Research On Image Generation Algorithm Based On GAN

Image generation occupies an important position in the practical technology of deep learning.Its purpose is to generate as large a detailed image as possible and to generate corresponding images as required to provide more reliable information.In view of the uncontrolled image generation of the image generation network model based on the generation adversarial network(GAN)and the long training time of the model,this paper proposes an improved image generation model based on the generative confrontation network.Taking the conditional generation adversarial network(CGAN)and the deep convolution generation adversarial network(DCGAN)for solving image generation as the benchmark,and the improvement of the benchmark model to make it play a higher performance in the task of generating target type images.The main research contents are as follows:(1)In this paper,CL-CGAN generation model is proposed to solve the problem of uncontrollable image samples.Based on the analysis of CGAN,the loss function,model structure and training method of CGAN are modified.An improved loss function network model is proposed to calculate the cosine similarity between the real image and the generated image.At the same time,the calculation result is combined with the random noise as the input vector to the generator,Control the direction of image generation to make it closer to the real sample.(2)In order to solve the problem that the training process of the model is too long and the image is distorted.In this paper,the HS-DCGAN generation model is proposed.Based on the deep convolution generation adversarial network,the Adam optimizer is used as the new optimizer of the network,and its hidden space is used to add the random noise to the hidden space between the hidden layers.Convolution neural network is used to improve the convergence speed of the model in the training process,and the hidden space is used to improve the image quality generated by the model.Compared with model one,the image generated by model two is clearer,and the training time of model two is less.(3)The MNIST data set,Cifar-10 data set and CelebA data set are used to experimentally verify the image quality generated by the two improved models and the convergence of the models.Experimental results show that the image generation model proposed in this paper can generate clearer black and white and color images,and effectively shorten the model training time.Figure 25 Table 5 Reference 63...