Research On The Construction Of Complex Generative Adversarial Network

In recent years,many breakthroughs have been made in the discriminant model represented by the convolution neural network(CNN).Recently,some researchers have extended the CNN from the real domain to the complex domain,and proposed the complex CNN,which has better performance than the real CNN.On the other hand,the generative model represented by the generative adversarial networks(GAN)has also received the attention of researchers at home and abroad.The GAN is composed of a generator and a discriminator.At present,most researchers use real CNN to construct generators and discriminators.This thesis mainly explores the use of complex CNN as the discriminator of GAN,and continuously optimizes the generator of GAN to alleviate the problem of mismatch between generator and discriminator parameters,so as to construct a complex GAN with better performance.The main contents of the thesis are as follows:(1)Construction of complex generative adversarial network.Based on deep convolutional generative adversarial networks(DCGAN),this thesis constructs the first kind of complex generative adversarial networks;based on conditional generative adversalal nets(CGAN),constructing the II and III types of complex generative adversarial networks.Firstly,we keep the generator in DCGAN unchanged and replace VGGNet in the discriminator with complex deep convolution residual network,which is an intermediate model structure between wide and shallow type and deep and narrow type and uses CReLU as complex activation function.The network obtained by this way is called the type I complex generative adversarial network.Secondly,aiming at the problem that the generator and discriminator mismatch in Type I GAN and the discriminator modeling in DCGAN is too free,This thesis proposes that the residual unit of Inception+Residual structure be used in the construction of CGAN network generator instead of ordinary real CNN.We call it type II complex generative adversarial network.Finally,in order to further improve the quality of image generation,we propose category conditional batch normalization(CBN)based on the idea of conditional batch normalization(CBN)of visual question answering(VQA)system in the field of natural language processing.We use Cat_BN to replace the standard batch normalization(BN)and build the GAN generator based on unit which is composed of Cat_BN and residual module.The discriminator is the same as the type II GAN.The constructed network is called the type III complex generative adversarial network.The validation of the proposed complex GAN is verified by the comparison experiments with real GAN.(2)Construction of complex generative adversarial network with self-attention mechanism.Aiming at the problem that GAN is difficult to model in some classes of multi-class datasets,the self-attention mechanism is applied to the generator and discriminator of GAN,and a complex GAN with self-attention mechanism is constructed.The experimental results show that it is effective to add a self-attention mechanism in the GAN,and it is better to apply the self-attention layer to the position of the advanced feature map.Subsequently,an experiment was designed to replace the self-attention module with the residual module with the same parameters.The results show that the performance improvement brought by adding the self-attention mechanism in GAN is not only caused by the increase of the depth and capacity of the model.Aiming at the instability of GAN training,two stable training techniques,spectral normalization(SN)and two time-scale update rule(TTUR),are applied to the proposed complex GAN with self-attention mechanism to further improve the performance of image generation.