Research On Generative Adversarial Networks Based On Attention Mechanism And Its Application

Generative Adversarial Networks(GAN)are a popular topic of research in machine learning,enabling mapping between data distributions by generating output information based on the features of the input information,and having the same data distribution as the input information.GAN has been widely used for tasks such as image-to-image translation,image synthesis,etc.The traditional generative adversarial networks often have problems such as high computational cost,unstable training process,too many convolutional layers as well as total parameters,and insufficient feature extraction of the input information,etc.To solve this problem,this thesis uses the attention mechanism to build a more effective GAN class model and apply it to the image-to-image translation task,the main work and innovation of this thesis are as follows:(1)A generative adversarial network,named AS-GAN,is proposed with attention mechanism in this thesis.Firstly,the network introduces the Adapting Belief(AdaBelief)optimizer and Spectral Normalization(SN)to stabilize the training process;secondly,the Global Context Block(GC Block)attention mechanism is used in the generator to improve its ability to retain valid information and suppress noisy information;finally,the generator uses Sub-Pixel Convolution(SPC)for upsampling to suppress artifacts in the generated samples.The model has better translation results compared to other similar models by qualitative and quantitative analysis on open datasets for image-to-image translation tasks.The model was applied to the translation task for both standard and satellite maps,and better results were achieved.(2)A generative adversarial network,named RGC-GAN,is proposed with RGC in this thesis.Firstly,A residual global context attention(RGC)mechanism based on GC Block is designed to capture the long-range dependencies;secondly,introduce the RGC into the generator to improve the detection of input information features;thirdly,the generator uses only three residual processes to decrease computational costs and reduce the total number of parameters;finally,a discriminator network is constructed based on residuals to reduce the number of discriminator parameters and to avoid the problem of accuracy degradation due to network depth increasing.The theoretical analysis shows that RGC-GAN reduces the computational cost and the total number of parameters compared to Cycle-Consistent Generative Adversarial Networks(CycleGAN).The network has better translation results compared to other similar networks by qualitative and quantitative analysis on open datasets for image-to-image translation tasks.The network is applied to the translation task of selfie images and anime images,and better results were achieved.