复杂背景下快速人脸检测与识别

【作者】 胡伏原；

【导师】 张艳宁；

【作者基本信息】 西北工业大学 ， 计算机软件及理论， 2004， 硕士

【摘要】 人脸的检测与识别的研究涉及模式识别、图像处理、生理学、心理学、认知科学等许多领域，与基于其它生物特征的身份鉴别方法以及计算机人机交互领域都有密切联系。由于各领域对人脸检测与识别技术的需求日益迫切(特别是安全领域)，这使得人脸检测与识别技术的研究已经成为当今的热点问题之一。 本文主要研究了复杂背景下快速人脸检测与识别的基本理论和关键技术。重点讨论了人脸表示、分类器的设计和复杂背景下快速人脸检测与识别系统架构问题。论文的主要研究工作和主要成果包括以下几个方面。 首先，研究了不同彩色空间下不同的肤色模型。通过实验表明，在HSI彩色空间的简单高斯模型和线性模型比较适合于复杂背景下的快速、准确的人脸检测。通过肤色检测能快速去除复杂背景，缩小检测的范围，为复杂背景下快速准确的人脸检测奠定了基础。但是肤色模型易受光照等外界因素影响，本文结合K-L变换提高了肤色检测的鲁棒性。 其次，研究了人脸表示及其快速算法。通过对PCA、LDA、Gabor和Like-Harr人脸表示方法的研究，分别为人脸检测和识别选取了较为稳定而准确的人脸表示；初步尝试了利用AdaBoost分类器进行特征选择从而消除冗余特征，并且提出了采用级联表示方法快速表征人脸，从而实现由粗到精、由简单到复杂的快速人脸表示，这样既提高了人脸的检测和识别的速度，还有利于检测率和识别率的提高。 随后，从三个方面研究了分类器的构造。首先研究了在人脸检测和识别中常用的分类器，比如符号函数、最近邻、神经网络、SVM、Adaboost等，选择了适合于人脸检测和识别的分类器，并提出了结合PCA特征和RBF进行人脸姿态的判别方法：其次研究了具有特征选择功能的分类器发计，这为人脸的级联表示提供了条什，也为快速准确的人脸检测和识别提供了可能；最后，对组合分类器设计进行了研究，提出了适于复杂背景下快速人脸检测和识别的有效分类器设计方案，这使得人脸检测和识别能够快速剔除不感兴趣区域，为复杂背景下实时人脸检测和大型人脸库的快速识别提供了可能。 最后，结合人脸的级联表示和组合分类器的设计思想提出了一个复杂背景下快速的人脸检测和识别系统架构，并实现了该系统。该系统实现了在复杂背景下快速的人脸检测和识别，在P4 1.6的机子上检测和识别的总耗时160ms，约为6帧／秒。实验结果验证了该系统的实时性、准确性和鲁棒性。更多还原

【Abstract】 Face detection and recognition research is related to the fields of pattern recognition, image processing, physiology, cognitive science and so on. At the same time, it is tied with other researches such as Biometric Authentication and Human-Computer Interaction. It has become one of the key issues because of its practical essentiality in many aspects, especially in the field of security.In the thesis, the fundamental theories and the key technologies of fast face detection and recognition from images with complex background are mainly researched. The topics such as face representation, classifier design and the design of Face Detection and Recognition System for images with complex background are discussed in details. Research work includes the following contents.Firstly, different skin models are researched in four different color spaces. The experiments show that Linear Model or Single Gaussian Model in HIS color space can get better results under complex background. With the help of skin models, it can get rid of the complexity of the background, confine the range of valid areas and lay a basis for fast and accurate face detection. However, it is easily affected by illumination. The robustness is improved by considering the K.-L transformation.Secondly, the face representation and its fast algorithms are summarized. There are some research approaches such as PCA, LDA, Gabor and Like-Harr etc to get features, and some stable and efficient facial features are obtained by these means. The method of selecting features by Adaboost is used to discard redundant features. Cascade representation is adopted to represent face rapidly to improve the fast face representation from coarse to fine, from simple to complex, and to improve the speed and accuracy of the processing.Subsequently, designs of the classifiers are described in three aspects. First of all, some common classifiers are firstly discussed such as SF(Sign Function), NN(Nearest Neighbour),Adaboost, SVM(Support Vector Machine), ANN(Artificial Neural Network) etc. Then appropriate classifiers for face detection and recognition are selected. After that, the method of detecting poses of faces by RBF and PCA is proposed. The classifier with ability of feature selection is studied to prepare for face cascade representation and to make it possible to detect and recognize face fast and accurately. Finally, construction of an array of classifiers is researched, and an effective method to design classifiers of fast face detection and recognition with complex background is presented, which is able to radically discard redundant areas and realize a robust real-time face detection designed for complex background and recognition system with large face database.Finally, a fast face detection and recognition system for images with complex background is proposed and implemented by combining face cascade representation and classifier design. The model system accomplishes face detection and recognition from images with complex background robustly and accurately. When applied, the system runs at 6 frames per second in P41.6G.更多还原