Deep Convolutional Generative Adversarial Networks On Face Image Generation

The last decade or so has seen rapid improvements in computer performance and a thriving explosion in artificial intelligence.Machine learning is an important part of this,and it has emerged as the fastest growing area of research,giving computers the ability to simulate human learning.Deep learning allows machines to simulate the human way of thinking and is the technology that enables machine learning.Generative adversarial networks(GAN)is a new deep learning model proposed in 2014,which generates new samples by learning the data distribution of the samples,the new samples will approximate the data distribution of the original samples,GAN has good adaptability to generate data in different dimensions,such as one-dimensional speech,two-dimensional images are of high research value.In this paper,we study the face image generation method of Deep convolutional Generative Adversarial Network(DCGAN),and improve the generation and discriminative models of DCGAN by combining similarity calculation and wavelet function,respectively,to further improve the effect of image generation.The specific research includes the following aspects.(1)A face map segmentation method based on similarity calculation is proposed by combining supervised and unsupervised learning models,thus realizing automatic face image segmentation based on support vector model,and a comparison test is conducted between image segmentation methods based on supervised and unsupervised models and this paper's image segmentation method,and through subjective and objective evaluation results it can be obtained that this paper's method effectively improves the performance and efficiency of image segmentation.(2)The basic principles and framework of GAN and its derivative model DCGAN are studied.In-depth study of the basic principles and components of the generative and discriminative models,so as to realize the training of DCGAN model,on which the basic experimental platform of face image generation based on DCGAN is constructed.(3)The face segmentation image calculated by similarity is added to the discriminant model of DCGAN,thus progressing to control the parameter adjustment of DCGAN.On this basis,the similarity-based DCGAN face map generation approach is presented.As opposed to the raw DCGAN,the method is improved in terms of both the brightness of the generated images and the accuracy of the recognized ones.(4)Inspired by wavelet neural networks,we added wavelet functions to the generators of DCGAN,and the model was adapted in terms of construction and parameters.Based on it,the face image generation method of wavelet DCGAN is proposed.By comparing with the original DCGAN face image generation results,the new way has increased the brightness,contrast and proper identification rate of the face generation maps.