Research On Generation Adversarial Networks Based On Classification Enhancement

With the improvement of hardware performance and the development of neural network in recent years,the research on Generative Adversarial Networks has entered a peak period.Various ideas and application schemes have injected vitality into GAN.GAN has excellent performance in image application,but it's undeniable that GAN still has many problems that are difficult to solve,especially the problem of mode collapse.The model collapse is manifested in two forms.The first is the lack of complexity in sample generation,and the second is the meaninglessness of sample generation.The problem of mode collapse hinders the further development and application of GAN.In this thesis,an improved model based on classifier is proposed to solve the problem of pattern collapse.The improved model is based on Wasserstein distance to measure distribution distance,and it uses convolutional neural network as network structure.The classifier is introduced to assist the discrimination process.The training process of the improved model is divided into pre-training and mixure-training.The discriminator and generator carry out confrontation training in all stages,while the classifier mainly trains the classification ability of sample images in the pre-training,and imposes punishment on the basis of discrimination loss based on the classification results of samples in the mixure-training.The discriminator loss of the improved model is composed of original GAN loss,gradient penalty,classifier attention penalty and complexity penalty.The degree of attention and complexity depend on the classification results of the generated samples by the classifier.The degree of attention is taken from the difference between the probability peak value of the generated samples in the classification and the threshold value of attention.It can constrain the gradient change direction of the generator and encourage the more similar generation to the real samples.Complexity is taken from the nformation entropy of generated distribution,and it takes the complexity of distribution as feedback to encourage the generator to generate more complicated samples.The performance of the model is improved by using the above two penalty to solve the two problems of mode collapse.We evaluate the performance of the improved model by the convergence rate,the similarity between the generated sample and the real sample,the quality of the generated image and the complexity of the generated distribution.The experimental results on the cifar-10 and CELEBA datasets confirmed our hypothesis that the performance of the improved model was improved at a small time cost in terms of generation quality,sample authenticity,and complexity.In addition,we set four sets of comparative experiments,including gradient penalty parameters,attention threshold,attention penalty parameters and complexity penalty parameters,to obtain the global optimal state of the improved model.