阈值控制下的融合股权多示例人脸鉴别技术  

Equity Multi-Instance Face Identification Based on Threshold Control and Fusion Feature

作　　者：邓剑勋[1] 熊忠阳[1] 曾代敏[2] 

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

出　　处：《模式识别与人工智能》2013年第1期14-19,共6页Pattern Recognition and Artificial Intelligence

基　　金：中国博士后科学基金项目(No.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项目(No.CDJXS11180001);中央高校基本科研业务费科研专项-自然科学类项目(No.CDJZR10100023);重庆市科委自然科学基金项目(No.2007BB2372)资助

摘　　要：多示例学习在区域图像检索中取得较好效果.其一票通过制在人脸鉴别中易导致误判,因五官之一相似,甚至都相似,两幅人脸仍可能不同.为适应特殊场景,提出股权多示例学习概念,某示例类在实验库中有不同股权,训练集特性可近似代表实验库特性;不同类示例的判别结果按示例类股权配比后,形成包的类别归属.其次引入整体特性作为特殊示例进行特征融合,引入整体示例股权阈值控制配比,防止五官类似而整体不同的情况;通过股权阈值选优提升识别率.在ORL和FERET图像集上进行的对比实验表明,该算法分类准确性优于传统算法.Image retrieval based on multi-instance learning （MIL）has great value in the field of regional image retrieval. The traditional voting mechanism in MIL is prone to misunderstanding, because the local similarity does not mean the overall similarity in face identification. Firstly, instance equity concepts of equity MIL are presented. Each kind of instance has different equity and the training set has the similar features to the test set. Therefore, the classification attribution for different packets can be obtained by the sum of results from multiplying every discriminant result and its instance equity. Secondly, the overall characteristic is considered as a special instance, and the overall sample equity threshold is used to control equity ratio. At the same time, the abnormal conditions, such as two persons have similar facial features, are prevented by means of the feature fusion. And the recognition rate is improved by the use of threshold control. The experimental results on the ORL and FERET show that the algorithm is feasibleand the performance is superior to other algorithms.

关 键 词：人脸鉴别 股权多示例学习 示例股权 特征融合 阈值控制 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]