Training Asymmetry SVM in Image Retrieval Using Unlabeled Data
Wensheng Wang, James J. Wu · 2009
Support vector machine (SVM) based relevance feedback (RF) schemes have been shown as a key technique for improving content-based image retrieval (CBIR) performance. However, SVM based RF schemes often suffer from small example problem. To address this problem, an asymmetry semisupervised SVM, AS3VM for short, has been proposed in this paper, which combines query point movement (QPM) technique with SVM and selects unlabeled examples for the positive and negative classes with different strategies. Concretely, in each round of RF, one virtual positive example is generated using QPM. Meanwhile, a random subset of unlabeled examples is selected as the candidate negative examples, and then QPM is used to it for data editing purpose. Finally, the virtual positive example and the edited unlabeled examples are exploited in conjunction with the labeled examples to train a SVM for image retrieval. Experimental results shows that the proposed scheme is more effective than other state-of-the-art approaches.