Scholarlink

Detecting irregularities in weather data serves multiple practical applications. For example, nowcasting focuses on predicting atmospheric conditions for the next 0 to 4 h, which is essential for effective emergency response and disaster management. Anomaly detection also plays a crucial role in forecasting extreme weather events. However, the complexity increases when considering both anomalies and varying weather conditions. A common approach to anomaly detection is weakly supervised video-level labeling, which aims to identify frames containing abnormal events and is typically framed as a multiple instance learning (MIL) problem. While existing methods perform well, the prevalence of negative instances significantly hinders their ability to detect positive instances, particularly rare abnormal segments. To address this, we aim to extract distinctive features by enhancing the observable differences between various classes using a single branch. We propose a novel method, Transformer Encoded Feature Video Anomaly Detection (TEF-VAD), which exclusively utilizes attention mechanisms, specifically Multi-Head Attention Learning. This approach combines feature magnitude learning loss, class-specific loss, and a TEF-VAD-enhanced MIL classifier training loss, thereby training a model to effectively identify positive examples and improve the MIL method’s robustness for detecting positive instances in abnormal videos. Our extensive experiments demonstrate that the MIL model enhanced by our Transformer method significantly improves sample efficiency and the detection of subtle anomalies, outperforming several state-of-the-art techniques on benchmark datasets like UCF-Crime and ShanghaiTech.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Robust Anomaly Detection Through Transformer-Encoded Feature Diversity Learning

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Kuldeep Biradar,
  • Dinesh Kumar Tyagi,
  • Ramesh Babu Battula,
  • P. Subbarao

摘要

Detecting irregularities in weather data serves multiple practical applications. For example, nowcasting focuses on predicting atmospheric conditions for the next 0 to 4 h, which is essential for effective emergency response and disaster management. Anomaly detection also plays a crucial role in forecasting extreme weather events. However, the complexity increases when considering both anomalies and varying weather conditions. A common approach to anomaly detection is weakly supervised video-level labeling, which aims to identify frames containing abnormal events and is typically framed as a multiple instance learning (MIL) problem. While existing methods perform well, the prevalence of negative instances significantly hinders their ability to detect positive instances, particularly rare abnormal segments. To address this, we aim to extract distinctive features by enhancing the observable differences between various classes using a single branch. We propose a novel method, Transformer Encoded Feature Video Anomaly Detection (TEF-VAD), which exclusively utilizes attention mechanisms, specifically Multi-Head Attention Learning. This approach combines feature magnitude learning loss, class-specific loss, and a TEF-VAD-enhanced MIL classifier training loss, thereby training a model to effectively identify positive examples and improve the MIL method’s robustness for detecting positive instances in abnormal videos. Our extensive experiments demonstrate that the MIL model enhanced by our Transformer method significantly improves sample efficiency and the detection of subtle anomalies, outperforming several state-of-the-art techniques on benchmark datasets like UCF-Crime and ShanghaiTech.