【作者】 张明；

【导师】 张彬；

【作者基本信息】 北京邮电大学 ， 电子与通信工程（专业学位）， 2021， 硕士

【摘要】 长久以来,人脸识别是计算机视觉领域中的一个研究热点。现有的人脸识别系统相比几年前已经有了质的飞跃,并且已经应用到了生活中的各个领域,例如安防监控、门禁人脸识别和手机人脸解锁等。但是,现有的人脸识别系统还有许多不足之处,例如,对不同域的人脸识别性能差异过大;在有遮挡、噪声干扰和过暗场景中,人脸识别性能显著下降;并且无法有效地应对检测性能过低的情况。而近年来风靡的强化学习通过结合深度学习模型,已经成为一个新兴的研究领域,且展现出非凡的性能。本文从强化学习的角度,分析了人脸识别模型性能下降的原因,主要由于传统人脸对齐方式缺乏鲁棒性。因此本文基于强化学习寻找解决方案。主要工作包括以下三点:一、在分布式场景下,本文设计并实现了一个新型的强化学习平台,并对主要模块的构成和实现进行了详细的阐述。同时对其运行性能进行了优化,对算法支持部分进行了模块化设计与实现。二、在强化学习应用平台上对人脸对齐任务进行了抽象,并实现了标准点搜索和相似变换矩阵搜索任务。对于标准点搜索,本文使用近似策略优化算法进行了搜索,并使性能得到了提升。对于相似变换矩阵搜索,本文通过间接搜索尺度变换系数,完成了对相似变换矩阵的改进。三、提出了渐进式搜索的方法,提升了强化学习的搜索性能。另外本文还在不同的强化学习算法下实现了搜索任务,并对比了它们对最终性能的影响。实验结果表明,本文提出的标准点搜索和相似矩阵搜索策略改进了现有对齐方式的不足,并在多个测试集上都有效地提高了人脸识别模型的性能。更多还原

【Abstract】 For a long time,face recognition has been a research hotspot in the field of computer vision.The existing face recognition system has made a qualitative leap compared to a few years ago,and has been applied to various areas of life,such as security monitoring,access control face recognition,and mobile phone face unlocking.However,the existing face recognition system still has many shortcomings,for example,the face recognition performance of different domains is too different;in the scenes with occlusion,noise interference and too dark,the face recognition performance is significantly reduced;and cannot effectively deal with the situation that the detection performance is too low.In recent years,the popular reinforcement learning has become an emerging research field through the combination of deep learning models,and has shown extraordinary performance.From the perspective of reinforcement learning,this paper analyzes the reasons for the decline in face recognition model performance,which mainly comes from the lack of robustness of traditional face alignment methods.Therefore,this article is based on reinforcement learning to find a solution.The main work includes the following three points:1.In a distributed scenario,this paper designs and implements a new type of reinforcement learning platform,and elaborates on the composition and implementation of the main modules.At the same time,its operating performance is optimized,and the algorithm support part is modularized and implemented.2.The face alignment task is abstracted on the reinforcement learning application platform,and the searching of standard point and the searching of similar transformation matrix are realized.For the standard point search,this paper uses the approximate strategy optimization algorithm to search,and the performance is improved.For the similarity transformation matrix search,this paper completes the improvement of the similarity transformation matrix by indirectly searching the scale transformation coefficients.3.A progressive search method is proposed to improve the search performance of reinforcement learning.In addition,this paper implements search tasks under different reinforcement learning algorithms,and compares their effects on the final performance.The experimental results show that the standard point search and similarity matrix search strategies proposed in this paper improve the shortcomings of the existing alignment methods,and effectively improve the performance of the face recognition model on multiple test sets.更多还原