Research On Face Frontalization Based On Generative Adversarial Networks

Face recognition is a common identification technology in life.Due to the influence of pose and illumination,the performance of the existing face recognition models wil]be seriously degraded in the unrestricted environment(such as the surveillance video).The pose is the main factor leading to the degradation of model performance.Face frontalization method can synthesize an identity-preserved frontal-view face image from a face image in a specific pose.It can be used as a preprocessing process in existing face recognition models to alleviate the influence of attitude change on model performance.The generative adversarial networks have a strong advantage in the task of face frontalization.The research content of this thesis is the face frontalization algorithm based on generating adversarial networks.In order to solve some problems of the existing advanced face frontalization algorithms and improve the face frontalization effect,so as to lay a foundation for further improving the accuracy of face recognition,this thesis proposes three face frontalization models:(1)A face frontalization generative adversarial network based on multi-attention mechanism(MA-GAN)is proposed.Due to the complex network structure,the training process of some existing face frontalization algorithms based on GAN is difficult and inefficient.And some algorithms require inputs of other prior knowledge of the face.It is always difficult and time-consuming to acquire this kind of data.This model integrates multi-attention mechanisms to solve the above problems while maintaining a good face frontalization effect.The attention mechanism introduced in the model is divided into three aspects.First,dual-attention module is added to the generator,so that the generator can focus on facial features and retain more identity information;Second,the face attention mechanism is constructed by four independent discriminators to emphasize the feature generation of the most discriminant region in the face.Third,a self-attention module is added to the network structure of the discriminator.It can enhance the discriminator’s ability to guide the generator.At the same time,a variety of loss functions are used during training to make the synthesized image more realistic.Experiments show the effectiveness of this model.(2)A face frontalization residual generative adversarial network based on deep feature encoder(DFE-ResGAN)is proposed.Under the premise of retaining the advantages of MAGAN,this model solves the problems of inaccurate outlines of the synthesized images.And further improves the generation effect of the model.The residual block can solve the degradation problem caused by the deepening of network layers.Based on this advantage,a residual generating adversarial network is proposed.By deepening the network layers,the model has a stronger fitting ability.At the same time,a deep feature encoder is proposed to add to the generator,which can enable the network to extract more abstract and relevant intermediate feature.Experimental results show that this model has better performance.(3)A face frontalization generative adversarial network based on cascade two generators(CTE-GAN)is proposed.The generation network of this model is composed of two generators in series.The two generators are set to different depths.They can effectively combine the advantages of shallow feature rich in detail information and deep feature rich in abstract information.Thus,this model solves the problem of DFE-ResGAN’s insufficient use of shallow feature information and further improves the face frontalization effect of the model.In addition,the model’s overall structure is relatively simple and doesn’t require inputs of other prior knowledge.Experiments prove the superiority of this model.