Metric learning with feature decomposition for image categorization

This paper proposes an approach named distance metric learning with feature decomposition (DMLFD) that can reduce the computational costs of distance metric learning (DML) methods as well as improve their performance for image categorization. We have analyzed that a high-dimensional feature space with limited training data will introduce difficulties to DML algorithms in both computation and performance. To tackle these difficulties, we decompose the high-dimensional feature space into a set of low-dimensional feature spaces with minimal dependencies. Then we perform DML for each low-dimensional feature space to construct a sub-metric and the sub-metrics are finally combined into a global metric. We conduct experiments on Corel and TRECVID datasets with different DML methods, and empirical results have demonstrated the effectiveness and efficiency of the proposed approach.