Multi-supervised CML for Small Sample Low-resolution Image Matching

The coupled metric learning(CML) is an effective method for low-resolution image matching. However, the existing coupled metrics generally use class label as the supervision, which easily leads to changes in the samples distribution after coupling transformation. These changes affect image matching accuracy seriously. To address this problem, we propose a multi-supervised coupled metric learning method fusing class label and distribution information. In this work, a novel multi-supervised objective function is constructed, which consists of the main and auxiliary objective functions. The class label in the main objective function plays key supervisory role, and the distribution information in the auxiliary objective function plays auxiliary supervisory role. Then, by solving the objective function, the samples with different resolution are transformed into a common space for distance measurement. Experimental results on two image datasets validate the efficacy of the proposed method.