An Efficient Fusion Method of Distance Metric Learning and Random Forests Distance for Image Verification

Generally, the distance metric learning method using sample classifier (such as Euclidean distance computing) can not achieve perfect classification performance for the image verification. Nevertheless, the random forests distance method (RFD) can overcome the shortcoming of distance metric learning since it can handle the heterogeneous data well. In addition, the distance metric learning method can reduce the training time of RFD because it can remove the data correlation. Therefore this paper proposes a fusion method of distance metric learning and random forests distance. We obtain a matrix M using the distance metric learning method and use it to linearly transform the sample space, then we classify new samples by RFD. We experiment on LFW, Pubfig and ToyCars datasets and the results show that our proposed fusion method outperforms the single distance metric learning method or RFD in the recognition accuracy; the training time of RFD is much less in the transformed sample space.