A Hybrid Distance Metric Learning for Image Ranking

Abstract

The Distance Metric Learning (DML) has been in attentive on image retrieval, but many of the previous methods are used for classification and clustering of the images. In this paper, we are focusing on designing the ordinal DML algorithms for image ranking purpose, hence the rank levels among the images can be well measured by us. A new hybrid approach is proposed in this paper in order to improve efficiency of existing system. Proposed approach is a hybrid algorithm of linear and nonlinear distance metric learning methods. First of all we present a linear ordinal Mahalanobis DML model which tries to preserve the local geometry information as well as the ordinal relationship of the data. Then a nonlinear DML method by kernelizing the above model is developed, here most of the real-world image data with nonlinear structures is considered. for further improvemrnt of the ranking performance, we derive a multiple kernel DML approach taken by the idea of multiple-kernel learning which performs different kernel operations on different kinds of features of image. Extensive experimental analysis demonstrates the relevant results.