基于主动学习SVM分类器的视频分类

Video genre categorization using SVM classifiers with active learning

提出一种基于主动学习SVM分类器的视频分类算法.该算法分为两个步骤:首先分析并提取与视频类型有关的十维底层视觉特征;然后用SVM分类器建立这些底层特征与视频类型之间的联系.在获取SVM分类器所需的训练样本时,采用主动学习的方法选择对SVM分类器最"有用"的样本提供给用户进行标注,用更少的训练样本获得与大量训练样本近似的分类效果,从而减轻用户标注负担.针对多类SVM分类的主动学习问题,提出用后验概率计算分类器对未标注样本的置信度进行样本选择.实验结果表明,主动学习算法与随机采样标注的被动学习算法相比,在相同的训练样本情况下能够获得更高的分类精度;而基于后验概率选择样本的主动学习要略好于传统的基于变型空间(version space)选择样本的主动学习.

A video genre categorization scheme based on SVM classifiers with active learning was proposed. The scheme consists of two steps： firstly, ten computable visual features related to video genres were analyzed and extracted; secondly, to acquire training samples for SVM classifiers, active learning was adopted to select the most informative samples for users to label. This can achieve comparable categorization accuracy with fewer training samples labeled. Moreover, to tackle the problem of selecting informative samples in the multi-class case, a new sample selection strategy was proposed. In this strategy, posterior probability was used to calculate the confidences of unlabeled samples. The informative samples were selected from the unlabeled samples according to their confidences. Experimental results show that active learning achieves good video genre categorization performance compared with the previous passive learning method, and active learning based on posterior probability performs better than the existing method based on version space.