基于EMD的融合特征快速多示例人脸识别算法  

Face Recognition Based on Improved Fast EMD-MIL Framework and Information Fusion

作　　者：邓剑勋[1,2] 熊忠阳[1,2] 曾代敏[1,2] 

机构地区：[1]重庆大学计算机学院,重庆400044 [2]重庆大学数理学院,重庆400044

出　　处：《四川大学学报（工程科学版）》2012年第2期99-104,共6页Journal of Sichuan University (Engineering Science Edition)

基　　金：中国博士后科学基金资助项目(20070420711);中央高校基本科研业务费科研专项-研究生科技创新基金(CDJXS11180001);中央高校基本科研业务费科研专项-自然科学类项目资助(CDJZR10100023);重庆市科委自然科学基金计划资助项目(2007BB2372)

摘　　要：在基于五官模版技术的人脸识别中,因光照、角度及缺少整体性考虑等情况,易产生异常示例,影响了算法精度。而适合多示例检索的EMD距离寻优路径较长,导致在人脸识别中应用受限。为此提出一种基于EMD距离的快速融合特征多示例人脸识别算法(IIFEMD-MIL)。针对异常点的问题,通过引入结合整体特征的融合多示例技术以及距离阀值,从而减少异常示例的产生并对超过阈值的示例予以平滑处理;针对寻优路径长的问题,将人脸五官之三结合整体示例为模版构建四示例的一一匹配,并进一步提出了融合快速EMD-MIL框架,缩短了寻优遍历路径。在ORL和MIT图像集上进行的比对实验表明,该算法执行效率和分类准确性优于其他同类算法。In order to solve the problems that a number of factors,such as brightness,shooting angle and the lack consideration of entirety,have a negative effect on many face recognition algorithms which were based on the facial features,and earth mover＇s distance（EMD）,which was fit for multi-instance learning（MIL）,runs slowly on large data sets,a fast face recognition algorithm was proposed.The algorithm was based on EMD-MIL framework and information fusion technique.Firstly,both information fusion MIL framework and distance threshold were used to make algorithm perform well.Secondly,relying on instances which be composed of the three facial features and the overall character,a fast EMD-MIL framework was bring out.The fast EMD-MIL framework runs more quickly than the old edition.Experimental results on the ORL and the MIT showed that this algorithm is feasible and the performance is superior to some other similar algorithms.

关 键 词：多示例学习(MIL) 人脸识别 推土机距离(EMD) 距离阈值 融合特征 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]