Semi-supervised LDA Based Method for Similarity Distance Metric Learning.

In recent years, computer vision technology has drawn much attention of people and been applied into many fields of human's living. Data classification/identification is a key task in computer vision. The similarity distance metric learning based method is wildly used to compare the similar positive pairs from dissimilar negative pairs. However, there are more and more challenging computer vision task have been proposed. Traditional similarity distance metric learning methods are fail to metric the similarity of these task due to the drastic variation of feature caused by illumination, view angle, pose and background changes. Thus, the existing methods are unable to learn effective and complete patterns to describe the appearance change of individuals. To overcome this problem, we proposed a novel semi-supervised (Linear Discriminant Analysis) LDA based method for similarity distance metric learning. The proposed method first learn a metric projection with traditional LDA method. The then test data are identified with the potential positive pairs to fine-turning the metric model by forcing the identified data to be close to the center of positive training data pairs. Finally, the proposed method are compared to some classic metric learning algorithms to demonstrate its effectiveness and accuracy.