基于人脸特征的身份识别

【作者】 高涛；

【导师】 何明一；

【作者基本信息】 西北工业大学 ， 生物医学工程， 2006， 硕士

【摘要】 生物特征是人的内在属性,具有很强的自身稳定性和个体差异性,因此是身份验证的理想依据。其中,利用人脸特征又是最自然直接的手段,相比其他生物特征,它具有直接、友好、方便的特点。人脸识别技术是计算机视觉与模式识别领域的研究热点,在身份验证、电视会议、人机界面、可视通信、公安档案管理、安全部门、基于内容的图像检索等很多方面有着广泛的应用。人脸识别过程一般可以分为人脸的检测和预处理、特征提取、匹配识别三个部分。本文对人脸的检测和识别技术进行了系统、详细的研究。 在人脸图像的预处理阶段,本文主要完成对人脸样本的图像增强、几何及灰度归一化工作。这些工作有效地改善了图像质量、降低了计算复杂度,从而提高了后续算法的收敛速度,并且使用了二维小波变换压缩数据。二维的小波分解具有对表情变化不敏感的特点,将图像的大部分能量集中到低频的子图像,高频部分则对应于图像的边缘和轮廓,可以很好地压缩和表征人脸图像的特征。 在人脸特征提取阶段,文中详细介绍了人脸图像几何特征和统计特征提取的方法,例如几何特征提取法,奇异值分解方法,线性鉴别分析,离散余弦变换,基于小波变换特征提取等方法。重点介绍了本论文所使用的主分量分析和定点独立分量分析的方法,定点独立分量分析是一种基于高阶统计信息的特征提取方法,收敛速度快克服了一般ICA收敛慢的缺点。 在人脸图像的识别阶段,文中详细介绍了最小距离法和支持向量基等识别方法。重点介绍了本论文所使用的神经网络的识别算法。径向基函数(RadialBasis Function,RBF)理论为多层前馈网络的学习提供了一种新颖有效的手段。RBF神经网络是一种性能良好的前向神经网络模型,RBF网络不仅具有良好的推广能力,而且计算量小,速度快。 最后,本文对ORL人脸库和YEL人脸库进行仿真,识别率能达到98%以上,明显比传统的方法提高了识别速度和识别率。更多还原

【Abstract】 The biological characteristics are the inherent attributes of human beings, which have strong self-stability and individual independency. They are the ideal source of information for ID verification. Compared with other biological characters, the patterns of face are natural and direct;therefore the technique of face verification is widely used as the main approach in ID verification. Human face recognition is one of the most active and challenging tasks for computer vision and pattern recognition. It can be widely applied in diverse fields such as personal identification, human-computer interface, teleconference, visual communication, criminal archive administration, security departments, content-based image retrieval, and etc. The process of recognition can be divided into three stages: face detection and preprocessing of face image, feature extraction, and recognition.The preprocessing mainly focuses on image enforcement, geometry normalization of face images, image intensity normalization. These will improve image quality, decrease computation complexity and therefore accelerate the convergence of algorithm. Because of its insensitivity to expression changes and ability to compress and express the characteristics of the face image, wavelet analysis is employed to compress the data of face image.In the feature extraction stage, methods of extracting geometrical and statistics are introduced in this paper, which include geometrical extracting method, singular value decomposition, linear discriminated analysis (LDA), discrete cosines transform, feature extraction based on wavelet analysis. And the method of fixed point ICA is proposed and discussed with emphasis, which is capable of acquiring the characteristics of face image and is much faster than normal ICA.In the process of image recognition, the classic methods are particular introduced, including nearest neighbor classifier, support vector machine (SVM), and etc. The high generalization ability and low computation cost of RBF makes it an excellent classifier designed for samples with large quantity.Experimental results on ORL and YEL face database denote that the newly proposed algorithm can increase the recognition accuracy to above 98% with low-computation cost. This indicates that the newly proposed algorithm is much more effective compared with some classical method.更多还原