A concept-based model for image retrieval systems

Content-based image retrieval systems are designed to retrieve images based on the high-level desires and needs of users. However, due to the use of low-level features, image retrieval systems are faced with the so-called semantic gap problem in describing high-level concepts. In order to address this critical problem, a new concept-based model is proposed in this paper. The proposed model retrieves images based on two conceptual layers. In the first layer, the object layer, the objects are detected using the discriminative part-based approach. The second layer, on the other hand, is designed to recognize visual composite, a higher level concept to specify the related co-occurring objects. In the proposed model, this concept is recognized by a new template structure including the appearance filters, constraints, and a set of parameters trained by latent SVM. Experiments are carried out on the well-known Pascal VOC dataset. Results show that the proposed model significantly outperforms the existing content-based approaches.