Mahalanobis PatchCore Enhances Industrial Anomaly Detection with Covariance Awareness
A novel technique known as Mahalanobis PatchCore enhances the detection of visual anomalies in industrial settings by integrating covariance awareness and compatibility with streaming data. This method builds on the PatchCore retrieval framework, which evaluates test images against a repository of normal patch characteristics. Unlike traditional Euclidean geometry, which overlooks feature correlations, Mahalanobis PatchCore employs a regularized covariance model in a reduced feature space and whitens embeddings, facilitating Mahalanobis retrieval through Euclidean nearest-neighbor search. Additionally, it features a bounded-memory, re-iterable training process that incrementally constructs the memory bank without the need to store all normal patches simultaneously, utilizing incremental dimensionality reduction. The research is available on arXiv under ID 2605.27748.
Key facts
- Mahalanobis PatchCore is a covariance-aware extension of PatchCore.
- It uses a regularised covariance model in reduced feature space.
- Whitening embeddings enables Euclidean nearest-neighbor search to implement Mahalanobis retrieval.
- The training pipeline is bounded-memory and re-iterable.
- It avoids storing all normal patches at once.
- The method is designed for streaming compatibility.
- Industrial visual anomaly detection is typically one-class.
- The paper is available on arXiv (ID 2605.27748).
Entities
Institutions
- arXiv