Source-linked AI summary
FuDU: A Fuzzy Dual-dimensional Uncertainty Framework for Streaming Active Learning in Industrial Defect Detection
Zhaoyang Wang, Haiyong Chen, Binyi Su, Xinwei Lyu
TL;DR
Reliable industrial defect detection is difficult with scarce annotations, class imbalance, and changing operating conditions. FuDU combines prototype-based global uncertainty, dual-entropy defect uncertainty, and fuzzy inference for expert-informed streaming sampling. It achieves 96.0% mAP and 99.3% recall using 15.4% of samples, while relying on parameters initialized from safety standards and tuned on a pilot validation set.
Problem
Industrial defect detection must remain reliable despite scarce labeled defects, severe class imbalance, limited operating scenarios, and adversarial or outlier risks.
Method
FuDU jointly estimates global and defect uncertainty with PGUQ and DeUE, then uses fuzzy inference to produce expert-informed sampling decisions.
Results
96.0% mAP and 99.3% recall are achieved using 15.4% of samples, surpassing existing sampling strategies.
Takeaways & Limitations
FuDU offers a practical expert-informed strategy for reliable, annotation-efficient streaming defect detection.
Takeaways & Limitations
Fuzzy linguistic-variable parameters are initialized from nuclear inspection safety standards and fine-tuned on a pilot validation set.
Abstract
from arXiv · showhide
Ensuring the reliability of deep learning models in real-time industrial defect detection is critical for high-stakes quality inspection. To mine uncertain samples within continuous industrial media streams, thereby enhancing the reliability of the detection system, this paper proposes a streaming active learning method based on the Fuzzy Dual-dimensional Uncertainty (FuDU) framework. Specifically, we first design a Prototype-based Global Uncertainty Quantification (PGUQ) module on the backbone to evaluate image-level uncertainty via normal/defective feature prototypes. A Dual-entropy defect Uncertainty Evaluator (DeUE) is then integrated into the detection head to quantify box-level uncertainty. Finally, by modeling uncertainty as systematic error, we propose a fuzzy dual-dimensional uncertainty-aware strategy that leverages fuzzy inference to fuse dual-dimensional uncertainties, enabling expert knowledge-driven adaptive sampling decisions. Comprehensive experiments demonstrate that FuDU is efficient and flexible, making it well-suited for challenging industrial inspection tasks such as the detection of nuclear fuel rod defects. Our code is publicly available at: https://github.com/wangzhaoyang-508/FuDU.
1 Introduction
Industrial defect detectors face adversarial risks near decision boundaries and outlier risks from novel defects or domain shifts. FuDU addresses these risks with dual-dimensional uncertainty estimation and fuzzy, expert-informed sampling, achieving strong accuracy and annotation efficiency.
- Motivation: Industrial inspection models require large, well-annotated datasets, yet deployed streams contain scarce defects and distribution shifts.These conditions expose models to decision-boundary ambiguity and out-of-distribution samples.
- Motivation: Active learning selects high-risk uncertain samples for expert annotation to iteratively optimize decision boundaries and improve domain generalization.The approach directly targets uncertain samples during inference rather than relying only on static training data.
- FuDU framework: FuDU combines PGUQ for image-level outlier uncertainty with DeUE for box-level adversarial uncertainty.PGUQ uses normal and defect feature prototypes, while DeUE uses localization and classification entropy.
- FuDU framework: Fuzzy inference maps global and defect uncertainties to linguistic variables and sampling decisions, allowing experts to interpret, validate, and calibrate the strategy.This design integrates domain expertise into adaptive streaming sample selection.
- Results: 99.3% recall is achieved with uncertainty-based sampling, while annotation cost falls below one-sixth of manual screening and mAP exceeds confidence-based sampling by 7.3%.The results also include validation on the public ELES dataset for cross-scenario generalization.
- Results: FuDU supports reliable active learning for high-stakes industrial visual inspection.Its stated application scope includes nuclear fuel rod defect detection.
2 Related Work
Industrial defect detection remains difficult because class imbalance, limited operating scenarios, and scarce labeled data constrain model reliability. Related work spans anomaly, adaptation, incremental, uncertainty, diversity, and fuzzy methods, but streaming industrial settings require practical uncertainty-guided selection.
- Industrial Defect Detection: General-purpose detectors face severe class imbalance and limited operational scenarios, while existing methods struggle to address adversarial and outlier risks simultaneously.These risks motivate specialized approaches for industrial inspection.
- Industrial Defect Detection: One-class anomaly detection uses normal prototypes and feature distances, but lacks defect categorization and is sensitive to thresholds.FuDU uses this paradigm as inspiration for PGUQ rather than adopting it unchanged.
- Active Learning: Object-detection active learning commonly uses uncertainty- or diversity-based sampling, including MC dropout, query-by-committee, evidential learning, clustering, and coreset selection.Most such methods are tailored to pool-based settings requiring offline processing.
- Fuzzy Control: Fuzzy control represents uncertainty with linguistic variables and IF-THEN rules, enabling interpretable fusion of heterogeneous uncertainty metrics and incorporation of expert knowledge.This is particularly relevant when black-box decisions are unacceptable in safety-critical inspection.
