Optimizing PatchCore for One-Shot Industrial Anomaly Detection: SEA Vision’s research at ETFA Portugal

Optimizing PatchCore for One-Shot Industrial Anomaly Detection 

 In modern manufacturing, the ability to detect product defects early is crucial to ensuring quality and efficiency. Yet training artificial intelligence systems to identify these anomalies remains a challenge. Traditional models require large sets of labelled images of both normal and defective products; an approach that’s often impractical in real-world production, where defects are rare and costly to reproduce. 

Addressing this challenge, the study “Optimizing PatchCore for One-Shot Industrial Anomaly Detection” explores how to make AI-based quality control more data-efficient. The paper was authored by Simona Zlatanova, Piercarlo Dondi, and Marco Porta from the University of Pavia, with Michele Antolini, Technical Program Manager at SEA Vision. 

Marco Porta, Michele Antolini, Piercarlo Dondi, Simona Zlatanova

The project originated during Simona Zlatanova’s internship at SEA Vision in 2024, where she developed her master’s thesis on anomaly detection in industrial vision systems under Antolini’s supervision. This collaboration between SEA Vision and the University of Pavia led to a paper that was later presented at IEEE ETFA 2025, one of the world’s foremost conferences on emerging technologies for factory automation. 

Tackling the One-Shot Challenge in Industrial Quality Control 

In modern factories, the early detection of anomalies, such as scratches, missing parts, or misalignments, is key to maintaining quality and reducing waste. However, training deep learning systems for defect detection typically requires thousands of labeled images, which is impractical in many real-world scenarios. 

This challenge defines the field of one-shot anomaly detection, where systems must learn to identify defects after seeing only a single image of a “normal” product. It’s akin to showing someone one perfect example of a product and asking them to identify flaws in every future instance. 

PatchCore: The Foundation of Modern Anomaly Detection 

The team’s work builds upon PatchCore, a state-of-the-art algorithm for industrial anomaly detection originally developed by researchers at the Max Planck Institute for Intelligent Systems, University of Freiburg, and Amazon Web Services (AWS). 

PatchCore examines small regions, or “patches” of an image, comparing them against known examples of normality to flag potential defects. Importantly, it doesn’t need to see defective samples during training, making it ideal for environments where defects are rare or costly to replicate. 

Some sample images from the MVTec-AD dataset. 

Optimizing PatchCore for One-Shot Learning 

To make PatchCore more effective in one-shot scenarios, the research team introduced several targeted improvements: 

1. Enhanced Feature Extraction with Anti-Aliased ResNet50 
   By replacing the standard ResNet50 backbone with its anti-aliased variant, the model gains robustness against image distortions caused by motion or shifts, issues common in production lines.
2. Smarter Dimensionality Reduction Using Gaussian Random Projection
   Instead of average pooling, the researchers employed a statistical method that compresses image data while preserving crucial information, enhancing the system’s discrimination power.
3. Image Alignment Optimization
   An optional preprocessing step aligns test images with the training image orientation, improving detection accuracy in certain categories. 

Some sample images from the MVTec-AD dataset. 

Results: Higher Accuracy, Smarter Detection 

The optimized system was evaluated using the MVTec Anomaly Detection (MVTec-AD) dataset, a benchmark collection featuring 15 industrial object categories such as bottles, pills, and screws. Using only one normal image per category for training, the optimized PatchCore achieved: 

• 85% image-level detection accuracy, up from 79% 

• 93% pixel-level accuracy, up from 92% 

• Noticeable gains in categories prone to rotation or texture variation 

The study demonstrates that even with minimal data, intelligent architectural choices and small preprocessing techniques, like rotation and brightness adjustments, can yield substantial improvements. 

A Step Toward the Future of Smart Manufacturing 

The optimized PatchCore architecture marks an important step toward practical, data-efficient anomaly detection in industrial environments. For sectors such as pharmaceuticals, where high precision and regulatory compliance are paramount, the ability to deploy accurate models from minimal data holds immense promise. 

Through collaborations like this, SEA Vision continues to strengthen its position as a bridge between academic research and industrial innovation, proving that fostering young talent is not only an investment in people, but also in the future of intelligent manufacturing. 

Article DOI:
https://doi.org/10.1109/ETFA65518.2025.11205697