基于仿生模式的人脸识别  

Face recognition on bionic pattern

作　　者：琚生根[1] 周激流[1] 何坤[1] 王刚[1] 

机构地区：[1]四川大学计算机学院,成都610065

出　　处：《四川大学学报（自然科学版）》2008年第1期65-70,共6页Journal of Sichuan University(Natural Science Edition)

基　　金：国家自然科学基金(60272095)

摘　　要：我们提出了基于"认识"事物的人脸识别.以同类样本全体的连续性规律为基本出发点把人脸识别看成人脸"认识"的问题而不是人脸分类划分,与以"最佳划分"为目标的传统统计模式识别相比,本文的方法更接近于人类"认识"事物的特性.同一人脸的分布在低维空间中具有一定的类聚性而不同人脸的覆盖范围相互交织.随着空间维数增加,同类样本内聚性减弱,不同类样本互斥性增强;空间维数继续增加,同类样本内聚性与不同类样本互斥性反而都会减弱.将待识别人脸在某一空间进行覆盖范围的识别.若属于多个侯选人脸覆盖范围,进而运用Fisher人脸识别出最终结果.对ORL人脸库实验表明,对于未经过训练的任意对象不会误识,识别率为97.5%.A ＂matter cognition＂ based face recognition model has been proposed. By taking continuity rule of samples of a same class as the starting point, face pattern recognition is considered as face pattern cognition instead of its classification. Compared with traditional best classification goaled statistic pattern recognition, it＇s more similar to the character of human cognition. A person＇s face distribution in low dimension space has a certain kind of cohesion, while face coverage of different people overlap. By the increase of space dimension, the cohesion of samples of a same class decrease, while the repel of samples of different classes increases. But as the increase of space dimension continue, both the cohesion of samples of a same class and the repel of samples of different classes decreases. Coverage of candidate faces recognition is processed in a certain space. If it belongs to several candidate face coverage, Fisher method can be applied to get the final result. Experiment based on ORL proves that, random object without training can be perfectly recognized.

关 键 词：模式识别 仿生学 连续性规律 多项式核函数 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]