Scaling Anomaly Detection with Segmentation Models

The paper addresses the challenge of anomaly detection in industrial product images, especially for large, complex, or high-resolution objects. Existing methods often fail to scale while preserving precision. The proposed approach, SADSeM (Scaling Anomaly Detection with Segmentation Models), leverages segmentation networks like Mask-RCNN. By combining segmentation maps with feature embeddings, SADSeM performs unsupervised anomaly detection. Since segmentation is robust across image sizes, the method scales effectively to high-resolution images while remaining competitive in simpler cases.

Recommended citation: Samele, S., Attore, F., Matteucci, M. (2025). "Scaling anomaly detection with segmentation models." 2025 Digital Signal Processing 105613.
Paper Link | Download Bibtex