Guilin University of Electronic Technology School of Information and Communication, Guangxi, Guilin 541000, China.
Comput Intell Neurosci. 2022 Oct 5;2022:8633955. doi: 10.1155/2022/8633955. eCollection 2022.
Anomaly detection in videos is a challenging computer vision problem. Existing state-of-the-art video anomaly detection methods mainly focus on the structural design of deep neural networks to obtain performance improvements. Different from the main research trend, this paper focuses on combining ensemble learning and deep neural networks and proposes an approach based on ensemble generative adversarial network (GAN). In the proposed method, a set of generators and a set of discriminators are trained together, so each generator gets feedback from multiple discriminators and vice versa. Compared with a single GAN, the proposed ensemble GAN can better model the distribution of normal data to better detect anomalies. In the experiments, the performance of the proposed method is tested on two public datasets. The results show that ensemble learning significantly improves the performance of a single detection model, which outperforms some existing state-of-the-art methods.
视频异常检测是计算机视觉领域中的一个具有挑战性的问题。现有的最先进的视频异常检测方法主要侧重于深度神经网络的结构设计，以获得性能的提升。与主要研究趋势不同，本文专注于结合集成学习和深度神经网络，并提出了一种基于集成生成对抗网络（GAN）的方法。在提出的方法中，一组生成器和一组判别器一起进行训练，因此每个生成器都可以从多个判别器中获得反馈，反之亦然。与单个 GAN 相比，所提出的集成 GAN 可以更好地建模正常数据的分布，从而更好地检测异常。在实验中，该方法的性能在两个公共数据集上进行了测试。结果表明，集成学习显著提高了单个检测模型的性能，优于一些现有的最先进的方法。
