1. 基于对比记忆网络的弱监督视频异常检测.
- Author
-
李文中, 吴克伟, 孙永宣, 焦 畅, and 熊思璇
- Subjects
- *
ANOMALY detection (Computer security) , *MEMORY , *CRIME , *LEARNING - Abstract
Anomaly detection aims to capture the discriminative features with limited training samples. However, when some anomalies share common compositional patterns with the normal training data, the model likely reconstructs the anomalies well, leading to the miss detection of anomalies. To mitigate this drawback, this paper proposed a novel contrastive memory network, which used the contrast learning framework to separate the sample features based on the autoencoder, and then designed a memory network to suppress the normal features similar to anomaly. This method proposed a two-stage framework for detecting abnormal events. In the first stage, the method used contrastive learning to increase the difference between normal features and abnormal features, and gained representation to be the augment memory and suppression memory of memory network. In the second stage, the model used augment memory to record multi-time normal behavior features, and used suppression memory to constrain the expression of pseudo anomaly items in the augment memory. The AUC value reached 83. 26% on UCF Crime datasets and 87.53% on Shanghai Tech datasets, which were 1. 14% and 2.43% higher than the existing methods. The results demonstrate that this method can efficiently predict the temporal localization of anomaly events. [ABSTRACT FROM AUTHOR]
- Published
- 2023
- Full Text
- View/download PDF