基于MB-LBP和改进的LFDA的人脸识别  

Face Recognition Based on MB-LBP and Improved LFDA

作　　者：齐鸣鸣[1,2] 向阳[1] 

机构地区：[1]同济大学计算机科学与技术系，上海201804 [2]绍兴文理学院元培学院，绍兴312000

出　　处：《计算机科学》2012年第6期266-269,共4页Computer Science

基　　金：国家自然科学基金项目(70771077);浙江省教育厅科研项目(Y201122544);绍兴文理学院元培学院科研项目(090610)资助

摘　　要：提出了一种基于多块LBP(Multi-scale Block Local Binary Patterns,MB-LBP)和改进的局部化的Fisher判别分析(Local Fisher Discriminant Analysis,LFDA)的人脸识别算法。该算法利用MB-LBP的局部和整体描述能力强化了标注样本的局部分析和训练样本的全局分析;以每个样本与同类其他样本的欧氏距离均值作为参数,克服了类内散度计算限制;通过参数融合训练样本的总散度信息保持样本的全局结构。实验表明,MB-LBP为局部保持分析和全局保持分析提供了良好的基础;在少量标注样本情况下,改进的LFDA的适应性和识别率明显优于LFDA。A algorithm on face recognition based on multi-scale block local binary patterns（MB-LBP） and improved Local fisher discriminant analysis（LFDA） was proposed,which strengthens local analysis for labeled samples and global analysis for training samples with the ability of MB-LBP for local and global description.The algorithm makes use of ave-rage Euclidean distance from every sample to other samples in the same class as the parameter to overcome the limit for computing within-class scatter and preserves global structure by inosculating the total scatter of training samples in the form of parameter.Experimental results show that MB-LBP provides good base for making analysis of local and global preserving,and improved LFDA has more obvious adaptability and recognition rate than LFDA when the number of labeled samples is small.

关 键 词：多块LBP 局部Fisher判别分析 全局结构 人脸识别 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]