ViCO: Vision-Augmented Time Series Anomaly Detection with Prototype-Guided Coordination
Abstract
Time series anomaly detection aims to identify time points or intervals that deviate from normal temporal behavior. In reconstruction-based detection, this requires errors that distinguish anomalies rather than merely accurate reconstruction. However, most existing methods rely solely on temporal modeling, while vision-augmented approaches focus primarily on forecasting, leaving the potential of structural visual cues for reconstruction-based detection underexplored. Moreover, bridging temporal and visual representations further remains challenging. To address these challenges, we propose a Vision-Augmented Coordination framework for time series anomaly detection, called ViCO. A decoupled visual teacher and temporal student characterize normal behavior from complementary views: tri-view visual representations capture structural regularities, whereas raw sequences model value-sensitive dynamics. Prototype-Guided Coordination retrieves normal structural prototypes and conditions the temporal decoder, translating visual structure into reconstruction guidance for anomaly scoring. Extensive experiments on five real-world datasets demonstrate that ViCO consistently outperforms strong baselines by up to 9.0%, while enabling teacher-free inference for lightweight deployment. Code is available at https://github.com/Cedarrr1/ViCO.