融合SIFT和MIL的红外人脸识别方法

Infrared face recognition method by integration of SIFT and MIL

针对红外人脸识别问题,提出一种新的基于尺度不变特征转换(SIFT)与多示例学习(MIL)相结合的算法。该算法将图像当作多示例包,SIFT描述子当作包中的示例,利用聚类的方法对训练集中的所有SIFT描述子进行聚类,建立"视觉词汇表",再根据"视觉字"在多示例训练包中出现的频率,建立"词-文档"矩阵,采用潜在语义分析(LSA)的方法获得多示例包(图像)的潜在语义特征,将MIL问题转化成标准的有监督学习问题,即在潜在语义空间用支持向量机(SVM)求解MIL问题。基于OTCBVS标准数据集的对比实验结果表明,所提算法是可行的,且识别率明显高于其他方法。

For the problem of infrared image face recognition, a novel algorithm based on SIFT feature and multi-instance learning （MIL） algorithm is proposed. Firstly, this algorithm regards image as a bag, and SIFT descriptor of the key points as instance. Then all the SIFT descriptors in the training set have been clustered by K-Means method, and regards cluster centers as ＂visual word＂ to build ＂visual vocabulary table＂; Secondly, according to the frequency of ＂visual word＂ in the training bag to establish a ＂word-document＂ matrix, then latent semantic analysis （LSA） method is used to obtain bag＇ s （image） latent semantic features, converts MIL problem to a standard supervised learning problem, which means to solve MIL problem use SVM in the latent semantic space. Experimental results on the OTBCVS image set show that the algorithm proposed is feasible, and the performance is superior to other algorithms.