Random forest based long-term learning for content based image retrieval

The use of content based image retrieval system in real life applications is limited because of the semantic gap between the low level and high level image features, used for the image similarity measure. Relevance feedback has been considered as a competent technique to overcome the semantic gap problem. However most of the relevance feedback based algorithms suffer from the problem of imbalanced dataset, which means that the numbers of irrelevant images are considerably larger than the number of relevant images for training the classifier. This imbalanced dataset problem causes the degradation in the retrieval results. In order to tackle this problem of imbalanced dataset, a long-term learning approach based on random forest classifier ensemble is proposed in this paper. The long-term learning relevance feedback approach collects the user feedback information for gaining the semantic knowledge of the database images. This knowledge is then learned by random forest classifier to improve the retrieval results. In the experimental evaluation it has been observed that there is a significant improvement in classification accuracy as compared to existing technique available in the literature. A precision of 93% has been observed in 9 iterations of the relevance feedback.