中国科学院机构知识库网格
Chinese Academy of Sciences Institutional Repositories Grid
基于Boosting学习的自动人脸识别算法研究

文献类型:学位论文

作者黄向生
学位类别工学博士
答辩日期2005-04-23
授予单位中国科学院研究生院
授予地点中国科学院自动化研究所
导师王阳生
关键词自动人脸识别 统计学习 Boosting Jensen-Shannon 散度 Gabor 小波 Automatic face recognition statistical learning Boosting
其他题名Automatic Face Recognition on Boosting Learning
学位专业模式识别与智能系统
中文摘要统计学习方法,特别是 Boosting 学习,在计算机视觉中得到广泛的应用。本文研究基于 Boosting 自动人脸识别算法。在理论上,本文提出一种改进的Boosting 学习算法;在实践上,本文将 Boosting 学习方法应用到人脸检测、人脸对齐、对齐评估和人脸识别中。主要贡献是: 首先,本文提出一种基于理论分析的特征选择准则的 Boosting 学习算法,称为 Jensen-Shannon (JS) Boost。采用相对熵(Relative Entropy)作为弱分类器性能评价的目标函数。我们推导表明,在 Boosting 每轮弱分类器选择学习中,使得 Jensen-Shannon(JS)散度最小的弱分类器正好是使得相对熵最小的弱分类器。由于 JS 散度比其它测度(例如 Kullback-Leibler 散度)计算稳定,所以 JS 散度能为两类之间提供一种更合适的相似性度量。因此,我们将 JS 散度与 Boosting 学习过程结合学习出更优化的弱分类器。将 JSBoost学习算法应用于人脸识别中。实验结果表明,JSBoost 能比其他 Boost 变种(例如 RealBoost,GentleBoost,KLBoost)得到更好的识别效果。 其次,Viola 和 Jones 提出一种称为 Haar-like 的特征,并将其应用到人脸检测中。后来有不少研究人员对 Haar-like 特征进行扩展,企图描述带有方向信息的特征。然而,这些方法始终不能很好地描述人脸的形状和边缘信息。obel 算子是一种简单有效的边缘描述算子。但 Sobel 算子本身对噪声比较敏感。我们引入一种称为 Sobel-like 特征。在特征构造过程中,我们采用图像块的平均值,而不是像素值,与 Sobel 算子进行卷积。然后,用 Boosting 学习这种 Sobel-like 特征来训练人脸检测模型。
英文摘要In this paper, I focus on automatic face recognition based on boosting learning. Statistical learning is widely used in computer vision, especially boosting. An improved boosting algorithm is proposed theoretically. Boosting learning algorithm is applied to face detection, face alignment, face alignment evaluation and face recognition with different feature space in practice. Its main contributions are: Firstly, a theoretically justified AdaBoost learning algorithm is proposed termed Jensen-Shannon (JS)-Boost. Relative entropy loss is used as objective function for learning weak classifiers. We show that for each iteration of AdaBoost learning, a weak classifier that maximizes JS divergence is the one that minimizes relative entropy. JS divergence provides more appropriate measure of dissimilarity between two classes because it is numerically more stable than other measures such as Kullback-Leibler (KL) divergence. We thus derive an optimal weak learner for AdaBoost, hence the JSBoost learning algorithm. JSBoost learning is demonstrated with an application in face recognition. Results show that classifiers learned using JSBoost produces better results than other AdaBoost variants such as RealBoost, GentleBoost and KLBoost. Secondly, Haar-like feature is proposed by Viola and Jones and extended by other researchers for face detection. But it is hard for Haar-like to describe the shape and edge information of face suitably. Sobel operator is a simple and useful tool to represent edge information. However, it is sensitive to noise. An efficient feature, namely Sobel-like feature, is introduced for face detection. During Sobel-like feature construction, blocks of image, instead of pixels, are convolved with Sobel operator. Then, Sobel-like feature is used to train a good model and test in face detection.
语种中文
其他标识符200218014603206
源URL[http://ir.ia.ac.cn/handle/173211/5846]  
专题毕业生_博士学位论文
推荐引用方式
GB/T 7714
黄向生. 基于Boosting学习的自动人脸识别算法研究[D]. 中国科学院自动化研究所. 中国科学院研究生院. 2005.

入库方式: OAI收割

来源:自动化研究所

浏览0
下载0
收藏0
其他版本

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。