In order to verify the effectiveness of the proposed algorithm Re-GAN,we compare Re-GAN with deep convolutional generative adversarial networks(DCGAN)and Wasserstein-GAN(WGAN)with different batches of samples on three datasets i.e.Cifar10,Cceleb A and LSUN bedroom.Two evaluation criteria,i.e.inception score(IS)and Fr'echet inception distance(FID)are adopted in our experiments.Experimental results indicate that: in the aspect of image generation,Re-GAN generates images of high quality and rich diversity;in the aspect of network training,ReGAN guarantees the training have better compatibility whether the batch is large or small and then make the training process more stable and the gradient disappearance relieved.Figure 31;Table 4;Reference 41...
Keywords/Search Tags:Image generation, deep learning, convolutional neural network (CNN), generative adversarial network(GAN), residual network(ResNet), group normalization(GN)
PDF Full Text Request
Related items