Kernel-based distance metric learning for content-based image retrieval

For a specific set of features chosen for representing images, the performance of a content-based image retrieval (CBIR) system depends critically on the similarity or dissimilarity measure used. Instead of manually choosing a distance function in advance, a more promising approach is to learn a good distance function from data automatically. In this paper, we propose a kernel approach to improve the retrieval performance of CBIR systems by learning a distance metric based on pairwise constraints between images as supervisory information. Unlike most existing metric learning methods which learn a Mahalanobis metric corresponding to performing linear transformation in the original image space, we define the transformation in the kernel-induced feature space which is nonlinearly related to the image space. Experiments performed on two real-world image databases show that our method not only improves the retrieval performance of Euclidean distance without distance learning, but it also outperforms other distance learning methods significantly due to its higher flexibility in metric learning.