Support top irrelevant machine: learning similarity measures to maximize top precision for image retrieval

Top precision is one of the most popular performance measures for content-based image retrieval task, while similarity function is the most critical component of a content-based image retrieval system. However, surprisingly, there is no existing similarity function learning method proposed to maximize the top precision measure. To fill this gap, in this paper, we propose the problem of maximum top precision similarity learning, and the first solution to this problem. The similarity is a linear function of the conjunction of features of a query image and a database image. To learn the similarity function parameter matrix, we propose to maximize the top precision measures of the training queries and also minimize the squared ℓ 2 norm of the parameter matrix. The optimization problem is translated to a quadratic programming problem with regard to the Lagrange multipliers of top irrelevant images. The proposed algorithm, named as support top irrelevant machine, is evaluated over four benchmark image databases and is advantage over other similarity learning methods measured by top precision is shown.