基于变分自编码器的人脸正面化产生式模型

The Generative Adversarial Network Based on Variational Auto-encoder for Face Frontalization

在人脸识别中,人脸姿态差异处理始终是一个难点。人脸正面化中大体可以分为两类:一是通过图形学方法生成正脸,二是通过产生式模型生成正脸。然而,第一类方法生成的正脸畸变会比较大,第二类方法普遍采用GAN网络,但是训练过程不易收敛。为解决以上问题,设计基于变分自编码器的产生式模型,采用β-VAE模型,学习隐空间与真实图片空间关系,提高图片生成质量;使用模拟+无监督(S+U)学习方法,提高模型在训练过程的稳定性。通过在Multi-PIE数据集上进行实验,模型在定量打分中取得了不错效果,能够生成更加真实的图片。实验证明,该模型能够有效解决传统方法中生成图片不真实和训练不稳定的问题。

Pose variations are always difficult to handle in face recognition.Face frontalization can be roughly divided into two categories.The first one is to obtain a frontal face by means of graphics.The second is to get a frontal face through the generative model.However,the frontal face images generated by the first method can be seriously distorted.The second method usually uses the GAN network.However,the process of training is not easy to converge.In order to solve the above problems,this paper proposes a generative network model based on variational auto-encoder and makes the following innovations.First,the model uses theβ-VAE to learn the relationship between hidden space and real picture spaces so that the quality of the generated images is improved.Second,the model uses the simulation+unsupervised(S+U)learning method to the stability of the training process.We conducted experiments on the Multi-PIE dataset,which achieved good results in quantitative scoring and could generate more realistic images.Through experiments,it is proved that the proposed model can effectively solve the problem of generating unrealistic images and unstable training in traditional methods.