【作者】 杨宇；

【导师】 胡红平；

【作者基本信息】 湖南大学 ， 信息与通信工程， 2017， 硕士

【摘要】 随着计算机科学的高速发展,人机交互已经融入了人们的工作与生活中,人工智能技术成为了计算机领域的研究热点和重点。人脸识别技术,正是人工智能领域的核心技术之一。经过近三十年的快速发展,人脸识别作为一种独特的生物特征识别方式,在社会公共安全和日常生活中有着广泛的应用。虽然受控环境下的人脸识别应用已经能基本满足要求,但在非限制条件下的应用仍存在许多问题。其中,如何提取有效的人脸特征和设计出鲁棒的分类算法是人脸识别系统有待解决的两个关键问题。本文基于深度学习算法,针对人脸图像的特征提取和鲁棒分类两个问题进行了深入研究,以深度学习模型为主体,提出一种结合局部模式的特征提取算法和一种基于动态随机森林的分类方法。主要完成了如下三部分工作:1)本文认真研究了大量的人脸识别算法,分析了现有人脸图像特征提取过程和分类过程中存在的问题,详细叙述了深度学习的发展历程,揭示了深度学习相对于浅层学习的优势所在,对深度学习的常用模型进行了举例说明,还详细介绍了局部量化模式、限制性玻尔兹曼机和深度置信网络的原理。2)本文提出了基于深度学习和局部量化模式的人脸图像特征提取方法。在利用深度学习模型提取人脸特征时,现有方法直接将人脸图像中的原始像素作为深度网络的输入,在非限制条件下由于图像受到表情变化、姿态差异、光照强弱等因素的干扰,网络会学习到不利的人脸特征表达。而且,现实环境中带标签的人脸样本数量较小,不足以用作训练样本充分调整网络结构。为了解决上述问题,本文首先利用局部量化模式对经Gabor滤波器预处理后的人脸图像提取低层局部特征,然后将提取到的局部特征组合成一个整体作为深度置信网络可视层的输入,训练网络,提取高层抽象特征。与现有方法相比,本文方法同时利用了局部描述子强大的图像可视化表达能力和深度学习模型自动提取特征的优势,克服了深度学习模型在特征提取过程中的缺陷,表现出更好的有效性和普适性。3)本文对深度网络结构进行优化,改进了网络结构,提出将动态随机森林作为深度学习模型的顶层分类器,替代固有的支持向量机和softmax分类器。首先用前文提出的深度模型提取样本的抽象特征,然后将提取到的特征输入到动态随机森林中进行分类,以此构成一种新的深度分类模型。通过实验验证,本文提出的深度分类模型提高了对人脸图像识别的准确率,并具有更好的鲁棒性。更多还原

【Abstract】 With the rapid development of computer science,human-computer interaction has been integrated into people’s work and life,artificial intelligence technology has become one of the most active and important research topics in the field of computer.Face recognition technology,it is one of the core technology in the field of artificial intelligence.After nearly three decades of rapid development,as a unique biometric identification,face recognition has a wide range of applications in the social public security and daily life.Although the applications in a controlled environment has been able to basically meet the requirements,there are still many problems for the application under unrestricted conditions.Among these problems,how to extract effective facial features and design a robust classification algorithm are the two key problems to be solved for face recognition system.Based on the deep learning algorithm,this thesis makes a deep research on feature extraction and robust classification of face images,and proposes a new feature extraction algorithm combining the depth model with local patterns and a new classification method based on dynamic random forests.The mainly completed work as the following three parts:1)This paper studies a lot of face recognition algorithms,analyzes the problems existing in the process of feature extraction and classification,elaborates the development of deep learning,reveals the advantages of deep learning relative to shallow learning,examples of commonly used models of deep learning,and introduces the principles of local quantized patterns,restricted boltzmann machine and deep belief networks in detail.2)This paper presents a novel face feature extraction method based on deep learning and local quantized patterns.When the features of face is extracted by deep learning model,the existing methods directly use the original pixels of face image as the network input,the network will learn some unfavorable feature representation in unconstrained environments where the image is affected by the change of expression,the difference of attitude and the intensity of illumination and so on.Moreover,in the real world,the number of labeled face samples is small and insufficient to be used as training samples to adjust the network structure.In order to solve the above problems,this paper first uses the local quantization model to extract low-level local features of the face images preprocessed by gabor filters,and then uses the intergral feature which is integrated by the local features as the visual layer input of deep belief network,trains the deep network,extracts high-level abstract features.Compared with the existing methods,our approach not only utilizes the local descriptor’s powerful performance of image visualization,but also takes advantage of deep learning’s automatic way in feature extraction,overcomes the defects of the deep network in the process of feature extraction,showing better effectiveness and universality than other methods.3)This paper optimizes and improves the structure of deep network,and proposes to use dynamic random forests as the classifier at the top of deep learning model,instead of the inherent support vector machine and softmax classifier.Firstly using the depth model proposed above to extract the abstract characteristics of the sample,and then input them into the dynamic random forests for classification,constitute a new depth classification model by this way.The experimental results show that the depth classification model proposed in this thesis improves the accuracy of face image recognition and has better robustness.更多还原