【作者】 王小明；

【导师】 刘锦高； 黄昶；

【作者基本信息】 华东师范大学 ， 通信与信息系统， 2010， 博士

【摘要】 人脸检测与人脸识别是模式识别与计算机视觉领域的重要研究课题,在公共安全、智能监控、视频会议、多媒体和数字娱乐等领域有广阔的应用前景。经过三十多年的研究,人脸检测与识别技术在可控环境中获得了很大的发展,在理想情况下已经能够取得可以接受的检测与识别的性能。但是在不可控环境中,由于受到多种因素的影响,如光照、姿态、表情、遮挡等,人脸检测和识别的性能会有明显的下降。人脸检测与人脸识别要真正走向实用仍然极具挑战性。本文对人脸检测与人脸识别中的人脸检测器训练,人脸特征抽取、人脸分类器设计等问题进行了研究,同时针对可变光照下的人脸检测与人脸识别问题,从图像增强和获取人脸光照不变特性的角度进行了深入的研究。论文的主要研究成果概括如下：(1)针对应用Adaboost算法进行人脸检测时误检率比较高的问题,提出了级联支持向量机的Adaboost-SVM人脸检测算法。Adaboost算法利用Haar小波,积分图和级联的思想实现了接近实时的人脸检测,但是对于复杂的背景存在着人脸误检率比较高的问题。本文首先利用Adaboost算法训练了一个人脸检测器,然后用该检测器对图像数据库进行自动的人脸检测,再对检测结果进行人工纠正分类,产生人脸正样本和负样本,利用这些样本训练基于SVW的人脸检测器。将这两个检测器级联起来,构成Adaboost-SVM人脸检测器。实验表明,Adaboost-SVM人脸检测算法在基本保持人脸检测率的情况下,误检率有明显的下降,误检窗口数最高下降达到82.91%。(2)当环境的光照比较复杂时,人脸检测的性能会有明显的下降,为了减少光照的影响,提出了快速的自适应图像增强算法,用于改善复杂光照下的人脸检测性能。本文深入研究了Retinex图像增强理论,针对多尺度Retinex算法存在的运行速度慢,增强后图像容易灰度化的现象,提出了一种改进的多尺度Retinex图像增强算法。同时结合对数变换和非线性变换,进一步提出了快速的自适应图像增强算法,与直方图均衡、单尺度Retinex和多尺度Retinex算法相比,自适应图像增强算法在人脸检测率和误检率方面都有明显的改善。(3)针对Gabor小波人脸特征表示数据维数过高的问题,研究了利用二维线性子空间进行特征降维的方法,并实现用支持向量机进行人脸分类的策略。人脸识别的过程中,利用多个尺度和不同方向的Gabor小波来表示人脸图像,并构成Gabor特征脸,然后利用二维线性子空间方法直接对Gabor特征脸进行特征降维,再采用支持向量机对人脸进行分类。实验表明,这种方法有效地提取了有利于分类的人脸特征,同时解决了Gabor小波的维数灾难问题,取得了良好的分类效果。(4)综述了复杂光照下的人脸识别算法,提出了形态学小波商图像算法。寻找具有光照不变性的人脸特征图像是解决复杂光照人脸识别的一个有效途径,对人脸图像进行形态学闭运算操作后,再在小波域去掉图像的高频成分,所得结果作为人脸图像的光照估计,将光照图像与原图像相除后得到一种商图像。实验结果表明,这种形态学小波商图像具有光照不变性,同时相比于SSR算法和形态学商图像,形态学小波商图像更能保留复杂光照下人脸图像的识别信息,识别效果更好。上述研究成果分别从降低人脸误检率的策略,人脸特征抽取和识别,可变光照下人脸检测和识别性能的提高等方面给出了具体的方案和实验结果,对于人脸检测和识别的理论研究和应用推广有一定的参考价值。更多还原

【Abstract】 Human face detection and recognition is an important research subject in the field of pattern recognition and computer vision. It has been widely used in many applications, such as public security, smart surveillance, video conference, multimedia and digital entertainment. After over 30 years of development, this technology has been developed steadily in controlled conditions and gains the better performance in an optimal situation. However, when it comes to uncontrolled conditions, such as different illumination conditions, pose variations, mixture of emotions and object shelter, the accuracy of face detection and recognition will dramatically decline. Therefore, the research faces enormous challenges in real-world applications.This paper does a lot of research on the training of the face detector, the face feature extraction, the design of the classifier, etc. In particular, this paper focuses on the problem of face detection and recognition in complex illumination environment and describes the work on face image enhancement and the obtaining of the invariant face features under illumination. The major contributions of this paper are as follows:First, in order to reduce the high false face detection rate of the Adaboost algorithm, the algorithm of Adaboost-SVM is presented. The Adaboost algorithm uses the Haar wavelet and the integral image and cascading to realize the real time face detection, its false face detection rate is high under complex background. In this paper, a face detector using the Adaboost algorithm(FD_Adaboost) is trained firstly and is used to detect faces in the face database,then, the face detection result is manual classified to two class:positive class and negative class,after that,the two-class training samples is used to train the face detector based on SVM(FD_SVM). Finally, FD_Adaboost and FD_SVM are cascaded to constitute the Adaboost-SVM face detector. Experiment results show that the Adaboost-SVM algorithm achieves good results on face detetion and obviously reduces the false face detection rate, the number of false detection window can be mostly reduced by 82.91%.Second, studies show that the performance of face detection will be obviously reduced under the complex illumination environment. In order to solve this problem, this paper presents a fast self-adaption image enhancement algorithm. To deal with the color distortion of enhanced image and reduce the computational complexity, the paper proposes an improved multi-scale Retinex algorithm. Meanwhile, the paper presents a fast self-adaption image enhancement algorithm that combines logarithmic transformation with nonlinear transformation. Compared with histogram equalization(HE), single-scale Retinex(SSR) and multi-scale Retinex(MSR), the self-adaption image enhancement algorithm increases the face detection rate and reduces the false face detection rate obviously..Third, in order to solve the problem of the high dimensionality of Gabor wavelet facial feature, the paper reduces the feature dimension by using two-dimensional linear subspace and classifies the face images using SVM. Firstly, the face image is processed by multi-scale and multi-directional Gabor wavelet to produce the Gabor EigenFace.Then, the feature dimension of the Gabor EigenFace is reduced by using two-dimensional linear subspace, finally, SVM is used to classify the face images. Experiment results show that this method not only acquires the facial feature effectively, but also solves the dimensionality curse caused by Gabor wavelet. The classification result is perfect.Fourth, this paper gives a review on face recognition algorithm under complex illumination environment and proposes a new morphological wavelet quotient image algorithm. The facial feature independent of illumination is contributed to improve the face recognition rate in complex illumination environment. In order to get the above facial feature, a morphologic closing computation operator is applied to process the face image, then, filters the high frequency components in wavelet domain. The result image(RI) is the lighting estimation of the source face image(SI).when RI and SI divide each other, the quotient image is independent of illumination. Compared with single scale Retinex and morphological quotient image (MQI), the new method has better performance on keeping face recognition feature under complex illumination environment and achieves a high recognition rate.In conclusion, This paper discusses the strategy of reducing the false face recognition rate, the face feature extraction and recognition and the improvement of the face detection and recognition under complex illumination environment, moreover, it puts forward specific proposals and experiment results which contributes to the research of the face detection and recognition and the practical application更多还原