Re-weighting Relevance Feedback in HSV Quantization for CBIR

Abstract

This research has implemented re-weighting relevance feedback technique in HSV quantization for CBIR with similarity measures using Jeffrey Divergence. The purpose of this study is to determine the best weight factor and the best feedback image selection for re-weighting relevance feedback technique in HSV Quantization so that can enhance the performance of the CBIR system. Performance measurement is evaluated based on precision, recall, and F-measure values obtained from testing results performed on Wang, Corel 10K, and GHIM 10K datasets. Based on the results of the research, the best feedback image selection technique is the selection of the image that closer to the threshold value. The best weight factor of Wang dataset is 0.40, while the best weight factor of Corel 10K and GHIM 10K dataset is 0.45. The re-weighting relevance feedback technique in HSV quantization on all datasets generally does not increase the average precision value, but always increases the recall value and the average F -measure value. On Wang dataset, the proposed technique can increase the average F -measure value up to 2.21 times better than the baseline, i.e. from 18.91 % to 41.86% which is much better than the existing method. On Corel 10K dataset, the proposed technique can increase the average F -measure value up to 1.62 times better than the baseline, i.e. from 17.78% to 28.86% which is much better than the existing method. While the GHIM 10K dataset, the proposed technique can enhance the performance of image retrieval 2.67 times better than the baseline, i.e. from 5.14% to 13.74%.