【作者】 张浩；

【导师】 徐巍华； 刘勇；

【作者基本信息】 浙江大学 ， 控制科学与工程， 2021， 硕士

【摘要】 人脸识别是人工智能和计算机视觉研究中的一大热点与重要落地场景,尽管近年来已取得长足进步,但其鲁棒性仍有待提升。在影响人脸识别鲁棒性的众多因素中,大角度姿态和伪造人脸攻击十分常见,会直接导致人脸信息部分丢失和人脸信息来源不可靠,尤其难以应对。尽管有方法通过检测人脸属性来控制识别前人脸的质量,从而提升人脸识别的鲁棒性,但其不足在于没有同时考虑大角度姿态和伪造人脸攻击给人脸识别鲁棒性带来的影响,姿态估计和防伪算法的精度不够好,以及对伪造人脸的新型方式不具备防御能力。本文提出更优的头部姿态估计和人脸防伪算法,实现基于姿态和真伪先验的鲁棒人脸识别技术,致力于提升人脸识别的鲁棒性。主要内容和研究成果如下:1.提出了一种以单帧RGB图像作为输入,无需人脸关键点,准确且轻量化的头部姿态估计算法。首先提出特征解耦模块,其中采用通道注意力机制来自适应调整不同角度类别特征图在通道层面的响应,从而为每个角度学习更加有区分性的特征。接着提出跨类别中心损失来约束不同角度类别在隐变量子空间的分布,可以获得更为紧凑和更特异的特征子空间。最后引入逆残差块和全局平均池化,极大降低模型参数量。该算法在多个公共数据集上精度优于最新的方法,且网络结构十分轻量化。2.设计了一种更为通用,泛化能力更强的人脸防伪算法。采用高分辨率网络提取人脸特征,实现对细节纹理信息更好的捕捉。提出将数据标签重整合为更细粒度的层级结构,从主干网络引出粗和精两部分输出,以层级监督的方式训练网络,使网络学习到真实人脸和各种伪造人脸更有区分性的特征。提出双流框架,设计输入采样器和质量不变损失来提升算法对质量变化的鲁棒性。该算法在多个公共数据集精度和泛化能力超出近年来多种方法,并在FaceForensics Benchmark排名第十。3.实现了 一种对姿态变化和伪造人脸攻击鲁棒的人脸识别技术。首先通过改进当前先进的人脸检测和人脸识别算法来权衡网络精度和模型参数量大小,完成1:N人脸识别的功能。接着构建姿态检测和真伪检测两个语义感知模块,在人脸识别过程中感知输入图像相应信息。最后实现完成基于姿态和真伪先验的鲁棒人脸识别技术,对检测到的属性先验做出应对,在图像和视频场景下能够进行鲁棒的人脸识别。更多还原

【Abstract】 Face recognition is a hot spot and important landing scene in artificial intelligence and com-puter vision research.Although considerable progress has been made in recent years,its robustness still needs to be improved.Among the many factors that affect the robustness of face recognition,large-angle pose and forged face attacks are very common,which will directly lead to partial loss of face information and unreliable sources of face information,which is especially difficult to deal with.Although there are methods to control the quality of the face before recognition by detect-ing face attributes,thereby improving the robustness of face recognition,the disadvantage is that it does not simultaneously consider large-angle poses and fake face attacks to bring the robust-ness of face recognition.As a result,the accuracy of posture estimation and anti-counterfeiting algorithms is not good enough,and the new ways of forging faces are not defensive.This paper proposes a better head pose estimation and face anti-counterfeiting algorithm to implement robust face recognition technology based on pose and authenticity priors,and is dedicated to improving the robustness of face recognition.The main content and research results are as follows:1.This paper proposes an accurate and lightweight head pose estimation network that takes a single frame of RGB image as input and does not require face key points.A feature decou-pling module is proposed,in which the channel attention mechanism is used to adaptively adjust the response of different angle category feature maps at the channel level,so as to learn more distinguishing features for each angle.Then,the cross-category center loss is proposed to constrain the distribution of different angle categories in the latent variable sub-space,and a more compact and specific feature subspace can be obtained.By introducing inverse residual block and global average pooling,the amount of model parameters is greatly reduced.The accuracy of this algorithm is better than the latest methods on multiple public datasets,and the network structure is very lightweight.2.A more accurate and generalized face anti-counterfeiting algorithm is designed.The high-resolution backbone network is used to extract image features,and high-resolution and low-resolution depth features are merged in the classification head to achieve better capture of image details.It is proposed to reintegrate data labels into a more fine-grained hierarchical structure,and draw two parts of output from the main network,coarse and fine,and train the network in a supervised way from coarse to fine level,so that the network can learn more distinguishing features among real faces and various fake faces.A Two-stream framework is proposed,and the input sampler and quality invariant loss are designed to improve the algorithm’s robustness to quality changes.The accuracy and generalization ability of this algorithm in multiple public data sets exceeds some methods in recent years,and ranks tenth in the FaceForensics Benchmark.3.A face recognition technology that is robust to pose changes and forged face attacks is re-alized.First,by improving the current advanced face detection and face recognition algo-rithms to weigh the network accuracy and the size of the model parameters,the 1:N face recognition function is completed.Secondly,two semantic perception modules are con-structed for posture detection and forgery detection to perceive the corresponding informa-tion of the input image during the face recognition process.Finally,a robust face recognition technology based on posture and authenticity priori is implemented,which is responded to the detected attribute priors,and robust face recognition can be performed in image and video scenes.更多还原