Research On Image Data Generation Technology Based On Generative Adversarial Network

Deep learning is research based on large-scale data.In the real world,it is difficult to obtain and label the tens of millions of data required for deep learning training,and human and financial resources are huge.Therefore,how to solve the lack of network training sample data is the main problem facing deep learning.The advent of generative adversarial networks has greatly promoted the research to solve the lack of sample data.However,when training the adversarial network,the model is prone to crash,the training process is unstable,and it is difficult to converge.All of these can result in poor quality images,inability to generate image samples,or inability to generate individual images.Therefore,in order to maintain the relevant performance of the neural network and avoid overfitting the network,a large amount of data needs to be obtained.In this paper,the generation technology of adversarial network in image data generation is studied.The main work and research results are as follows:(1)Aiming at the problems of unstable training and difficult convergence for generative adversarial networks,this paper analyzes the reasons for the difficulty in training generative adversarial networks and proposes a generative adversarial network based on spectrum constraints.This method introduces spectral norm normalization.Layering,limiting the gradient to the 1-Lipschtiz condition,slows down the network convergence speed and improves the stability of the model.(2)Aiming at the problem that the quality of the generated image is not high,a generative adversarial network based on residual structure is proposed.This method introduces the residual structure to deepen the network depth of the generator model and the discriminator model,and improves the generator model and the discriminator.The ability of the model to obtain deep features enhances the network's ability to express,thereby improving the quality of the images generated by the network.(3)Aiming at the validity of the generated image data for dataset expansion,a small-scale dataset expansion experiment was performed under the same experimental platform and environment.After expanding by different multiples,this paper analyzes the recognition accuracy of the original dataset on the classification network.Experiments show that generating image data can improve the convergence of the network,improve the recognition accuracy of the network,and reliably and effectively alleviate the problem of insufficient image data.In summary,this paper has conducted in-depth research and discussion on the generation of adversarial networks in image data generation,and has proposed an improvement method for the problems of unstable training,easy to collapse and low quality of generated images,which is proved by experiments.The effectiveness of the method and the generated data significantly improve the ability to identify the network.