Abstract
Industrial Video Anomaly Detection (VAD) presents unique challenges that distinguish it from conventional surveillance tasks. While standard models focus on dense, unpredictable movements like pedestrians or traffic, industrial environments are characterized by sparse object distribution, strict periodicity, and high sensitivity to lighting noise. We introduce INFP (Industrial Video Anomaly Detection via Frequency Prioritization), an encoder-decoder framework designed to leverage the rhythmic "heartbeat" of production lines. By shifting analysis into the frequency domain, INFP effectively isolates product trajectories and remains robust against industrial lighting fluctuations.
Introduction: The Industrial VAD Challenge
Most research in Video Anomaly Detection focuses on "conventional" scenes where objects are dense and move with high entropy. In contrast, industrial production lines exhibit three distinct characteristics that often cause standard models to fail:
- Distribution Sparsity: Products typically occupy only a small fraction of the frame, making them difficult for global spatial models to track.
- Equidistant Periodicity: Products appear at highly regular intervals. Conventional models fail to exploit this temporal regularity.
- Light Sensitivity: Industrial settings are prone to flickering or low-light conditions, creating pixel-level noise that models often misinterpret as anomalies.
Methodology: Frequency-Domain Prioritization
The INFP framework addresses these challenges by prioritizing frequency-domain information through two primary components: the Trajectory Filter and the Multi-feature Fusion Module.
Trajectory Filter
In a factory setting, the background is largely static (low-frequency energy), while moving products generate distinct high-frequency energy. The Trajectory Filter utilizes Fourier Transforms to assign higher weights to the fixed paths where products travel. This effectively "mutes" irrelevant background noise and focuses the model's attention on the product's motion path without requiring manual labels.
Multi-feature Fusion Module
To capture the complex nature of industrial video, this module extracts and integrates three feature types:
- Spatial Features: Standard visual representations of the frame.
- Temporal Features: Captured via temporal shifts to understand motion.
- Frequency Features: Extracted using Fast Fourier Transform (FFT) to capture the unique periodic signature of the production line.
These features are fused and processed through a contextual integration step to mitigate "pixel jitter" caused by shifting light, ensuring stable anomaly detection even in harsh environments.
Experimental Evaluation
The model was evaluated on the IPAD (Industrial Period Video Dataset), the first dataset specifically curated for industrial VAD, containing over six hours of footage across 16 device types. It covers real-world anomalies such as speed variations, structural deformities, and incorrect product angles.
Key Results
INFP established a new state-of-the-art (SOTA) benchmark on the IPAD dataset:
- Performance: Achieved an average AUC of 71.8%, significantly outperforming conventional models like conAE (56.7%) and specialized baselines like IPAD_VAD (68.6%).
- Ablation Gains: The Multi-feature Fusion module and Trajectory Filter contributed 7.2% and 5.9% boosts to the AUC, respectively, validating the frequency-prioritization approach.
- Robustness: In low-light tests, INFP maintained stable performance while baseline models experienced sharp declines in accuracy due to noise.
Conclusion and Future Work
INFP demonstrates that the frequency domain provides a superior "fingerprint" for capturing the periodicity of industrial processes. By filtering for high-frequency energy and fusing multi-modal features, the model achieves high robustness to lighting and sparsity. Future research will focus on adapting these techniques to non-linear production tasks and integrating Large Language Models to provide natural language explanations for detected anomalies.
Summary updated