【毕业学位论文】面部特征点定位方法研究-计算机应用技术博士论文

摘要摘要在计算机视觉和图像感知研究领域内，目标的检测和定位是一个研究的热点问题，而且许多相关问题的研究都依赖于目标检测的结果和定位的精度，因而得到越来越多研究者的重视。在人脸相关的领域内，面部特征点定位作为人脸图像分析问题的一个关键步骤，为后续的人脸识别、姿态估计、表情分析、人脸动画等提供了技术保证。但是由于在很多情况下人脸图像存在着姿态变化，表情变化，光照变化等一系列复杂变化，受此影响人脸图像特征点定位的精度还不能满足实际应用的要求，所以研究出速度快，精度高，适用性好的面部特征点定位方法就成为了人脸相关的研究中的当务之急。本文正是针对这一重要研究课题，进行了深入的理论研究和应用实践。对面部特征点定位和形状提取问题的研究，主要的研究内容集中在两个方面：一个是建立独立特征点的局部纹理模型，另一个是建立所有特征点的全局形状模型。本文从这两个方面着手提出了新方法，取得了新成果。本文的主要研究成果包括：(1)提出基于双向级联结构的特征点检测和定位方法。所有面部特征点总体的定位准确度是和每一个独立的特征点的定位精度分不开的，尤其是眼睛，鼻子，嘴等一些关键特征点的定位，这些特征点为更多其他的面部特征点提供初始的搜索范围，使得定位算法更稳定。本文通过这种双向级联结构的分类器设计来对特征点的局部纹理建模，这样可以更加精确，同时也更加快速地检测和定位这些关键特征点的位置。所谓双向是指在训练过程中，对正例样本和反例样本交替进行重采样的过程。与传统方法相比本文方法具有以下几个优点：可以应对大规模的数据集；可以处理存在复杂变化的正反例样本；而且无论是在训练过程，还是在测试过程，该算法都能快速地过滤大量的“易分”样本，获得了较高的执行效率。(2)提出一种增强的主动形状模型方法。该方法是在单个特征点独立检测的基础上，引入了全局的形状模型，并且将每个特征点检测器的输出值作为检测到这个特征点的信度。将那些检测信度比较低的特征点作为缺失点进行处理，这些点的最终位置不通过特征点检测器给出，而是通过信度比较高的特征点来预测得到。这种预测是以最大化形状概率为目标，解析地得到不可信特征点的位置，因而当检测器输出的信度较低或者是检测不到特征点时，算法的定位精 I 哈尔滨工业大学工学博士学位论文度不会受到太大的影响，可以应对存在缺失特征点的情况。通过以上的设计，本文将传统的主动形状模型进行了有效地增强。(3)提出基于贝叶斯推理的人脸形状提取方法。在该方法中，本文将单个特征点的分布和全局形状的分布用概率描述的形式来表达，统一到贝叶斯推理的概率框架之中。特征点定位的目标转化为最大化形状的后验概率，然后通过迭代求解的方式使得所有特征点最终收敛于正确的面部特征点位置。在推理的过程中，本文对控制形状的两种参数（包括几何参数和形状参数）分别进行优化：根据特征点位置移动和这两种参数变化之间的关系分别采用了梯度上升和高斯牛顿优化算法。并且根据算法的推导结果定义了一种称为“概率梯度提示”的变量，人脸形状正是在这种变量的驱动下不断的更新，从而完成面部特征点的定位和人脸形状的提取。(4)提出基于人脸坐标回归模型的特征点定位和形状提取方法。基于上述的特征点检测方法和形状建模方法，本文进一步提出基于人脸坐标系的人脸形状表示，将面部特征点定位问题转化为人脸坐标的对应问题。在这个人脸坐标系的定义下，本文提出了基于人脸坐标回归的特征点定位和形状提取方法，对图像上的局部表观使用回归算法得到对应的人脸坐标，并且结合形状模型设计了使得所有特征点的回归误差之和最小的目标函数，最后通过迭代优化的方式求解最佳的形状。本文提出的以上几个算法和设计在多个人脸数据库上进行了实验和验证，实验结果表明本文提出的面部特征点定位方法是有效的和可行的。关键词：面部特征点定位；分类器设计；增强主动形状模型；形状优化；坐标回归 II AbstractAbstractIncomputervisionandimageunderstandingfield,thedetectionandlocalizationofatargetobjectwithinanimageisanactivetopic. Infacerelatedresearches,faciallandmarklocalization act as a key step in face image analysis. Many other related researches, suchas face recognition, pose estimation, expression analysis and face animation are dependon the results of the localization. Therefore facial landmark localization attracts moreand more attention. However, due to many real world effects, including pose variations,expression variations and lighting conditions, the results of facial landmark localizationare not accurate enough. Therefore it is necessary to develop accurate, efficient andcapable methods for facial landmark localization in face related researches. Aiming atthis target, we carried out theoretical studies and practical applications.There are mainly two aspects in facial landmark localization and face shapeextraction: one is the local texture modeling for a single landmark, the other is the globalshape modeling for all landmarks. We implement new methods in the two aspects andachieve good results. In this thesis, the main contributions are summarized as follows:(1) This thesis proposes bidirectional cascaded classifiers for facial landmarkdetection and localization. The overall accuracy of landmark localization depends onthe individual landmark detectors, especially some key landmarks which provide theinitial positions for the other landmarks make the localization algorithms more stable.We exploit the bidirectional cascaded classifier to model the local texture and combinethe feature extraction and pattern classification methods. Here,“bidirection”refers tothe procedure that adopts bootstrap to resample the positive and negative samples in turnsin the training phase. Compared with the traditional methods, the method of this thesisis of the following advantages: First, it can cope with large scale data sets. Second, itcan deal with complicated variations of the positive and negative training samples. Third,no matter in the training stage or the testing stage, the algorithm can rapidly reject largenumbers of simple samples, which brings high efficiency.