Boosting-Based Relevance Feedback for CBIR

摘要

In this research, we implemented Boosting-based Relevance Feedback (BRF) technique for Content-Based Image Retrieval (CBIR) system. The BRF technique follows two stages. In the first stage, the system returns the results of image retrieval based on the dissimilarity measure using Jeffrey Divergence with threshold 0.15. In the second stage, the system returns the results of the image retrieval based on the prediction of the BRF model which is generated based on the user's feedback image. With the same procedure, every feedback generates a BRF model that corresponds to the user's feedback images. In this study, we compare existing three Boosting algorithms, i.e.: AdaBoost, Gradient Boosting, and XGBoost. We consider the performance of application from precision, recall, F-measure, and accuracy value. The best BRF technique is XGBoost on fourth feedback, based on the results of experiments that conducted on the Wang Dataset. The BRF technique using XGBoost enhances the average precision value by 18.82%, the average recall value amount 173.32%, the average F-measure value amount 94.97%, and the average accuracy value amount 4.15% compared with the baseline. The BRF technique using XGBoost achieves the best performance on both the average recall and F-measure value compared to the most recent methods.