Research On Face Frontalization Based On Generative Adversarial Network

Face recognition technology has developed rapidly in recent decades,receiving widespread attention.With the burst of deep learning method,many breakthroughs have been made.The face recognition rates based on deep learning have exceeded human eyes.However,most of these studies were under the assumption of frontal face or near frontal face,so there have some limitations.The thesis solves pose-invariant face recognition(PIFR)problem by generating frontal face,and proposes two novel pose normalization methods.The main contents of the thesis are as follows:1.A face frontalization algorithm based on Boundary Equilibrium Generative Adversarial Network(BEGAN)is proposed.In order to solve the problems of training difficulty and mode collapse caused by unreasonable distance measurement in traditional generation adversarial network,the following changes are made.Firstly,the discriminator is designed as an auto-encoder and the probability value is changed into energy-based continuous value,which is the same as BEGAN.In this way,the training process is more stable and efficient,the diversity of generated samples is also increased.Secondly,a feature extraction module(Light-CNN)is added into the network,as the third part of the network structure.It works against the generator together with the discriminator to retain the face identity consistency.Finally,compared with other algorithms,it adopts a simpler and more efficient network structure,avoiding the multi-scale feature extraction operation,which reduces the computational burden greatly.The experimental results on public datasets show that the algorithm can greatly improve the face frontalization effect,which proves that the algorithm is effective and can provide a valid solution for the pose-invariant face recognition problem.2.A wide-pose face frontalization algorithm based on Pose-Aided Generative Adversarial Network(PA-GAN)is proposed.Adding the pose parameters of the rotated face can accelerate the convergence of the network by increasing the constraints.Firstly,in order to get more accurate pose parameters,a pose estimation algorithm using discriminative regression random forests on RGBD images is designed.Secondly,the two-pathway network is changed into single-path structure.The two-pathway structure algorithm has to localize the face landmarks to get the detailed area of the face,which limits its application.Finally,the algorithm uses the idea of BEGAN to design the network structure,avoiding many problems brought by the traditional network structure such as Deep Convolutional Generative Adversarial Network(DCGAN).The experimental results on public dataset show that the algorithm can effectively frontalize the rotated face with a wider range of pose changes.3.A pose robust face recognition system is implemented.The system offers functions of real-time face storage,face detection,landmark extraction,pose normalization,face recognition and so on.It has good interactivity.Moreover,it can display the face recognition results in short time.The system has great significance for further research of pose-invariant face recognition.