Kernel-based similarity learning

The concept of similarity measurement is is systematically proposed. Although there are some previous related works on this issue, similarity learning has not attracted the attention it deserves. We formulate the problem of similarity measurement learning and propose a framework to solve it. In our framework, the similarity measure is the distance of the samples in some feature space, therefore to learn the similarity measure is to learn a feature mapping function. Previous works are surveyed and they are integrated in this framework. Some applications of similarity measure learning are discussed, including fingerprint, face recognition and content-based image retrieval. A kernel-based method to learning the similarity measure is proposed and experimental result is given and discussed.