Reproduce “Sequential PatchCore: Anomaly Detection for Surface Inspection using Synthetic Impurities” on OpenResearch

About this paper

Runzhou Mao, Juraj Fulir, Christoph Garth, Petra Gospodnetić

This research introduces Sequential PatchCore, a memory-efficient adaptation of PatchCore for anomaly detection on high-resolution images, along with a procedural method for generating photorealistic synthetic water stains. The approach enables training on consumer-grade GPUs and demonstrates the impact of synthetic impurities on model robustness, while proposing a defect-wise recall metric for industrial evaluation.

Key insights

• Incorporating synthetically generated impurities into training data is crucial, as their presence can impact anomaly detection performance, particularly if not properly handled through finetuning.
• The iterative coreset construction of Sequential PatchCore enables the training of high-resolution image-based anomaly detection models on memory-constrained consumer-grade hardware.
• Coreset melding provides an efficient mechanism for transfer learning and adapting models, effectively bridging the domain gap between synthetic and real data or integrating knowledge from different sources.

Summary by alphaXiv.