A Relevance Feedback System for CBIR with Long-Term Learning

Relevance feedback has been developed to improve retrieval performance effectively in Content Based Image Retrieval (CBIR). This paper introduces a relevance feedback system for CBIR with both short-term relevance feedback and long-term learning. In short-term relevance feedback, query reweighting algorithm, support vector machines (SVM), and genetic algorithm are adopted. In long-term learning, the expanded-judging model with index table is used for analyzing the historical log data. Experimental results show that among short-term feedback algorithms, the SVM gets the best feedback results, and for the use of our proposed expanded-judging model in long-term learning, the recall of the retrieval system is improved more than 30% in average.