Research And Implementation Of Generative Adversarial Network

The Generative Adversarial Networks(GAN)was first proposed by a group of researchers led by Ian Goodfellow in 2014.The antagonistic generation network fits the distribution by constructing two competive models.The generation model takes noise as input and uses the neural network to generate samples.The discriminant model simultaneously samples from the generator model and the real training datasets,and uses the gradient descent to optimize the model parameters.Finally generate similar real sample can be sampled from the generator.Based on the original Generative Adversarial Networks theory and the deep convolutional neural network,this thesis studies and solves the problems existing in the complex computer vision scenes of the Generative Adversarial Networks application,and designs a variety of improvements.A deep convolutional neural network model applied to complex image content.The main work of this thesis has the following parts:1)To analyze the difficulty of convergence of the original Generative Adversarial Networks and the causes of pattern collapse during training.According to the analysis results,this thesis proposes three improvements,including 1)The derivative transformation of the discriminator input,making it more difficult for the discriminator to find the hyperplane between the implicit distribution represented by the generator and the real distribution of real dataset.To stabilize the training process and solve the convergence problem in complex scenes;2)Introduce the experience playback queue,make the data random in timing while training the discriminator,and maintain the sampling data independently and identically to reduce model collapse;3)Adding additional coding information in the generator and using the discriminator as a decoder,so that the image produced by the generator has more adjustable changes.2)To study the application of the Generative Adversarial Networks in image superresolution.By describing the image information repair as the conditional probability expectation maximization under a given sample,this thesis designed an improved Generative Adversarial Networks model.At the same time,the superresolution task is decomposed into two parts: image denoising and detail reduction.It can realize end-to-end model block training and two-part smooth fusion in the model.In the evaluation,this thesis also designed a baseline evaluation network to evaluate the results,and the model achieved good results in both Google Street View and the natural photography data set.3)To study the application of the Generative Adversarial Networks in the image domain migration problem(or image translation problem).This thesis combines the improved Generative Adversarial Networks and realizes training with unpaired images by constructing the loop constraint loss function.In the experiment,the performance of the network under different conditions on the CELEBA data set was studied,and the conversion results were analyzed by the gender determination network.The results show that the model can be effective in most cases.