Research And Applications Of Generative Adversarial Network Based On Complete Representation Learning For Multi-view Face Image Generation

As an indispensable part of face analysis and recognition,face image has attracted much attention.Generating multi-view face image samples with high precesion and fidelity can expand face data set to a certain extent,and alleviate the problem of training data demand in model training.In recent years,the emergence of generative adversarial networks(GAN)has significantly improved the authenticity and diversity of face image generation,and achieved good results in multi-view face generation tasks.However,the research shows that in the process of generating face image by random sampling from low-dimensional hidden space,the result of sampling the mapping space which is not covered by the training data set is not satisfactory.That is to say,the hidden space representation of the data learned by the model is incomplete.On the basis of learning from the complete space representation,this paper proposes an improved generative adversarial network based on the complete space representation(ICRGAN)and ICR-GAN with spectral normalization(ICR-GAN-SN).The above model adopts the dual channel structure of learning the complete hidden space representation of real data to ensure that the learning of hidden space is complete and not limited to the training data set.Based on the above model,a multi-view face image generation system is designed and implemented,which is applied to face matching.The specific research work of this paper is as follows:Firstly,the Improved generative adversarial network based on the complete space representation is proposed.In this model,the dual channel structure of CR-GAN model is excavated to give full play to the guiding role of different channels to the generator composed of encoder E-decoder G in improving the image generation quality.In the aspect of reconstruction of multi-view face images of real samples,ICR-GAN adds constraints to ensure that the generated image is the real reconstruction of the input samples,and guides the generator on the reconstruction path to generate multi-view face image with better retention of the object identity;Similar discriminant score constraints are added to the objective function of discriminator D on the generation path to identify the same object samples,so as to ensure that discriminator D fully considers face feature analysis when discriminating authenticity.The experimental results show that the ICR-GAN improves the discriminator's discriminant score for the same object,and the model achieves better results when generating random images and multi-view face images than CR-GAN.Secondly,the ICR-GAN based on Spectral Normalization(ICR-GAN-SN)is proposed.This model aims at the problem that the quality fluctuation of the image generated by the generator is not stable during the model training of CR-GAN and ICR-GAN.It analyzes the optimization scheme that satisfies the 1-lipschitz condition in the objective function of the discriminator.Instead of using the original Gradient Penalty,it applies the Spectral Normalization to both discriminators D on the two channels,which limits the changes of neural network layer parameters in discriminator D.The experimental results show that in ICR-GAN-SN,the fluctuation amplitude of the objective function of discriminator D and encoder G decreases,the convergence speed of the two networks is obviously accelerated,the training process of the model is more stable,and the generation effect of multi-view face for a single face image is improved.Finally,multi-view face image generation system is designed and implemented.The system mainly includes training modules of different models,multi-view face generation module and face matching module.The system can select different models to generate different multi-view face images by inputting a single-view face image.According to the different input data sets,the model parameters can be retrained.The face matching module is set up in the system.The face image generated by the model is compared with the face database,and good matching results are obtained.