Long Term Learning in Image Retrieval Systems Using Case Based Reasoning

Relevance feedback is a powerful tool emerged to boost the retrieval performance of content based image retrieval (CBIR) systems. Short term learning (STL) and long term learning (LTL) are two learning methods of relevance feedback scheme. This paper presents a long term learning method in CBIR systems adopting case based reasoning (CBR) which is called Case-based LTL (CB-LTL). The method has two stages of learning and reasoning. In the learning stage, information extracted from retrieval sessions is saved as cases and in the reasoning stage, information of cases is utilized to improve the results of the retrieval sessions. The main components of CB-LTL method are 'key of query' which represents the desire of the user, a 'trigger function' which is used to find a similar case with a query, and 'semantic frame' which is a structure for saving cases. In the proposed method, cases are recorded in the case knowledge base using both low level and high level features. The information of the relevance feedback and short term learning are employed as high level features. In this paper, the general approach of CB-LTL is produced and an example of the method is implemented in a CBIR system with the similarity refinement based STL. To evaluate the proposed method, a comparative study with the "virtual feature based" LTL method is performed based on the Corel image dataset. The experimental results validate the effectiveness of Case-based LTL method empirically.