3 Methodology
FuDU combines prototype-based image-level uncertainty, dual-entropy box-level uncertainty, and fuzzy inference to assign adaptive sampling probabilities in streaming industrial inspection. The framework updates both detector and prototype library with annotated uncertain images after each detection batch.
- Framework overview: FuDU produces image-level uncertainty U_g from PGUQ and defect-level uncertainty U_d from DeUE, then fuses them to determine each image’s sampling probability.After each detection batch, collected uncertain images are annotated and used to update the detector and prototype library.
- Prototype-based global uncertainty: PGUQ estimates global uncertainty from distances between an image feature prototype and normal and defective prototype sets, with higher U_g indicating a low-density region.Prototype sets are initialized with K-means and treated as learnable during active learning to accommodate domain variation.
- Prototype-based global uncertainty: The PGUQ contrastive loss pulls features toward their ground-truth prototype and pushes them from the opposite prototype, while dispersion regularization prevents prototype collapse.The total training loss combines detection, contrastive, and dispersion losses, with λ1 = 1.0 and λ2 = 0.1.
- Fuzzy uncertainty-aware sampling: Fuzzy inference maps normalized U_g and U_d to linguistic uncertainty levels and sampling actions ranging from Do Not Sample to Must Sample.The safety-first rule base assigns Must Sample when either uncertainty is Very High and uses Mamdani fuzzification, rule activation, and aggregation.
- Fuzzy uncertainty-aware sampling: FuDU uses fuzzy logic to convert quantitative uncertainty into expert-rule-based sampling actions that balance annotation efficiency with inspection reliability.The strategy prioritizes false-negative and false-positive samples expected to maximize detector gains while reducing sampling of confident normal samples.
4 Experimental evaluation
Experiments evaluate FuDU on streaming nuclear fuel rod inspection data through baseline comparisons, uncertainty ablations, sampling-strategy tests, and flexibility and generalization studies. FuDU combines strong detection performance with limited annotation and stable adaptation across settings.
- Experimental setup: The evaluation uses 2,000 unlabeled images in four streaming batches, with separate labeled training and held-out test subsets covering four defect categories.Experiments repeat streaming order and batch partitioning with five random seeds and report mAP50, precision, recall, F1, and sampling ratio.
- Baseline comparison: Random sampling improves each round by less than 1 mAP with precision and recall below 20%, while confidence-based methods still miss false negatives.QBC and SPENet achieve high recall, whereas PPAL underperforms because of limited scenario adaptability; Oracle is costly because it labels every image.
- Global uncertainty: k = 50 yields the highest mean mAP of 92.2% with a 15.4% annotation ratio, balancing feature-space coverage and prototype updateability.k = 10 produces 92.0% mean mAP with 40.5% annotations, whereas k = 100 produces 88.5% mean mAP with 9.8% annotations.
- Global uncertainty: Static learnable prototypes raise mean mAP from 90.8% to 92.2% and reduce annotation ratio from 21.5% to 15.4% compared with re-clustering after each round.The authors associate prototype stability with better adaptation to domain shifts and avoidance of distribution oscillations.
- Defect uncertainty: The default DeUE weighting w1 = 1, w2 = 2 delivers 92.2% mean mAP with a 15.4% annotation ratio, while over-weighting localization entropy raises annotation cost.Using w1 = 1, w2 = 4 maintains 92.1% accuracy but increases the annotation ratio to 17.3%.
- Sampling strategies: 96.0% mAP and 99.3% recall are achieved with 15.4% annotated samples, outperforming the compared sampling strategies under identical settings.Hard thresholds are sensitive to rigid partitions, while fuzzy rules provide finer-grained sample ranking; Sigmoid reaches 97.6% recall but uses 23.4% annotations.
- Flexibility and generalization: FuDU maintains stable detection performance when the stream is divided into ten 200-image batches and generalizes across detector architectures and the public ELES dataset.Cross-architecture experiments compare FuDU-based and confidence-based active learning, while ELES transfer retains the fuzzy rule base and sampling-probability mapping.
5 Conclusion
FuDU addresses reliable defect detection with limited annotations through stream-based active learning that combines global and defect uncertainty with expert-informed fuzzy sampling. It reports 96.0% mAP and 99.3% recall using 15.4% of samples, supporting practical real-time quality inspection.
- Conclusion: FuDU combines PGUQ and DeUE within a fuzzy dual-dimensional uncertainty framework for expert-informed sampling under limited annotations.PGUQ assesses image-level uncertainty, while DeUE evaluates defect-level uncertainty.
- Conclusion: 96.0% mAP and 99.3% recall are achieved while requiring only 15.4% of samples, surpassing the compared strategies.The conclusion presents this result as evidence of FuDU’s state-of-the-art detection performance at reduced annotation usage.
- Conclusion: FuDU offers a practical pathway for real-time quality inspection systems to move from passive updating toward active self-evolution.This consequence is stated within the paper’s industrial defect-detection scope.