Image Generation Oriented Deep Convolution Generation Adversarial Network

In recent years,research in the field of artificial intelligence,especially in machine learning,has made great progress.Thanks to the improvement of computing power,the popularity of information tools and the accumulation of data,the urgency and feasibility of artificial intelligence research have been greatly improved.In this wave of artificial intelligence,the machine learning method represented by statistical machine learning and deep learning is one of the main research directions.Prior to the introduction of Generative adversarial networks(GAN),the generation of model-based models from the perspective of machine-understanding data was mostly trained using Markov chains,which was less efficient and greatly limited the possibility of practical application of the model.With the introduction of GAN,as an important part of deep learning,it is bringing revolutionary changes in the field of computer vision.GAN can extract features from images well and restore them,so GAN is very suitable for image generation.Not only that,GAN has a good performance in many fields such as language processing and chess competition,and has great application prospects.GAN realizes the process of understanding from model to creation,and its birth means an advancement in the field of artificial intelligence.Nowadays,the generation of confrontation networks has spawned many different branches,such as deep convolution generative adversarial networks(DCGAN),Wassertein generative adversarial networks(WGAN),and boundary balance generation.Different models such as Boundary equilibrium generative adversarial networks.Among them,DCGAN can be regarded as a combination of convolutional neural network and generative confrontation network.It has many advantages of convolutional neural network,such as local perception and weight sharing,which greatly reduces the parameter quantity and has strong robustness and fault tolerance.ability.Other network models are basically evolved from DCGAN,so DCGAN has high research value.In this paper,the deep convolutional generation anti-network is used as the entry point to carry out research.Firstly,the development history of GAN and the current research progress at home and abroad are expounded.Next,the discriminator and generation model structure of the deep convolution generator against the network are analyzed in detail,including the convolutional layer,the transposed convolutional layer and the fully connected layer.At the same time,the learning mechanism of DCGAN and its advantages and disadvantages are analyzed in detail.In order to study the application of DCGAN in the field of image generation,this paper designs and implements a 5-layer structure deep convolutional generation confrontation network based on tensorflow,and uses CelebA to expose the portrait dataset to train it.Finally,the model is obtained and used.Image generation.Through the analysis of the experimental results,it is found that the image generated by DCGAN is not ideal.In order to increase the direct fitting degree between the generated image and the real image,this paper proposes a combination of convolution kernels of size 3 and 5,stacking the feature maps to form a new feature map,and extracting the characteristics in the original image through different receptive fields.Information to increase the diversity of features extracted or restored by the model,and to add structural similarity index(ssim)to the loss function as a new constraint of the loss function to enhance the generated image from three aspects:image composition,brightness and contrast.the quality of.Finally,a verification experiment was designed to compare the image quality generated by the original DCGAN and the improved DCGAN.The judgment was made through a residual network.The feasibility and effectiveness of these improvements were verified,and the future work was carried out.Summary and outlook.