采用元学习的弱监督视频异常检测方法

Weakly supervised video anomaly detection method based on meta-learning

针对现有弱监督类异常检测方法只考虑准确性而忽略对未知场景的泛化能力,导致模型转移到新场景时性能不佳的问题,提出了一种基于元学习的弱监督视频异常检测方法。该方法的核心思想是通过元学习训练一个自适应模型,通过设计多个任务使模型快速适应新的场景。构建了一个两阶段的视频异常检测框架。在内循环阶段,通过减少任务内部损失函数提高基础检测器的检测精度;在外循环阶段,使模型适应不同任务,提高模型的内部表示能力,使其易于在新的场景中快速微调。所提出方法可在不降低已有模型准确性的前提下提高模型对未知场景的泛化能力,大幅减少模型转移新场景时的迭代次数与训练时间。在UCF-Crime数据集、XD-Violence数据集和UCSD Ped2数据集上的实验结果表明,新方法的训练迭代轮数分别降低到105、125和135。

Video anomaly detection usually involves many unknown scenarios,and current weak supervision methods only consider the accuracy of anomaly detection and ignore the generalization ability of unknown scenarios,resulting in poor performance when the model is transferred to a new scenario.To address the generalization problem of the model,this paper proposes a meta-learning based method.The core idea of the method is to learn an adaptive model through meta-learning and make the new model adapt to a new scenario quickly by designing multiple tasks.This method builds a two-stage video anomaly detection framework.In the inner phase,the detection accuracy of the basic detector is improved by reducing the internal loss function of the task.In the outer loop phase,the model is adapted to different tasks and the internal representation of the model is improved,so that it is easy to fine-tune quickly in new scenarios.The new method improves the generalization ability of the model to unseen scenarios without reducing the accuracy of the existing method,and greatly reduces the number of iterations and training time when the model transfers to the new scenarios.The number of training iterations on UCF-Crime dataset,XD-Violence dataset and UCSD Ped2 dataset is reduced to 105,125 and 135 rounds respectively.