深度学习驱动的异常检测:一份综合综述

需积分: 36 12 下载量 149 浏览量 更新于2024-07-09 收藏 714KB PDF 举报
"这篇论文深度探讨了深度学习在异常检测中的应用,由GUANSONG PANG、CHUNHUA SHEN(阿德莱德大学)、LONGBING CAO(悉尼科技大学)和ANTON VANDEN HENGEL(阿德莱德大学)撰写。文章回顾了深度异常检测的研究,提出了全面的分类体系,涵盖了3个高级类别和11个细粒度方法类别,并分析了这些方法的关键直觉、目标函数、基础假设、优缺点以及如何应对挑战。此外,还讨论了未来可能的研究机会和新视角。" 深度学习在异常检测中的应用是一个日益重要的研究领域,尤其是在各种研究社区中。异常检测,也称为离群值检测或新颖性检测,旨在识别与正常模式显著不同的数据点。传统方法虽然有效,但面对复杂性和挑战时往往力不从心。近年来,深度学习的出现为解决这些问题提供了新的可能性。 论文首先对深度异常检测进行了全面的分类,这包括三个高层次的类别:(1) 基于模型的方法,这些方法利用深度神经网络构建数据的复杂表示;(2) 基于统计的方法,利用深度学习来学习数据的统计特性;(3) 基于生成的方法,如生成对抗网络(GANs),通过生成数据来捕捉正常行为的分布。每个高层次类别下又细分为多个子类别,比如在基于模型的方法中,可以有自编码器(Autoencoders)、卷积神经网络(CNNs)等。 作者们详细讨论了这些方法的核心思想,例如,自编码器通过学习数据的压缩表示,然后在解码过程中检测重构误差来识别异常。CNNs则擅长处理图像数据中的局部特征,从而发现不寻常的模式。此外,他们还分析了各种方法的基础假设,如假设正常数据遵循一定的分布,或者认为异常数据是低概率事件。 论文进一步比较了各种方法的优点和缺点。例如,基于模型的方法通常适用于高维数据,但可能对训练数据的质量和数量敏感;而基于生成的方法能够生成高度逼真的数据,但训练过程可能更复杂且容易陷入局部最优。 针对深度学习在异常检测中面临的挑战,如数据稀疏性、不平衡性和实时性,论文提出了未来的研究方向,包括改进模型的泛化能力、开发适应小样本的深度学习框架,以及优化在线学习和实时检测算法。此外,跨领域应用和多模态数据融合也被认为是未来的重要研究领域。 这篇论文全面概述了深度学习在异常检测中的应用现状,并指明了未来的研究趋势,对于理解这一领域的最新进展和潜在问题具有很高的参考价值。
177 浏览量
Intelligent video systems and analytics represent an active research field combining methods from computer vision, machine learning, data mining, signal processing and other areas for mining meaningful information from raw video data. The availability of cheap sensors and need for solving intelligent tasks facilitate the growth of interest in this area. Vast amount of data collected by different devices require automatic systems for analysis. These systems should be able to make decisions without human interruption or with minimal assistance from a human operator. Video analytics systems should understand and interpret a scene, detect motion, classify and track objects, explore typical behaviours and detect abnormal events [1]. The application area of such systems is huge: preventing crimes in public spaces such as airports, railway stations, or schools; counting objects at stadiums or shopping malls; detection of breaks or leaks; smart homes for elderly people maintenance with fall detection functionality and others. Behaviour analysis and anomaly detection are essential parts of intelligent video systems [2,3]. The objectives of anomaly detection are to detect and inform about any unusual, suspicious and abnormal events happening within the observed scene. These may be pedestrians crossing a road in a wrong place, cars running on the red light, abandoned objects, a person fall, a pipe leak and others. Decisions made by a system should be interpretable by a human therefore the system should also provide information about typical behaviours to confirm its decisions. This thesis develops machine learning methods for automatic behaviour analysis and anomaly detections in video. The methods allow to extract semantic patterns from data. These patterns can be interpreted as behaviours and they are used as a basis for decision making in anomaly detection.