CORC  > 自动化研究所  > 中国科学院自动化研究所  > 毕业生  > 博士学位论文
题名人脸检测和识别若干问题的研究
作者金洪亮
学位类别工学博士
答辩日期2006-01-06
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
导师卢汉清
关键词人脸检测 人脸识别 基于人脸素描图的合成和识别 人脸认证系统 Face Detection Face Recognition Face Sketch Synthesis and Recognition Face Identification System
其他题名Research On Several Issues of Face Detection and Recognition
学位专业模式识别与智能系统
中文摘要人脸分析是生物特征鉴别技术的一个重要方向,与其他生物特征如虹膜、指纹相比,人脸具有主动性、非侵犯性等优点,因而受到研究者的广泛关注。人脸分析包括人脸检测、人脸器官定位、人脸识别和认证、人脸跟踪和人脸配准等主要研究方向。 本文主要针对人脸检测和人脸识别中的若干问题进行了深入的研究,其中涉及到人脸检测模型的建立、人脸局部特征的抽取、Gabor人脸识别算法中的不同尺度和方向下的特征的选择问题、基于多级相似度的人脸识别算法、人脸素描图的重建和基于人脸素描图的识别等内容。本文的主要的工作和贡献有: 1.结合人脸的皮肤色模型,提出了一种基于支持向量机的彩色图像人脸检测算法。旨在削弱人脸检测算法对负样本的依赖性。 2.提出了一种基于局部特征的人脸检测算法,即改进的局部二进制模式(Improved Local Binary Pattern)。该算法的特点有:(1)它是光照鲁棒的,在光照条件较差的情况下也能获得较好的人脸检测效果;(2)计算代价小,不仅使得训练步骤的时间大大缩短,而且可以做到在320*240大小的图像上进行实时检测。 3.针对不同尺度和方向的Gabor特征的不同特点以及对人脸识别性能的贡献不同,本文提出了两种新的基于Gabor特征的人脸识别方法,即:(1)分别用子空间方法对不同尺度和方向的Gabor特征进行分析,然后采用Boosting学习进行融合。(2)对不同尺度和方向的Gabor特征进行核特征描述,结合基于矩阵的核判别分析学习用于识别的判别特征。实验证明了两种方法的有效性。 4.提出了一种非线性的人脸素描图的合成和识别算法。受局部线性内嵌(LLE - Locally Linear Embedding )思想的启发,本文提出了一个基于局部几何保持的人脸素描图合成方法。针对画家画素描图时带来的局部微小扭曲以及伪素描图中的局部模糊特性,本文采用了基于核判别分析的素描图识别算法。 5.开发了一个人脸认证的原型系统平台。本文参与的主要工作包括开发用于大数据库的人脸检测分布式训练系统,人脸检测算法的选择,人眼定位算法以及部分人脸识别的工作。 本文提出的人脸检测和识别算法可以用来建立一个人脸分析系统。
英文摘要Face analysis is an important part in biometrics. Compared with other biologic characters such as iris, fingerprint, face has the virtue of initiative and nonaggressive, so it attracts more attention of the researches. Face analysis includes face detection, facial feature location, face recognition and identification, face tracking, face alignment and etc. In this thesis, we mainly research some issues on face detection and face recognition, including establishment of face detection model, extraction of local facial feature, Gabor feature selection of different scales and directions, multi-similarity based face recognition approach, face sketch synthesis and recognition. The main contributions of this thesis include: 1. Combining the skin color model, we propose a one class SVM based algorithm for face detection, which aims at weakening the dependance on non-face examples when training a detection model. 2.Propose a face detection approach based on local features, i.e. Improved Local Binary Patterns($ILBP$). The new features have the virtue of: (1)Illumination-robust, good performance will be also obtained in bad lighting condition. (2) The computation cost of the binary pattern is very low, so the training time is shorten greatly. In an image with size 320*240, we can do real-time face detection. 3.Gabor features of different scales and directions have different characteristics and make different contributions to recognition. In this thesis, we bring out two Gabor-based face recognition framework, i.e.: (1)We analysis Gabor features of different scales and directions, and use Boosting learning to merge them. (2) Using kernel methods to describe these Gabor features, and learning the discriminate features using matrix based kernel discriminant analysis. The experiments prove the effectiveness of our approaches. 4. We propose a nonlinear face sketch synthesis and recognition approaches. Inspired by Locally Linear Embedding (LLE), we develop a local geometric preserving based face sketch synthesis appraoch. Aiming at the distortion brought by the artist and the local blurs in pseudo-sketch, we use kernel discriminant analysis approach for face sketch recognition. 5.Establish a face identification proto-type system, including the distributing training system of face detection model, the selection of face detection algorithms, eye location approaches and part of the face recognition work. The algorithms brought in this thesis can be used in a face analysis system
语种中文
其他标识符200218014603208
内容类型学位论文
源URL[http://ir.ia.ac.cn/handle/173211/5887]  
专题毕业生_博士学位论文
推荐引用方式
GB/T 7714
金洪亮. 人脸检测和识别若干问题的研究[D]. 中国科学院自动化研究所. 中国科学院研究生院. 2006.
个性服务
查看访问统计
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

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。


©版权所有 ©2017 CSpace - Powered by CSpace