Unsupervised Anomaly Detection for Manufacturing Quality Control

Introduction

In manufacturing quality control, manual defect inspection becomes impractical at scale. While computer vision techniques like semantic segmentation and object detection [1] have revolutionized automated inspection, they face a significant challenge: new product lines lack the extensive labeled training data needed for these models. This data scarcity can compromise quality control effectiveness for novel products.

Project Objectives & Role

In this internship project, I led the evaluation of unsupervised anomaly detection approaches for manufacturing quality control. The key objectives were:

1. Evaluate and compare ten different anomaly detection models, focusing on:
   • Processing capability for non-square images
   • Heat map generation quality
   • Model size and inference speed
   • Detection accuracy (recall)
2. Identify the most effective model for production deployment

Methods

We evaluated ten different anomaly detection models, focusing on their ability to process non-square images and produce meaningful heat maps. The models were trained on defect-free (OK) samples and tested on defective (NG) samples.

Results

Three models emerged as top performers:

EfficientAd [2] demonstrated superior performance with:

• Most accurate heat map generation
• Smallest model size (30 MB)
• Highest recall score (0.55607)
• Effective localization of anomalous regions

Discussion

The unsupervised models serve multiple purposes:

• Generate annotation data for supervised models
• Assist in data quality inspection
• Identify anomaly signals including:
  • Defects
  • Lighting variations
  • Layout differences
• Flag potentially defective samples without pre-existing annotations
• Social Value: By improving defect detection, we enhance product quality and safety, ultimately benefiting consumers and manufacturers.

Personal Contribution

In this internship project, I led the evaluation of unsupervised anomaly detection approaches for manufacturing quality control. The key objectives were:

1. Evaluate and compare ten different anomaly detection models, focusing on:
   • Processing capability for non-square images
   • Heat map generation quality
   • Model size and inference speed
   • Detection accuracy (recall)
2. Identify the most effective model for production deployment
   • Role: My leadership and analytical skills were instrumental in identifying the best-performing model, contributing to the project’s success and its positive social impact.

References

[1] D. Weimer, B. Scholz-Reiter, and M. Shpitalni, “Design of deep convolutional neural network architectures for automated feature extraction in industrial inspection,” CIRP Annals, vol. 65, no. 1, pp. 417–420, 2016.

[2] K. Batzner, L. Heckler, and R. König, “EfficientAD: Accurate Visual Anomaly Detection at Millisecond-Level Latencies,” arXiv.org, Mar. 25, 2023.

[3] H. Deng and X. Li, “Anomaly Detection via Reverse Distillation from One-Class Embedding,” arXiv.org, 2022.

[4] T. Defard, Aleksandr Setkov, A. Loesch, and Romaric Audigier, “PaDiM: a Patch Distribution Modeling Framework for Anomaly Detection and Localization,” arXiv, Nov. 2020.