节点文献
彩色图像序列的人脸检测、跟踪与识别研究
Study on Face Detection, Tracking and Recognition in Color Image Sequences
【作者】 夏思宇;
【导师】 夏良正;
【作者基本信息】 东南大学 , 模式识别与智能系统, 2006, 博士
【摘要】 人脸在社会交往中扮演着十分重要的角色,是人类在确定一个人身份时所采用的最普通的生物特征,研究人脸识别及其相关技术具有十分重要的理论和应用价值。彩色图像序列的人脸检测、跟踪与识别技术是随着计算机技术的高速发展和视频监控等应用的需要在近几年才逐渐成为一个研究热点。与基于静态图像的技术相比,彩色图像序列提供了更加丰富的信息,比如颜色信息、运动信息等等,但是彩色图像序列中的人脸检测、跟踪与识别有着更高的要求,如能够适应成像条件、光线和复杂场景变化,对图像中的人脸的姿态、遮挡、时间变化等能进行有效的处理,而且应具有较高的计算效率等等。因此,对彩色图像序列的人脸检测、跟踪与识别的研究还是一个任重而道远的研究课题。本论文针对彩色图像序列这个范围,就人脸检测、跟踪与识别技术中的核心技术与关键问题展开研究工作。其主要研究内容与创新点包括:1.提出了一种彩色图像增强改进算法。主要从人眼对物体颜色的感知特性出发,在Retinex理论与算法的基础上,针对迭代参数的选取作了改进。改进的算法基于采用一种迭代截止条件,对参数进行自适应选取,避免了人工设定,并同时减少了整个算法的运算量。实验结果表明,改进后的算法对彩色图像的颜色、亮度、对比度处理的结果符合人眼视觉系统的感知特性,且改善了图像的偏色情况。应用于彩色人脸检测的预处理后,人脸正确检测率得到提高。2.依据人眼视觉特性,提出一种“由粗至精”的人脸肤色区域检测方法。首先从彩色序列图像中提取运动目标区域,以剔除无关背景,然后在运动目标区域中检测肤色特征(粗检测),接着结合运动目标的边缘特征,对肤色区域进行形态学处理(精检测),最后检测得到人脸肤色区域。3.提出一种Self-Skin肤色检测算法。这种算法抛弃了传统的采用事先通过大量肤色样本统计得到的肤色模型进行肤色检测的思路,而是针对单幅图像中的肤色分布,在色度空间中进行区域分割,同时结合了肤色统计信息,有效的克服传统肤色模型方法中的“过检测”问题。实验表明,Self-Skin肤色检测算法鲁棒性好,能应用于复杂背景,且对光线变化不敏感。4.提出了一种彩色图像序列的人脸检测与跟踪方法。该方法将人脸检测与人脸跟踪这两个步骤有效的结合在一起,而不是分成两个相对独立的部分。其思想是先通过肤色检测算法得到候选人脸区域,然后通过Condensation滤波跟踪算法对候选人脸区域进行跟踪,在跟踪过程中提出了一种基于
【Abstract】 Human face is our primary focus of attention in social intercourse, playing a major role in conveying identity and emotion. Research on the face recognition technology has great theoretical and practical values. The study on face detection, tracking and recognition is becoming an active research subject with the development of computer technology in recent years. Compared with still image, color image sequence provides more information, such as color information, motion information, and so on. However, it should be more robust to different illumination conditions, complex background, face occlusion, and so on. Additionally, it demands lower computational cost. The study on face detection, tracking and recognition in color image sequence is still a challenging task.This thesis focuses only on the key technologies of face detection, tracking and recognition in color image sequence. The main contributions of this thesis are as follows:1. An improved color image enhancement algorithm is presented. Based on Retinex theory, this improved algorithm can automatically determine values for the parameter, instead of looking for fixed value, and decrease the amount of computation. The results on color images show the validity of this improved algorithm. After applying this algorithm to preprocessing in face detection, the detection rate increases slightly.2. A coarse-to-fine approach to face skin detection in color image sequence is proposed. At first, motion regions are extracted from color image sequence in order to discard background region. Once the regions of interest are located, skin color detection is used to get skin color regions. Then edge information and mathematical morphology method are integrated to progressively restrict the regions to smaller areas, as candidate face skin regions.3. A novel self-skin algorithm is proposed for skin detection. The idea of this algorithm is using the watershed method to cluster the pixels in the YCbCr color space, instead of applying conventional skin color model. The experimental results show that this algorithm is robust to different illumination conditions and complex background, and its reliability excels
【Key words】 Face Detection; Face Tracking; Face Recognition; Color Image Enhancement; Skin Color; Support Vector Machine; Contourlet Transform;