摘要
提出一种改进的事件检测算法,通过交叉特征学习实现相关样本的自适应利用。首先将相关性水平看成是有序标签,利用标签候选集中相邻两个相关性标签的最大容限准则进行模型学习。然后采用多核学习理论来定义标签加权问题,通过交叉特征预测来更新标签候选集合。重复上述步骤直到算法收敛为止,将最终获得的统一检测器用于事件检测。利用大规模TRECVID 2011数据集来测试本文算法,实验结果表明,就平均精度和Pmiss值而言,本文算法的检测性能优于当前其他算法。
It is difficult to realize the complex event detection using the relevant samples while multiple features are available. Relevant samples share certain positive elements of the event,but have no uniform pattern due to the huge variance of relevance levels among different relevant samples. The existing detection schemes lack consideration of the correlation between the event features,and weaken the accuracy of event detection. In this paper,an improved algorithm is proposed which adaptively utilizes the relevant samples by cross-feature learning. Firstly,the relevance levels are treated as ordinal labels,and we learn the model with the maximum margin criterion between the consecutive relevance labels from a label candidates set. Then the label weighting problem is formulated based on the multiple kernel learning theory,we update the label candidate set from the prediction of cross-feature. The procedure is repeated until convergence and the final unified detector is used for event detection. We test our algorithm using the large scale TRECVID 2011 dataset,experimental results show that the detection performance of the proposed algorithm is superior to other current algorithms in terms of average accuracy and Pmiss value.
出处
《实验室研究与探索》
CAS
北大核心
2016年第5期141-145,244,共6页
Research and Exploration In Laboratory
关键词
复杂事件检测
相关样本
交叉特征学习
标签候选集
平均精度
complex event detection
relevant samples
cross-feature learning
label candidates set
average accuracy