Reviewing visual anomaly detection
Anomaly detection extracts non-conforming patterns based on expected behavior to create a reasonable model for observing unseen data. Visual anomaly detection refers to identifying and localizing anomalous regions in imagery data using machine learning (ML) approaches, such as supervised, semi-supervised, and unsupervised techniques to surface anomalies through visual inspection. Please refer to Chapter 1 for comparisons of these techniques.
There are two approaches for assessing visual anomalies:
- Image-level: This approach evaluates whether the whole image is normal or abnormal.
- Pixel-level: This approach detects abnormal regions in an image to determine whether an image is normal or abnormal. The pixel-level method is widely used in industrial fault detection and medical diagnosis. However, this approach can be challenging for mission-critical applications due to large variations and ambiguous boundaries.
Generally, most computer...