Discriminative probabilistic kernel learning for image retrieval

Learning the content-level similarity over images is a fundamental problem in content-based image retrieval, and is highly challenging due to the variance within images. Such variance requires the similarity measure to be adaptive enough. In this paper, we propose a similarity learning approach based on the probabilistic modeling of images. First, we derive the similarity measure, free energy score space kernel (FESS kernel), from the probabilistic models. FESS kernel is essentially the function of the observed data, hidden variable and model parameters, where the hidden variables are very informative and are absent in previous methods. Then, we propose a discriminative learning approach for FESS kernel, encouraging the similarity to take a large value for the image pair with the same label and to take a small value for the image pair with distinct labels. The learned similarity fully exploits the data distribution and class label, and inherits the adaptability. We evaluate the proposed method on three databases. The results validate its competitive performance.