Fast Online Svdd Based On Support Vectors Merging

Online support vector data description (SVDD) performs excellently when dealing with novelty detection problems. This method only uses the support vectors (SVs) as an initial training set and incrementally updates the SVDD classifier when receiving new samples. Therefore, as the SVs number tends to grow along with the number of training samples, the update time grows exponentially. To tackle this problem, we propose a new online SVDD method based on SVs merging procedure in this paper. SVs merging procedure bound the size of training set during online learning, thus the proposed method maintains a constant update time. Experimental results of the banana datasets, Mnist datasets and Cifar 10 datasets show that the proposed method achieves higher training speed than both the canonical online SVDD method and the incremental SVDD method, while maintaining a comparable classification accuracy.