| Many machine learning and pattern recognition algorithms rely on a distance metric. Instead of choosing a metric manually, a more promising approach is to learn the metric from data automatically. Besides some early work on metric learning for classification, more and more efforts have been devoted in recent years to learning a distance metric under the semi-supervised learning setting. Semi-supervised learning is a learning paradigm between the supervised and unsupervised learning extremes. Algorithms of this class usually solve the classification or clustering problems with the aid of additional background knowledge. While there has been a whole set of interesting ideas on how to learn from data with supervisory information, we focus our study on semi-supervised learning in the metric learning context.; In this thesis, we propose a series of novel methods for semi-supervised distance metric learning with additional information in the form of pairwise similarity and dissimilarity constraints. More specifically, metric learning in nonparametric and parametric forms, kernel-based metric learning, and metric learning based on manifold structure are presented in turn. We apply our methods to some real-world applications, such as content-based image retrieval and color image segmentation. Experimental results show that our proposed methods outperform previous metric learning methods. |
