Class-specific Mahalanobis distance metric learning for biological image classification

Distance metric learning (DML) is an emerging field of machine learning. The basic idea behind DML is to adapt the underlying distance metric to improve the performance for the pattern analysis tasks. In this paper, we present the use of DML techniques to improve the classification accuracy of k-Nearest Neighbour classifier (kNN) used for biological image classification tasks. The distance metric learning technique is used for learning the Mahalanobis distance metric. The learning problem is cast into a Bregman optimization problem that minimizes the LogDet divergence subject to linear constraints. We propose the class-specific Mahalanobis distance metric learning for further improvement of the performance of the kNN classifier. Results of our studies on benchmark data sets demonstrate the effectiveness of the distance metric learning techniques in classification of biological images. © 2012 Springer-Verlag.