Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Domain-independent detection of known anomalies

Jonas Bühler, Jonas Fehrenbach, Lucas Steinmann, Christian Nauck, Marios Koulakis

TL;DR

This work addresses industrial anomaly detection under the constraint of sparse, known anomaly types across unseen objects by introducing a domain-generalization-on-sparse-classes task. It creates three cross-domain datasets derived from MVTec AD (hole, cut, color) to benchmark performance and introduces two embedding-based methods, Labeled PatchCore and SEMLP, alongside strong baselines. SEMLP achieves the best average image-level AUROC (87.2%) and often outperforms MIRO and PatchCore, highlighting the effectiveness of per-embedding MLP classifiers over coreset-based distance scoring in this setting. The open datasets and proposed approaches offer a practical pathway for deploying robust anomaly detection in diverse industrial contexts, with future work focusing on threshold-free operation and deeper analysis of failure cases.

Abstract

One persistent obstacle in industrial quality inspection is the detection of anomalies. In real-world use cases, two problems must be addressed: anomalous data is sparse and the same types of anomalies need to be detected on previously unseen objects. Current anomaly detection approaches can be trained with sparse nominal data, whereas domain generalization approaches enable detecting objects in previously unseen domains. Utilizing those two observations, we introduce the hybrid task of domain generalization on sparse classes. To introduce an accompanying dataset for this task, we present a modification of the well-established MVTec AD dataset by generating three new datasets. In addition to applying existing methods for benchmark, we design two embedding-based approaches, Spatial Embedding MLP (SEMLP) and Labeled PatchCore. Overall, SEMLP achieves the best performance with an average image-level AUROC of 87.2 % vs. 80.4 % by MIRO. The new and openly available datasets allow for further research to improve industrial anomaly detection.

Domain-independent detection of known anomalies

TL;DR

This work addresses industrial anomaly detection under the constraint of sparse, known anomaly types across unseen objects by introducing a domain-generalization-on-sparse-classes task. It creates three cross-domain datasets derived from MVTec AD (hole, cut, color) to benchmark performance and introduces two embedding-based methods, Labeled PatchCore and SEMLP, alongside strong baselines. SEMLP achieves the best average image-level AUROC (87.2%) and often outperforms MIRO and PatchCore, highlighting the effectiveness of per-embedding MLP classifiers over coreset-based distance scoring in this setting. The open datasets and proposed approaches offer a practical pathway for deploying robust anomaly detection in diverse industrial contexts, with future work focusing on threshold-free operation and deeper analysis of failure cases.

Abstract

One persistent obstacle in industrial quality inspection is the detection of anomalies. In real-world use cases, two problems must be addressed: anomalous data is sparse and the same types of anomalies need to be detected on previously unseen objects. Current anomaly detection approaches can be trained with sparse nominal data, whereas domain generalization approaches enable detecting objects in previously unseen domains. Utilizing those two observations, we introduce the hybrid task of domain generalization on sparse classes. To introduce an accompanying dataset for this task, we present a modification of the well-established MVTec AD dataset by generating three new datasets. In addition to applying existing methods for benchmark, we design two embedding-based approaches, Spatial Embedding MLP (SEMLP) and Labeled PatchCore. Overall, SEMLP achieves the best performance with an average image-level AUROC of 87.2 % vs. 80.4 % by MIRO. The new and openly available datasets allow for further research to improve industrial anomaly detection.
Paper Structure (23 sections, 3 figures, 5 tables, 2 algorithms)

This paper contains 23 sections, 3 figures, 5 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. Anomaly Detection
  4. Domain Generalization
  5. Dataset
  6. Methods
  7. Generation of new datasets
  8. Extension of PatchCore: Labeled PatchCore
  9. Simplified Approach: Spatial Embedding MLP
  10. Experiments
  11. Setup
  12. Results
  13. Conclusion
  14. Appendix
  15. Labeled PatchCore
  16. ...and 8 more sections

Figures (3)

  • Figure 1: SEMLP passes the input image to a backbone. The intermediate features are extracted and concatenated to create embeddings. Each embedding is passed to an MLP which will classify single embeddings as normal or anomalous. The image is considered anomalous if at least one embedding is classified as anomalous.
  • Figure 2: Classification using multiple coresets in a two-dimensional scenario: yellow is the coreset containing embeddings of good images and green the coreset with anomalous embeddings. The red circle will be classified based on its distance to the next embedding of each coreset ($d_0$ and $d_1$), after weighting the distances to the next neighbors in their respective coreset like in rothTotalRecallIndustrial2022 (dotted lines).
  • Figure 3: Example images of the custom defined anomaly types (source of individual images: bergmannMVTecADComprehensive2019bergmannMVTecAnomalyDetection2021)