(2) This thesis proposes the enhanced active shape model to solve the landmarkmissing problem. We introduce a global shape model to integrate the individual landmarkdetectors and take the outputs of the landmark detectors as the final confidence of the III 哈尔滨工业大学工学博士学位论文landmarks. The landmarks which are detected with a low confidence are considered asmissing landmarks. In this situation, the positions of these landmarks are predicted byother reliable ones together with the global shape model but rather than the outputs of thelandmark detectors themselves. The prediction of the unreliable landmarks is derivedby maximizing an objective function of shape probability analytically. Therefore thelocalization algorithm can solve the missing landmark problem and the accuracy will notbe affected too much when some landmarks are detected with large errors. Using thisstructure, the active shape model is enhanced effectively.(3) A face shape extraction method based on Bayesian inference is proposed in thisthesis. The distributions of individual landmarks and the global shape are representedas the probability distributions and integrated into the Bayesian inference framework.The goal of localizing the landmarks is then transformed to maximize a posteriori. Twotypesofparameterswhichcontroltheshape,includingthegeometricparameterandshapeparameter,areoptimizedseparatelyduetotherelationshipbetweentheshapeinstanceandparameters. Gradient ascent method and Gaussian-Newton method are used respectivelyin the optimization procedure. A measurement of probability gradient hints (PGH) isdefined according to the derivation of the algorithm. The face shape is updated iterativelyand driven by PGH to achieve the goals of facial landmark localization and face shapeextraction.(4) We propose a facial landmark localization and face shape extraction methodbased on the face coordinate regression. On the basis of the above mentioned frameworkof landmark detection and shape modeling, we further represent the face shapes in aface coordinate system. The problem of facial landmark localization and face alignmentis fulfilled by corresponding face coordinates in each face image. A facial landmarklocalization method is proposed under this face coordinate system via face coordinateregression. The regression method reveals the face coordinates according to the localappearance. Theoptimalshape isobtainedbyiterativelyoptimizing anobjectivefunctionwhich combines the regression error and shape constraints.The above algorithms are verified on several face databases. The experimentalresults validate the effectiveness and feasibility of our methods for facial landmarklocalization. IV AbstractKeywords: facial landmark localization; classifier design; enhanced active shape model;shape optimization; coordinate regression V 目录目录摘要. IAbstract . III插图. XIII表格. XIX算法. XXI第1章绪论. 11.1课题研究背景. 11.1.1问题提出. 11.1.2研究意义. 21.1.3主要挑战. 21.2面部特征点定位问题综述. 31.2.1面部特征的表示方法. 41.2.2面部特征的纹理模型. 41.2.3人脸形状的形状模型. 101.2.4纹理模型和形状模型的关联. 131.2.5人脸形状的求解方法. 161.2.6本节小结. 171.3本文主要研究内容. 181.3.1主要贡献. 181.3.2论文的组织结构. 19第2章基于双向级联结构的面部特征检测和定位. 212.1引言. 212.2基于Haar特征和AdaBoost分类器的目标检测方法. 222.2.1 Haar特征提取. 222.2.2 AdaBoost分类器. 232.2.3用于面部特征点定位. 252.3双向级联结构分类器. 262.3.1问题分析和基本思路. 262.3.2分类器设计. 272.3.3分类器训练过程. 282.3.4分类器决策过程. 29 VII 哈尔滨工业大学工学博士学位论文2.4实验与分析. 302.4.1实验数据库. 302.4.2性能评价准则. 302.4.3实验项目. 312.5本章小结.