Research On Equivalent Model And Applications Of Generative Adversarial Network

Since it was proposed in 2014,the Generative Adversarial Network(GAN) has gained extensive attention in many fields and has become one of the hottest research fields of artificial intelligence.GAN's thought is mainly inspired by the zero-sum game in game theory,and the model training is realized by the confrontation between the generator and the discriminator.At present,the most successful application fields of GAN are computer vision,such as image generation,image translation,image super resolution,etc.,and it has also made breakthroughs in natural language processing and other fields.However,GAN is still in the stage of development,and at present,there are still some problems in GAN,such as unstable training and mode collapse.Therefore,it is of great significance for research of the existing problems and applicable fields of GAN.In this paper,an equivalent model of original GAN is proposed based on the analysis of the problems of original GAN in the training process,and applied to insulator fault recognition.Based on CycleGAN,a CycleGAN model for image defogging is proposed.The research contents of this paper are summarized as follows:Based on the basic principle and training mechanism of the original GAN,the original GAN training process is analyzed by using the experimental results of the original GAN;Then based on the theory of signal detection and estimation,the equivalent model is derived for the discriminator and generator of the original GAN,and obtaining the conclusion that the derivation result of generator is consistent with the Wiener-Hopf integral equation.Finally,experiments were carried out on the MNIST and CIFAR-10 experimental databases to verify the effectiveness and advantages of the equivalent model.Based on the original GAN equivalent model proposed in this paper,for the problem that the small sample library similar to the insulator is easy to over-fit the neural network model,the original GAN equivalent model is used to amplify the insulator sample library,so as to improve the correct recognition rate of defective insulators by neural networks.Finally,the effectiveness of this method is verified by comparing with other methods.Based on the basic principle of CycleGAN,the CycleGAN model for image defogging was constructed.By improving the structure of CycleGAN and introducing the perceptive loss function into the loss function of the model,the fogging processing of the image with fog is realized.And by comparing with the experimental results of several typical defogging algorithms,the superiority of the CycleGAN defogging model proposed in this paper is proved.