A Robust and Efficient Doubly Regularized Metric Learning Approach

A proper distance metric is fundamental in many computer vision and pattern recognition applications such as classification, image retrieval, face recognition and so on. However, it is usually not clear what metric is appropriate for specific applications, therefore it becomes more reliable to learn a task oriented metric. Over the years, many metric learning approaches have been reported in literature. A typical one is to learn a Mahalanobis distance which is parameterized by a positive semidefinite (PSD) matrix M. An efficient method of estimating M is to treat M as a linear combination of rank-one matrices that can be learned using a boosting type approach. However, such approaches have two main drawbacks. First, the weight change across the training samples maybe non-smooth. Second, the learned rank-one matrices might be redundant. In this paper, we propose a doubly regularized metric learning algorithm, termed by DRMetric, which imposes two regularizations on the conventional metric learning method. First, a regularization is applied on the weight of the training examples, which prevents unstable change of the weights and also prevents outlier examples from being weighed too much. Besides, a regularization is applied on the rank-one matrices to make them independent. This greatly reduces the redundancy of the rank-one matrices. We present experiments depicting the performance of the proposed method on a variety of datasets for various applications.