华中科技大学学报（自然科学版）2025，Vol.53Issue(5)：171-177,7.DOI:10.13245/j.hust.250330
基于特征差异学习的弱监督视频异常检测算法
Weakly supervised video anomaly detection algorithm based on feature difference learning
摘要
Abstract
A video anomaly detection algorithm based on feature difference learning was proposed to address the issue that existing weakly supervised video anomaly detection algorithms based on multi instance learning mainly focued on learning the discriminative features of anomaly examples,while ignoring the guidance information of normal patterns.First,a multi-scale time feature fusion network was constructed to extract local time dependency information across multiple time spans,and local time information was utilized to assist attention mechanisms in capturing the global time dependency of video segments.Then,a sorting loss constrained by feature differences was designed.By utilizing the correlation between anomaly and normal at the feature level,the selection of abnormal fragments was defined as the degree of difference between them and normal,which could improve the accuracy of selecting anomalous segments.Finally,the proposed sorting loss and the classification loss were employed to train the entire network model.Experimental results show that the proposed algorithm achieves accuracy of 86.40%and 84.26%on the UCF-Crime and XD-Violence dataset,respectively,effectively improving the performance of video anomaly detection.关键词
视频异常检测/多示例学习/多尺度特征融合/交叉注意力机制/特征差异学习Key words
video anomaly detection/multiple instance learning/multi-scale feature fusion/cross attention mechanism/feature difference learning分类
计算机与自动化引用本文复制引用
唐俊,张印,王科,鲍文霞..基于特征差异学习的弱监督视频异常检测算法[J].华中科技大学学报（自然科学版）,2025,53(5):171-177,7.基金项目
安徽省自然科学基金资助项目(2308085QF228). （2308085QF228）
