Research On Image Generation Method Based On Deep Learning

The generated adversarial network consists of a generation model and a discriminant model.The confrontation training is its greatest advantage and characteristic.It has been deeply researched and highly concerned by the artificial intelligence academia and industry since its introduction,and has become a hot research direction in the field of deep learning.Generating adversarial network promotes the research and application of image generation and repair,text generation,video prediction and style transfer,and also proves its great potential in image processing.However,in practical application,the generated adversarial network has such problems as unstable training,slow convergence speed,mode collapse,uncontrollable sample generation and lack of quantitative evaluation criteria,so that the generated samples cannot achieve satisfactory results.By improving the model structure,activation function and loss function,the quality of generated samples can be improved effectively.Based on the Pytorch deep learning model,the structure of the deep convolution generated adversarial network is improved,and the convolution operation and the maximum pool operation are alternately used in the discriminator.During training,constraint conditions are added into discriminator and generator to enhance network controllability,make data generation process purposeful and assist sample generation.Scaling exponential linear units(SELU)was used as the activation function instead of Relu and Leaky Relu functions which were commonly used at present.Using the least squares loss function instead of the cross entropy loss function,a conditional deep convolution generation adversarial network(ls-c-dcgan)based on the least squares loss function is proposed.The experimental results on CelebA data set show that the model structure designed in this paper is reasonable and the generated pictures are realistic.