Study And Application On The Structure Improved Deep Convolutional Generative Adversarial Networks

In recent years,deep learning has achieved rapid development.Generative Adversarial Network(GAN)is a popular generation model in deep learning.Different from the traditional generation model,in addition to the generation network,GAN includes a discriminant network which can distinguish samples in the model structure.There is a confrontational relationship between the generative network and the discriminative network.GAN has been widely studied and applied in the field of image and vision.In addition to the field of image generation,it is also a practical and effective method to use model of GAN for image recognition to replace traditional data enhancement methods.However,the problem of unstable training and difficulty in convergence during the test are important factors affecting the development of GAN.Therefore,it is of great significance to study ways to improve the GAN network structure and the usage of GAN for image recognition.By analyzing research results on deep learning and GAN development at home and abroad,GAN basic ideas,GAN structural composition,and training methods of GAN,in view of such problems as inadequate feature extraction of original GAN,uncontrollable sample generation,mode collapse and mode loss in the training process,the following main work is carried out in this thesis:Deep Convolutional GAN(DCGAN)integrates Convolutional Neural Networks(CNN)for unsupervised training based on traditional GAN.GAN can be extended to Conditional Conditional GAN(CGAN)by adding conditions.Combining the advantages of DCGAN and CGAN,Conditional-DCGAN(C-DCGAN)is established.The stronger feature extraction ability of convolutional neural network is utilized to generate samples with conditional assistance,and the structure is then optimized and improved for image recognition.The experimental results show that the method can effectively improve the recognition accuracy of images.This thesis proposes a new method that takes advantage of the spectral normalization and global normalization in an improved C-DCGAN.The global weight obtained after spectral normalization adds Lipschitz constraints to the model discriminator.Group normalization normalizes input to the hidden layer,speeding up training and improving the quality of the generated samples.Using the global average pooling layer instead of the fully connected layer to avoid overfitting problems.This improved method is applied to image generation and image recognition of two data sets,CIFAR-10 and SVHN respectively.The experimental results show that,compared with the existing methods,this method can not only generate better images,but also improve the accuracy of image recognition.Finally,based on the two improved algorithms C-DCGAN and SICS-GAN,this paper conducts the algorithms on the MNIST dataset and deploys the algorithms on the Web.Through the handwriting panel of the front-end page,the online handwritten digit recognition system is realized.