【作者】 王杨峰；

【导师】 付永庆；

【作者基本信息】 哈尔滨工程大学 ， 通信与信息系统， 2008， 硕士

【摘要】 人脸识别是当前人工智能和模式识别领域的一个研究热点。特征提取是人脸识别的关键,直接影响识别结果。本文主要对特征提取算法进行研究。一维方法特征提取时运算量大,特别是图像很大时。二维方法特征提取直接,速度快,但提取出的特征是矩阵,特征数量大,影响分类速度。结合两者的优点,提出一种基于2DPCA(two-Dimensional Principal ComponentsAnalysis)和DLDA(Direct Linear Discriminant Analysis)的人脸识别算法,先用2DPCA处理原始图像,降维后进行DLDA处理,并深入分析了特征向量参数选取的问题。实验结果表明,识别率和分类速度均有很大提高。传统特征脸方法需要大量人脸样本且当人脸样本增加时需要重新训练“特征脸”,不适合人脸样本较少和实时性要求比较高的场合。针对这一问题,提出一种基于全局和局部特征的LBP(Local Binary Pattern)人脸识别算法。该方法无须训练“特征脸”,直接抽取人脸图像的局部和全局直方图特征作为图像特征进行识别。实验结果表明,该方法具有较好的实时性和识别率,且对光照不敏感。本文还研究了单样本人脸识别问题。专门研究了增加训练样本和利用单样本进行人脸识别的方法。在ORL人脸库中验证了这些方法的有效性。更多还原

【Abstract】 Face recognition is one of active research topics in the fields of pattern recognition and artificial intelligence. And feature extraction is the key of human face recognition, which directly influences the recognition result. This paper takes a research on the feature extraction algorithm.Computational amount is large when using the one-dimensional method of feature extraction, especially for large images. The two-dimensional method extracts feature directly and rapidly, but the features must be expressed with matrixes. Therefore the classification speed is affected by too many features. Hence, by combining their virtues, a new approach is proposed to recognize human faces based on 2DPCA (two-Dimensional Principal Components Analysis) and DLDA (Direct Linear Discriminant Analysis). First, 2DPCA is used to deal with original images, and then DLDA is used to compress the feature. Also the parameter selection of eigenvector is analyzed deeply in this paper. The results show that both recognition rate and classification speed are greatly improved.The traditional eigenfaces methods demand large face samples, and eigenfaces have to be retrained when additional face samples are added. Thus, it is not suitable for face recognition with requirements of fewer samples and high real time ability. For this problem, a LBP face recognition method based on combination features of global and local features is presented. Without re-training the eigenfaces, it directly extracts the face histogram features of global and local as feature for recognition. The experimental results indicate that the method has a higher recognition rate, better real time character and it was not sensitive to illumination condition.The single training image per person is also discussed in this paper. Both the methods of increasing training samples and using single training image for face recognition are especially researched. The feasibility of those methods is verified in ORL face database.更多还原