Research On Image Generation Method Based On Generative Adversarial Network

Deep learning is based on deep neural networks,seeking and establishing the mapping relationship between data and tasks.It has been highly concerned and widely used in tasks such as computer vision,natural language processing,and intelligent robots.However,although deep learning has achieved great success in many fields,as the depth of the network and the complexity of the network structure increase,in order to maintain the generalization ability of the network and avoid overfitting of the network,deep network learning and training requires a lot of Label data.This means that on the one hand,a large amount of data needs to be obtained,and on the other hand,it takes a lot of manpower and material resources to label the data.This undoubtedly greatly increases the difficulty of deep learning and limits its application to a certain extent.In view of this,this article studies the data generation technology required for deep learning.Generative adversarial networks have strong theoretical advantages and broad development prospects.For random input data,generative adversarial networks have the ability to generate diverse data samples.In deep learning application scenarios,a large amount of labeled data is often required,so the researchers hope that the generation model can generate data samples that meet the conditions and have diversity according to the scene requirements;in addition,the researchers also hope that the data samples generated by the generation model have Higher resolution.In response to these problems,this article did the following:Based on a large amount of literature reading and model test verification,combined with the advantages of the current image generation field model,the generation model structure based on independent learning rules is determined.The main ideas are as follows: first,the fusion of the residual network(Res Net)and the auto encoder(Auto encoder)to form a new GAN model;second,the introduction of independent learning rules,the discriminant model and the generated model are trained at the same ratio,and the two The learner sets different learning rates;third,add jump connections between different layers,and feature information can be directly transmitted across layers to enhance feature transmission.Therefore,a robust neural network model can be trained,and experimental verification on multiple data sets shows that the model can generate smooth and realistic images.Further,in order to be able to generate higher resolution images,reduce the entanglement between features and improve the applicability of the model in real scenes,a dynamic generation model based on the self-attention mechanism andpattern generator was determined.The main ideas are as follows: first,add a self-attention module to the discriminant model to obtain global image information,and generate a mapping network and style module in the model to control the style of the generated image at different resolutions;second,make a new data set based on the home scene,And verified the function of the model to generate images on this data set,which is reproducible on the data set obtained in the real scene.The experimental verification of the model on multiple classic data sets and non-classical data sets(made home scene data sets)shows that the model has better interpolation properties to achieve unsupervised generation of ultra-high resolution images.Similarly,experiments based on non-classical data sets show that the model has the ability to generate samples infinitely,and a small number of training samples can also generate a large number of diverse and realistic images during the training process.