Maximum similarity based feature matching and adaptive multiple kernel learning for object recognition

In this thesis, we perform object recognition using (i) maximum similarity based feature matching, and (ii) adaptive multiple kernel learning. Images are likely more similar if they contain objects within the same categories, so how to measure image similarities correctly and efficiently is one of the critical issues for object recognition. We first propose to match features between two images so that their similarity is maximized, and employ support vector machines (SVMs) for recognition based on the maximum similarity matrix. Secondly, given several similarity matrices (kernels) created by different visual information in images, we propose a novel adaptive multiple kernel learning technique to generate an optimal kernel from all the kernels based on biconvex optimization. These two new approaches are tested on the most recent image benchmark datasets and their results are impressive, equalling or bettering the state-of-the-art results.