Towards Total Recall in Industrial Anomaly Detection (2106.08265v2)

Abstract: Being able to spot defective parts is a critical component in large-scale industrial manufacturing. A particular challenge that we address in this work is the cold-start problem: fit a model using nominal (non-defective) example images only. While handcrafted solutions per class are possible, the goal is to build systems that work well simultaneously on many different tasks automatically. The best performing approaches combine embeddings from ImageNet models with an outlier detection model. In this paper, we extend on this line of work and propose \textbf{PatchCore}, which uses a maximally representative memory bank of nominal patch-features. PatchCore offers competitive inference times while achieving state-of-the-art performance for both detection and localization. On the challenging, widely used MVTec AD benchmark PatchCore achieves an image-level anomaly detection AUROC score of up to $99.6\%$, more than halving the error compared to the next best competitor. We further report competitive results on two additional datasets and also find competitive results in the few samples regime.\freefootnote{$^*$ Work done during a research internship at Amazon AWS.} Code: github.com/amazon-research/patchcore-inspection.

• The paper presents PatchCore, a novel framework that leverages patch-level features from pre-trained networks to address cold-start anomaly detection.
• It employs coreset subsampling to efficiently manage memory banks, achieving image-level AUROC scores up to 99.6% on benchmark datasets.
• The methodology supports robust low-shot learning, enabling accurate industrial defect detection even with minimal nominal training samples.

Overview of PatchCore for Industrial Anomaly Detection

The paper "Towards Total Recall in Industrial Anomaly Detection" proposes PatchCore, an innovative approach to cold-start anomaly detection, focusing on identifying defects in industrial imagery. The central challenge addressed is the cold-start scenario, wherein only nominal images are available for model training—this is particularly pertinent in situations where defective samples are rare or costly to obtain. PatchCore stands out by leveraging memory banks of patch-level features extracted from pre-trained convolutional neural networks, such as ResNet and WideResNet, employed without task-specific fine-tuning.

Methodology

PatchCore's operational foundations lie in several key components:

1. Patch-Level Features: The approach begins by transforming nominal images into a collection of patch features. These patches capture intermediate-level feature representations, balancing the retention of spatial resolution and the avoidance of site-specific biases inherent in ImageNet-based models. By utilizing mid-level hierarchies, PatchCore significantly enhances detection robustness.
2. Coreset Subsampling: Given the potential size of memory banks, efficiently managing computational resources becomes crucial. PatchCore employs coreset subsampling, which strategically reduces the number of necessary patches without appreciably diminishing detection performance. This method distinctly outperforms naive random subsampling, optimizing for memory bank representativeness.
3. Anomaly Scoring Mechanism: Anomalies are identified through nearest-neighbor computations within the memory bank, measuring the distance between test patch features and their nearest nominal counterparts. The approach includes a reweighting mechanism accounting for the rarity of nominal features, refining the anomaly score calculations and ensuring comprehensive defect detection.

Results

Experiments conducted on the MVTec Anomaly Detection benchmark and other specialized datasets showcase PatchCore's effectiveness:

• Performance Metrics: PatchCore achieves superior image-level detection AUROC scores, peaking at 99.6% on MVTec AD, demonstrating a significant error reduction compared to existing methodologies. Its segmentation capabilities are similarly notable, with pixelwise AUROC reaching 98.2%.
• Efficiency: Through coreset subsampling, the model operates with reduced inference time while maintaining performance metrics. The ability to scale with larger image resolutions is particularly advantageous for practical applications.
• Low-Shot Learning: Remarkably, the model maintains robust performance even when trained with drastically reduced nominal datasets, outperforming state-of-the-art alternatives in low-shot scenarios.

Implications and Future Directions

PatchCore's framework presents several promising implications, both theoretically and practically. Its reliance on pre-trained models ensures broad applicability across various anomaly detection tasks without the need for domain-specific retraining. This characteristic underlines its utility in real-world industrial environments where computational resources and labeled data are limited.

Future research could explore extending PatchCore's capabilities by integrating model adaptation techniques to further enhance specificity to nominal data semantics, potentially elevating anomaly detection performance. Additionally, investigating ensemble strategies or hybrid approaches combining PatchCore with adaptive fine-tuning may offer further improvements.

In summary, this work marks a step forward in industrial anomaly detection, demonstrating that leveraging pre-trained networks with intelligent memory and computational strategies can yield state-of-the-art results in challenging cold-start scenarios.