【作者】 孙顺利；

【导师】 吴国伟；

【作者基本信息】 大连理工大学 ， 软件工程， 2007， 硕士

【摘要】 人脸自动识别技术就是利用计算机分析人脸图像，从中提取有效的识别信息，用来辨认身份的一门技术。它使用了人体本身所固有的生物特征，是与传统方法完全不同的崭新技术，具有更好的安全性、可靠性和有效性，越来越受到人们的重视。人脸识别主要有三个部分组成：图像预处理，特征提取和分类识别。本文的图像预处理主要采用了直方图修正，并利用小波变换来降低图像的维数。特征提取方面使用PCA方法进行特征提取，最后使用距离分类法进行图像识别。主成分分析(PCA)方法作为最成功的线性鉴别方法之一，目前仍然被广泛应用于人脸等图像识别领域，但传统的PCA方法由于提取的是图像的全局特征，因此受光照条件和人脸表情变化影响比较大，造成识别效果不是太好。经过分析发现，当人脸表情和光照条件变化时，一般仅有部分人脸区域变化明显，而其它部分变化不是太大，甚至没有变化。基于这种情况，本论文提出了分块PCA方法，在对图像进行特征提取前，首先对图像进行划分，然后再对子图像使用PCA方法提取特征向量，在特征提取时根据各个子图像在整个图像中的重要性不同，分别赋予不同的权值，最后构造分类器，进行识别。在使用PCA方法进行特征提取时，由于要先将图像转化为一维向量，这样会使图像的向量维数过高，对以后的特征提取造成困难。在此基础上，本论文进一步提出了分块2DPCA方法，与分块PCA方法不同的是，在对原图像进行划分以后，对子图像使用2DPCA方法进特特征提取，2DPCA方法由于直接利用原图像的二维矩阵构造图像的散布矩阵。使得总体散布矩阵的维数远远低于利用PCA方法得到的协方差矩阵。这样可以很大程度上提高特征提取的速度并降低特征提取的复杂度，进一步提高特征提取的准确性，使得整体识别速度提高并可以在一定程度上提高识别率。最后，在ORL人脸库上进行了实验，验证了改进的PCA算法优于传统的PCA算法。更多还原

【Abstract】 Automatic face recognition technology is analyzes the person face picture using the computer, extract the effective identification information, uses for identify the status of a technology. It is a totally brand-new technique different from traditional methods because it adopts the inherent organism’s characteristics of human body. More and more people pay great attention to it as it is safer, more reliable and effective. Face recognition consists of three parts: Preprocessing, feature extraction and classification. In this thesis, we use histogram equalization to modify the picture and use wavelet transform to reduce the dimension of picture in the processing. Then use PCA extract feature and finally use distance classification to get the result of recognition.As the most successful method of linear differential, principal component analysis (PCA) method is widely used in identify areas, such as face recognition. But traditional PCA is influenced by the light conditions and facial expression changed because it extract the global feature of the image, so the recognition effect is not very good. We know when expression and light conditions changing, only some face region obvious changed, and the other changed is not too much, even no changed. So based on this situation, this thesis presents a method of block-PCA. In this method, we first segment the picture and then extract eigenvector using the PCA. In extracting feature, for each of the parts in accordance with the different role in the overall image give the different weights. In the finally, we construct classification and get the result of recognition. When we use PCA method to extract feature, we should convert the image to one-dimensional vector, so it make the vector dimension of image too much high and induce difficulty to extract feature. Because of this, this thesis presents another method, it is block-2DPCA. After we segmented the image, we use 2DPCA to extract feature in the sub-image space. The matrix dimension of 2DPCA is far below dimension of the covariance matrix using PCA, because it uses the two-dimensional matrix of original image directly to make the covariance matrix. So is can be greatly improved the speed and accuracy, reduce the complexity of feature extraction. In the finally, it improved the speed of recognition and the recognition rate.In the finally, the experimental result show the improved method of PCA is superior to method traditional PCA.更多还原