奇异值分解的人脸识别算法

【作者】 赵慧琳；

【导师】 王林泉；

【作者基本信息】 上海海运学院 ， 计算机应用技术， 2002， 硕士

【摘要】 随着信息技术的发展，社会管理日趋电子化，自动化。在这样一个庞大的社会网络体系中，系统安全十分重要。准确地鉴别个人身份是各安全系统的必要前提。人脸是区别人的重要器官，利用人脸特征进行身份验证又是最自然直接的手段，相比其它人体生物特征它具有直接、友好、方便的特点，易于为用户所接受。因而在公安(罪犯识别等)、安全验证系统、信用卡验证、医学、档案管理、视频会议、人机交互系统等方面有着巨大的应用前景，人脸识别越来越成为当前模式识别和人工智能领域的一个研究热点。 人脸识别的研究涉及模式识别、图像处理、生理学、心理学、认知科学等领域，早在二十世纪六七十年代即引起了研究者的强烈兴趣。人脸识别可以分为：人脸的检测与定位、特征提取和识别三部分。 本文中在人脸的检测与定位部分，首先根据人脸图像的边缘检测图中的人脸内轮廓线进行垂直定位。然后在垂直定位后的灰度图中根据人眼的水平结构特征进行人脸以及人眼的水平定位。接着根据人眼的水平位置剪切人眼部分的二值化图像对它进行垂直投影，定位出人眼的精确位置；根据人眼信息得到标准化的图像。实验证明本方法正确定位率可达88.6％。 在特征提取部分，采用的是基于奇异值分解的特征提取办法，将人脸图像矩阵的奇异值作为特征向量。 在识别部分，构造了一个层次的距离分类器进行人脸的识别。设Y为待识图像，首先求Y的特征向量记为Y，计算第i类样本图像特征向量的中心记为X~i，求出这两个向量的距离。根据计算出的距离进行判定，我们就把待识图像归为距离最小的那一类。实验证明该方法可以取得76％的识别率。更多还原

【Abstract】 With the development of information technology, social management tends to be electricized and automated. In such a large social network system, system security is of vital importance with the premise to make out personal identification. As face is an important part for identification, it is most direct and natural to identify a person by means of face features. This kind of way is characterized by convenience and friendliness compared with other means and will be easily accepted by without stirring up defense. Naturally, face identification has bright application future in the field of criminal distinguishing, safety check system, credit card identification, medicine, file management, video meeting, people-machine exchange system, and will become a hotspot in the recent research of pattern recognition and artificial intelligence.Face Recognition(FR), having aroused great interest of the researchers in the early 1960’s, involves pattern recognition, image process, physiology, psychology, recognition science and so on. It can be separated into three parts: face detection and orientation, feature extraction and recognition.In the part of face detection and orientation, vertical orientation will be made according to inner outline within edge extraction picture of the face image. After that, according to the horizontal level features of human eye, face and human eyes will be horizontal orientated in the gray image. Then, the binary image of the eye will be cut in accordance with the eye horizontal level, which will be vertically projected to orientate the accurate position of human eyes. The human eyes will in turn accurately position the face to further get standard image. Experiments prove the accuracy reaches to 88.6%.Face feature extraction will apply the method based on singular value decomposition which takes singular value decomposition of the matrix as feature vector.A distance classifier has been built up for face recognition. If Y is an image to be recognized, the feature vector Y will first be worked out, and the feature vector center of i sample class will be noted as Xt. Then the distance between those two vectors can be calculated for judgements. We classify the image into the sample class that has the shortest distance. Such kind of method has proved 76% accuracy.更多还原