【作者】 杨春友；

【导师】 楼新远；

【作者基本信息】 西南交通大学 ， 计算机软件与理论， 2013， 硕士

【摘要】 近年来,基于生物特征的身份认证技术得到了飞速发展,人脸检测作为生物特征认证技术的一个重要应用,越来越受到人们的重视。人脸检测在视觉图像处理、人机交互、表情识别、视频监控以及基于内容检索等领域有着很重要的作用和实际应用价值。当前,已有许多关于人脸检测算法的研究,但是由于人脸本身具有相当的复杂性,很难找到一种算法能够适应各种条件下的检测任务,更多的研究是基于多种算法的有效结合,从而提高算法的综合检测性能。本文对当前主要的人脸检测算法的优缺点进行了简述和分析,实现了肤色检测与级联支持向量新颖检测(SVND)结合的人脸检测方法。首先对基于肤色的人脸检测算法进行了研究。通过对肤色在不同彩色空间中的聚类性分析,选择在H-CbCr彩色分量上进行简单肤色模型的分割,然后在YCbCr彩色空间中建立高斯模型对简单肤色模型的分割区域进行二次分割,并在肤色分割的过程中加入边缘检测,从而将大片连通肤色或类肤色区域进行分离。在对图像进行二值化时,选择改进的最大类间方差(Otsu)阈值分割法,通过滤波、形态学处理,以消除噪声的影响,最终较好地分割出人脸候选区域。然后对基于SVND的人脸检测方法进行了研究和分析。详细介绍了二分类SVM的检测原理、1类SVM检测方法,并实现了级联SVND的人脸检测算法。将样本图像的像素作为特征值进行训练和测试,会导致样本维数过高而增加计算的复杂度,针对这一问题,本文先对样本图像进行主成分分析(PCA)。通过在MIT人脸数据库及真实图像上对级联SVND使用不同特征(像素特征和PCA特征)情况下的分类性能进行了比较,验证了PCA降维处理在人脸检测中的有效性。最后针对肤色检测和基于PCA特征的级联SVND在人脸检测中各自的特点,本文将这两种方法结合起来进行人脸检测,实验验证了本文采用的方法提高了人脸检测的检测率并在一定程度上降低了误检率和检测时间。为了充分利用肤色检测的结果,对级联SVND的检测窗口大小进行调整,进一步减少了检测时间。更多还原

【Abstract】 Recently, the identity authentication technology based on biometric has been developed rapidly. Face detection as an important application of biometric authentication technology has been paid more and more attentions. Face detection plays a very important role and practical value in the visual image processing, human-computer interaction, facial expression recognition, video surveillance, content-based search, etc. Many face detection algorithms have been proposed. Because of the complexity of the face, it is difficult to find an algorithm, which can adapt to all kinds of conditions of detection task. More research is based on a variety of effective combination of the algorithms to improve the detection performance of the algorithms. In this paper, the description and analysis about the advantages and disadvantages in current main face detection algorithms has been given. A face detection method combined with the skin color detection and the novel cascade support vector (SVND) is implemented.First, face detection algorithm based on skin color is studied. Through the analyis about clustering of skin color in different color space, we select the H-CbCr color components of the color model for simple division. Then the Gaussian model is set up in the YCbCr color space of simple skin color model segmentation area for secondary division, and the edge detection is added in the process of color segmentation, thus large areas of communicating color or skin color is separated. During the binarization of image, choose the improved between-cluster variance (Otsu) threshold segmentation method for binary, by filtering, morphological processing, in order to eliminate the influence of noise. Face candidate region is better splited up at last.Then, the face method based on SVND is studied and analyzed. Principle of the two classification support vector machine (SVM) method and one-class SVM detection method are introduced in details. The cascade SVND face detection algorithm is realized. The pixel values of the sample image for training and testing will lead to the increasing of the dimension of the sample and the complexity of the calculation. In order to solve this problem, the principal component analysis (PCA) is applied to sample image firstly in this article. Based on the MIT face database and real images, for a cascade SVND, we give the performance comparisons for different characteristic (Pixel features and PCA features) and verify the PCA dimensionality reduction in the effectiveness of face detection.Finally, for the characteristics of cascade SVND based PCA feature and skin detection in face detection, this paper combines these two methods for face detection. Improvement in the detection rate of face detection and reduction in the false detection rate and detection time are verified through experiments. In order to make full use of the skin color segmentation results, the size of cascade SVND detection window is adjusted and the detection time is reduced further.更多还原