融合距离度量学习和SVM的图像匹配算法

Fusion Method of Distance Metric Learning and SVM for Image Matching

目前度量学习方法通过有限样本数据学习得到新度量后,采用简单的分类器(如直接欧式距离计算)通常不能达到最佳分类效果.SVM作为一种经典的分类器,具有优秀的线性和非线性分类能力,可以弥补距离度量学习方法的不足.对此,提出一种应用于图像匹配的融合距离度量学习和SVM的(DML-SVM)算法.首先,利用度量学习方法得到的线性变换矩阵,将样本变换到新的特征空间,降低特征各维度之间的相关性,调整特征各维度的权重;然后通过SVM对新特征进行线性或非线性分类.通过在LFW,Pubfig,Toycars三个图像数据库上的测试结果表明:融合方法的分类能力优于度量学习和SVM算法各自单独使用时的性能,且融合算法对训练样本数量具有很强的鲁棒性,即使只有少量训练样本(180个)时,融合算法仍然能具有较高的分类能力.

Generally,the distance metric learning method trained with limited samples can not achieve the best classification perform- ance for the image matching by using a simple classification （ such as Euclidean distance computing }. SVM that is a classic classifier with excellent linear and nonlinear classification ability can make up for the lack of distance metric ,learning methods. This paper pres- / / ents a fusion method of distance metric learning and SVM （ DML-SVM } for image matching. First, the feature extracted from images is transformed into a new feature space by the transformation matrix obtained from DML, which reduces the correlation between each dimension of feature ~ and optimizes the weight of each dimensions. Then the linear or non-linear SVM classifier is used to classify samples. The experimental results on the LFW, Pubfig and ToyCars databases show that the fusion method improve the recognition ac- curacy compared to single distance metric learning method or SVM, and the fusion method is robust to the number of training samples. Even if using only a small amount of training samples （180） ,the fusion method still has good classification performance.