Learning to Combine Ad-hoc Ranking Functions for Image Retrieval

Along with the success of "bag of visual words" scheme in content based image retrieval (CBIR), various technologies in text information retrieval realm have been transferred into image retrieval system and obtain promising performance. However, how to select the suitable ranking technology, such as ranking model, for a specific image database is still an open question. Because most ranking models are data-dependent, it is hard to find an optimal model for all the applications. In this paper, we propose to resolve this problem for CBIR with the learning to rank approach which has been widely utilized in text retrieval. Specifically, we consider several well performed ad-hoc ranking models and use their ranking scores to construct the ranking features for the Ranking SVM framework. To best preserve the spatial structures existed in the visual words of image, we split the image into different size blocks, and design the ranking features with a pyramid approach from large blocks to small blocks. Experimental results on both Oxford and Image Net databases demonstrate the effectiveness of proposed method compared with the performance that individual ranking model is adopted. Moreover, the proposed method brings little computational burden to the system and the efficiency analysis proves its scalability.