Leveraging Concept Association Network for Multimedia Rare Concept Mining and Retrieval

Automatic high-level semantic concept detection is a crucial step for multimedia data management, indexing, and retrieval. It is well-acknowledged that semantic gap poses a great challenge in multimedia content-based research. It becomes even more challenging when the concept of interest is extremely rare in the training data sets because of the poor modeling for the positive instances. In this paper, a Concept Association Network (CAN) is trained by selecting significant links to capture the strong associations among different concepts using association rule mining (ARM). By taking into account of the correlations and credibilities of reference concept nodes, the advantages of the reference nodes are utilized. Experimental results using TRECVID 2010 data sets show that by utilizing the proposed framework, the Mean Average Precision (MAP) values of all the concepts are improved, and the significant improvement of the MAP values of the rare concepts further attests the promising results.