【作者】 张力；

【导师】 刘乐元；

【作者基本信息】 华中师范大学 ， 通信与信息系统， 2020， 硕士

【摘要】 现在许多基于深度学习的人脸识别算法已经在现实生活中得到了广泛的应用,但是这些人脸识别的算法大多数需要待识别的人脸具有比较正面的姿态,如果人脸的姿态比较歪或者是整个侧面人脸,则会导致人脸识别错误率大大提高,甚至导致人脸识别系统的崩溃。于是头部姿态影响人脸识别系统的无法正常运转的问题变成了亟待解决的问题。国内外的许多研究者也开始慢慢的研究姿态对于人脸识别的影响,研究的解决办法主要分两大类,一种是通过学习一个对姿态具有鲁棒性的卷积神经网络,另一种通过生成对抗网络的方法进行人脸正面标准化的过程。本文提出的方法是一种渐进式姿态归一化生成对抗网络的人脸合成算法(PPN-GAN),与其他的合成算法相比,本算法合成的图像具有更高的质量,合成之后的人脸识别也具有更高的精度。由于其他的合成算法都是一步法,不管侧面人脸的角度有多大,通过设计不同的生成对抗网络及损失函数,将这些侧面人脸直接合成正面人脸。本文提出的算法则是多步法,通过渐进式的生成策略,逐渐的将侧面人脸合成正面人脸。因为将侧面人脸合成正面人脸的这个问题,本身就是一个病态的问题,是一个逆问题,由于侧面人脸到正面人脸的信息丢失太多,导致很难一步直接从侧面人脸合成正面人脸,一步法合成图像会出现难以训练而且图像质量很差的问题。渐进式姿态归一化生成对抗网络的人脸合成算法,通过寻找多个中间的人脸来过度,我们逐渐的将侧面人脸恢复到中间的人脸,再从中间的人脸去恢复到下一个中间的人脸,直至最后恢复到正面人脸。除了渐进式策略之外,我们在图像空间和特征空间都引入身份判别器,两个身份保留损失都合并到GAN中以保存身份。这种渐进策略主要从两个方面提高了正面人脸合成的性能:(1)分解不确定信息量,从而减轻GAN面临的学习负担。(2)在合成人脸的中间视图时可以探索更多的数据先验知识,这有助于恢复不确定信息。本算法主要有以下三个创新点:(1)我们提出了一个渐进式姿态归一化生成对抗网络的姿态归一化的框架,这个框架用于人脸正面合成。该框架可以在大姿态下生成具有高质量的、身份一致的正面人脸图像;(2)我们将图像空间和特征空间都进行了身份保留,并纳入了我们的损失函数中,这样有助于在训练的时候忠实地保存身份信息不变;(3)我们设计了一个“路由选择”器,在测试阶段,我们可以根据“路由选择”器来决定PPN-GAN的合成路径,这样也有利于PPN-GAN的人脸正面合成。实验结果也表明,我们的方法不仅产生令人信服的感知结果,而且在大姿势下的人脸识别方面也优于最新方法。更多还原

【Abstract】 Nowadays many deep learning-based face recognition algorithms have been widely applied in real life,but most of these face recognition algorithms require that the face to be recognized has a more positive posture.If the posture of the face is relatively lateral or it is one of the entire side face,it will lead to a greatly increased error rate in face recognition,and even cause the crash of the face recognition system.Therefore,the problem that the head posture affects the normal operation of the face recognition system has become an urgent problem.Many researchers at home and abroad have also begun to study the impact of poses on face recognition gradually.The solutions of the problem are divided into two categories,one is the learning of a convolutional neural network that is robust to poses,and the other is a process of face normalization by generating adversarial networks.The method proposed in this paper is a face synthesis algorithm for Progressive Pose-Normalization Generative Adversarial Network(PPN-GAN).Compared with other synthesis algorithms,the image synthesized by this algorithm has higher quality,and the face recognition after synthesis also has higher accuracy.The other synthesis algorithms are one-step methods which regardless of the angle of the profile image.By designing different generative adversarial networks and loss functions,these profile images are directly synthesized into the frontal view image.On the contrary,the algorithm proposed in this paper is a multi-step method.Through a progressive generation strategy,profile images are gradually synthesized into frontal faces.The problem of synthesizing a profile image into a frontal image itself is a morbid problem,and also an inverse problem.Because too much information is lost during the synthesis of a profile face to the frontal face,it is difficult to synthesize the frontal face directly from the profile face.Meanwhile,synthesis through one-step method makes the model difficult to train and the quality of the synthesized image is low.Progressive Pose-Normalization Generative Adversarial Network for frontal face synthesis transits by searching for multiple intermediate faces.We gradually transform the profile face to the middle face,and then transform from the middle face to the next middle face,until it finally returns to a positive face.In addition to the progressive strategy,an additional identity discriminator and identity aware losses in both the image and feature spaces are also incorporated into the GAN for identity preserving.This progressive strategy improves the performance of frontal face synthesis mainly in two aspects:(1)The amount of uncertain information is decomposed,so that the learning burden faced by GAN is reduced.(2)More prior knowledge from data can be explored during synthesizing the intermediate views of a face,which is helpful for recovering the uncertain information.The main contributions of this Method are threefold:(1)We propose a progressive pose-normalization framework based on GAN for frontal face synthesis,which can recover photorealistic and identity-preserving frontal faces from face images under extremely large pose.(2)We incorporate an additional identity discriminator and identity-aware losses for both the image and feature spaces into the GAN architecture,which helps faithfully preserve identities.(3)We introduce a sample-specific "pose router" to direct different testing samples into corresponding stages of the PPN-GAN,so that optimized frontal faces can be synthesized by the PPN-GAN.Experimental results show that our method not only produces compelling perceptual results but also outperforms the state-of-the-art methods on face recognition under large-pose.更多还原