Source-linked AI summary
Trustworthy Visual Quality Inspection under Data Scarcity in Manufacturing
Panagiotis Sapoutzoglou, Jessy Ribaira, Martin Kanounnikoff, Bas Tijsma, Christian Geiß, Maria Pateraki
TL;DR
Manufacturing inspection must operate with scarce defect data while distinguishing reliable automated decisions from uncertain ones. The paper combines diffusion-generated defects, uncertainty-aware classification, and staged checks; preliminary results suggest improved calibration and performance, with online deployment supported by the pipeline’s latency.
Problem
Defect scarcity and unquantified prediction uncertainty make trustworthy automation difficult when manufacturing inspection must handle rare defects.
Method
The paper uses diffusion-generated defects, Bayesian and ensemble uncertainty estimation, and a staged pipeline of segmentation, anomaly detection, and classification.
Results
Preliminary results show better Bayesian calibration, stronger three-class performance as training data increases, and pipeline latency suitable for online applications.
Takeaways & Limitations
Uncertainty estimation can let the pipeline defer decisions it cannot trust to human review while retaining automated inspection for other units.
Abstract
from arXiv · showhide
Automated visual inspection in manufacturing aims to replace slow and inconsistent manual checks, but its economic value depends on whether its decisions can be trusted enough to automate routine inspection while reserving human expertise for ambiguous cases. In production-line settings, defective samples are scarce, since the process is optimized to produce good parts, which limits any learning-based inspector trained on real data alone. Compounding this, defect decisions emitted as hard labels with no confidence estimate carry an asymmetric cost: a false reject wastes a good product, while a false accept may increase the risk of undetected defects progressing through the production process. We address both problems by mitigating data scarcity through the generation of synthetic defective samples with a diffusion model, and meeting the need for confidence-aware decisions with a Bayesian classifier that defers ambiguous units to human review rather than misclassifying them. These components are embedded in a staged pipeline of successive, complementary checks. We evaluate how synthetic augmentation affects classification and localization on a test set of real defects, and examine the system's trustworthiness at three points: the decision, the synthetic data, and the pipeline structure. This work-in-progress reports preliminary results suggesting that diffusion-generated defects, combined with uncertainty-aware classification, can lower the cost of reaching a trustworthy, deployable inspection model under data scarcity.
1 Introduction
Manufacturing inspection remains difficult to automate because defects are scarce and automated predictions may be confidently wrong outside the training distribution. The paper therefore combines synthetic data, staged inspection, and uncertainty-aware classification to support selective human intervention.
- Manual final inspection creates bottlenecks through labor cost, inspection time, and inter-operator variability.
- Defect scarcity limits supervised learning, while unsupervised detection avoids defect annotations but typically cannot categorize defects.
- The proposed pipeline combines region-of-interest segmentation, unsupervised anomaly detection, and uncertainty-aware defect classification.
- Synthetic defect generation is evaluated for its impact on defect classification and localization using real production defects.
- Bayesian neural networks and deep ensembles are assessed for predictive performance and calibration in low-data regimes.
2 Related work
Prior work addresses scarce defect data through synthetic generation and defect detection through supervised or unsupervised learning. The paper builds on diffusion-based generation and uncertainty estimation to support classification under limited coverage.
- GAN-based synthetic defects have improved recognition when real defect samples are limited, including in production settings.
- DualAnoDiff jointly generates full images and anomaly regions, producing pixel-aligned masks that support classification and localization.
- Supervised methods can detect and classify defects but require costly labeled defect data, whereas unsupervised methods learn from defect-free samples and flag deviations.
- Epistemic uncertainty is important in inspection because sparse training coverage can produce overconfident wrong predictions under distribution shifts.
- The pipeline combines synthetic-defect training with EfficientAD anomaly detection and uncertainty-aware classification.
3 Proposed pipeline
The proposed system separates offline model preparation from online inspection. It generates synthetic defects and trains complementary models before applying progressive checks that can defer uncertain cases.
- The offline stage generates synthetic defect images and trains the relevant inspection models, while the online stage applies progressive checks to incoming units.
- DualAnoDiff generates synthetic defects from 50 real images across fingerprint, interrupted_print, and smudged categories.
- DualAnoDiff jointly renders a complete anomalous object and isolated defect region, keeping the synthetic anomaly and mask consistent through shared attention.
- EfficientAD operates on segmented chassis images, trains exclusively on defect-free samples, and produces an anomaly map and scalar anomaly score.
- Defect classification estimates aleatoric and epistemic uncertainty using deep ensembles and an EfficientNet-B3 Bayesian approximation.
- Synthetic samples are illustrated with their corresponding generated masks.
4 Experiments and results
Preliminary experiments compare Bayesian and deep-ensemble classifiers under limited data and measure calibration, uncertainty, latency, throughput, and pipeline deployment characteristics. The Bayesian approach shows stronger calibration and multi-class performance, while the pipeline’s total latency supports online use.
- 1.000 accuracy was achieved by both classifiers on the 16-image binary test set across all experimental configurations.The authors caution that one error would materially reduce this estimate because the test set is small.
- Up to two orders of magnitude lower Brier and ECE values gave the Bayesian classifier better calibration than the deep ensemble.
- 0.833 accuracy and 0.830 F1-score at 500 images per class, together with 0.977 versus 0.930 AUROC, favored the Bayesian approach on the three-class task.
- Deep Ensemble had lower latency and higher throughput, whereas Bayesian inference incurred overhead from Monte Carlo sampling and its architecture.
- The pipeline’s total latency makes it well-suited for online applications and production-line deployment.
- Table 1 compares Deep Ensemble and Bayesian models across dataset configurations, with calibration and uncertainty metrics multiplied by stated factors.
5 Conclusion and future work
The work-in-progress pipeline addresses scarce defect data and untrustworthy inspection decisions through synthetic generation and uncertainty-aware human deferral. Preliminary offline results suggest improved classifier accuracy and calibration, while end-to-end online evaluation remains future work.
- Synthetic generation addresses defect-data scarcity at near-zero marginal cost relative to collecting rare real defects.
- Preliminary results suggest synthetic augmentation improves classifier accuracy and calibration.
- Uncertainty estimation lets the pipeline defer decisions it cannot trust to human operators while preserving routine automation.
- The reported results are preliminary and limited to the offline stage; ongoing work will evaluate the end-to-end online pipeline.