Source-linked AI summary
MAOL: Morphology-Aware Ordinal Learning for Fine-Grained Industrial Defect Severity Grading
Zhaoyang Wang, Haiyong Chen, Binyi Su, Kun Liu, Kun Wang, Xianen Zhou, Atik Shahariar
TL;DR
Fine-grained industrial defect severity grading must account for ordered labels, morphology-related cues, and noisy detector-generated instances. MAOL addresses these issues with morphology-aware ordinal learning, adaptive class-conditional thresholds, and localization-perturbation training. It consistently outperforms rule-based, classification, and ordinal baselines across clean-ROI and predicted-instance settings, ranking third in the IDA 2026 Challenge.
Problem
Industrial severity grading is challenging because labels are ordered, morphology-sensitive, and evaluated on noisy predicted instances rather than perfectly aligned training ROIs.
Method
MAOL combines instance-level ordinal learning, explicit morphology descriptors, class-conditional adaptive ordinal thresholds, and prediction-aware localization perturbation.
Results
MAOL consistently outperforms rule-based methods, classification-based models, and existing ordinal baselines under clean-ROI and predicted-instance settings, achieving 0.7965 (S2 metric) and ranking third in the IDA 2026 Challenge.
Takeaways & Limitations
Explicitly modeling ordinal structure, morphology, category-dependent boundaries, and predicted-instance noise provides a unified approach for practical industrial severity grading.
Abstract
from arXiv · showhide
Fine-grained defect severity grading is essential for industrial inspection, yet remains challenging due to the ordinal nature of severity labels, the strong dependence on morphology-related cues, and the train-test discrepancy between clean annotated instances and noisy predicted instances in two-stage pipelines. We propose MAOL, a Morphology-Aware Ordinal Learning framework for fine-grained industrial defect severity grading. MAOL formulates severity grading as an instance-level ordinal learning task, incorporates explicit morphological features to enhance representation learning, introduces class-conditional adaptive ordinal thresholds to model defect-specific grading boundaries, and employs prediction-aware training via localization perturbation to improve robustness to imperfect predicted instances. Extensive experiments under both clean-ROI and predicted-instance settings demonstrate that MAOL consistently outperforms rule-based methods, nominal classification models, and existing ordinal baselines, especially in the predicted-instance setting. The proposed approach ranked third in the IDA 2026 Challenge on Fine-Grained Severity Grading for High-Precision Manufacturing.
I. INTRODUCTION
Fine-grained severity grading supports industrial quality decisions but is difficult because severity labels are ordered, morphology-dependent, and affected by noisy predicted instances. MAOL addresses these challenges with morphology-aware ordinal modeling, adaptive thresholds, and prediction-aware training, achieving strong results across clean and predicted ROI settings.
- Severity grading provides fine-grained quality information for screening, rework decisions, and production control.
- Rule-based grading relies on handcrafted morphology thresholds and performs poorly on complex severity distinctions.Under clean ROI, the rule-based method achieves 0.5597 Accuracy and 0.6421 Quadratic Weighted Kappa (QWK).
- Standard classification ignores ordered severity relationships, making adjacent-level and cross-level errors indistinguishable.
- Predicted ROIs introduce localization errors, background contamination, and incomplete defect regions, creating a train–test discrepancy for downstream grading.
- MAOL combines instance-level ordinal learning, explicit morphology descriptors, class-conditional adaptive thresholds, and localization-perturbation training.
- 0.935 Acc. and 0.9672 QWK are achieved by MAOL under clean ROI, while experiments report particularly significant improvements under predicted-instance conditions.The method also achieves a final score of 0.7965 (S2 metric) in the IDA 2026 Challenge, ranking third among participants.
II. RELATED WORK
Industrial defect research has emphasized detection and segmentation more than severity assessment. Existing severity methods commonly use handcrafted rules or classification, while explicit modeling of ordered severity labels remains limited.
- Industrial visual inspection research has advanced from handcrafted features to CNN- and Transformer-based detection, segmentation, and anomaly localization.
- Rule-based severity methods quantify area, shape, contrast, grayscale intensity, or defect regions using predefined formulas or thresholds.These approaches depend heavily on human experience and generalize poorly to complex appearance variations.
- Most existing severity studies formulate grading as standard classification or application-specific scoring, with limited explicit treatment of ordered labels.
- This work targets instance-level, fine-grained industrial severity grading while additionally addressing predicted-instance noise in two-stage systems.
B. Ordinal Regression and Ordered Visual Recognition
Ordinal regression models the ordering of severity labels more directly than standard classification or distance-aware classification. MAOL builds on CORAL by adding morphology-aware representation, category-dependent grading boundaries, and prediction-aware training.
- Ordinal regression distinguishes adjacent-level errors from larger cross-level mistakes by explicitly modeling label ordering.
- Distance-aware classification incorporates class-distance-dependent penalties, whereas CORAL uses cumulative thresholds and ordinal binary targets.
- CORAL enforces rank consistency across ordinal thresholds so predictions follow a valid monotonic ordering.
- MAOL extends ordinal learning with morphology-aware representation, class-conditional grading boundaries, and prediction-aware training for practical deployment.The framework operates on detector-segmented instances and downstream ROIs in a two-stage inspection pipeline.
A. Overall Framework
MAOL treats defect severity as an instance-level ordinal prediction task using ROI appearance, morphology, and category information. Its adaptive CORAL head combines shared severity ordering with category-specific decision boundaries.
- Instance-level ordinal grading: Each detected defect instance is cropped to a 64×64 ROI and assigned one of four ordered severity levels.The levels are Acceptable, Marginal NG, NG, and Gross NG.
- Instance-level ordinal grading: MAOL processes ROI appearance with ResNet18 while extracting explicit morphology descriptors in parallel.The framework then fuses these features with defect-category information before ordinal prediction.
- Adaptive CORAL head: Severity labels are converted into K−1 ordinal binary targets, preserving their ordering across thresholds.For four levels, the encodings are 0→[0,0,0], 1→[1,0,0], 2→[1,1,0], and 3→.
- Adaptive CORAL head: The adaptive CORAL head combines a shared latent severity axis and globally shared thresholds with class-conditional threshold offsets.The offsets are generated from the defect-category embedding, allowing category-dependent grading boundaries.
- Adaptive CORAL head: During inference, the predicted severity is obtained by counting passed ordinal thresholds, while training uses binary cross-entropy over rank logits.This design models shared ordering alongside category-adaptive decision boundaries.
C. Morphology Branch and Class Embedding
The morphology branch supplements ROI appearance with explicit structural cues, while a learnable class embedding supplies category context. Feature fusion combines these signals for more reliable ordinal severity grading.
- Motivation: ROI appearance may miss severity-related structural cues because resizing can distort geometric and intensity properties.Relevant properties include defect size, shape, and contrast.
- Morphology branch: MAOL extracts morphology descriptors in parallel with ResNet18 ROI features and projects them into a low-dimensional embedding.The morphology representation captures defect size, geometry, and contrast.
- Class embedding: The defect category is mapped to a learnable embedding and incorporated into the fused instance representation.This adds category-level context alongside appearance and morphology features.
- Feature fusion: The fused representation jointly leverages appearance, scale-consistent morphology cues, and category-level context for ordinal severity grading.A learnable fusion module produces the final representation used by the grading framework.
D. Perturbation-aware Training
Two-stage inspection creates a distribution shift because grading uses detector-predicted ROIs with localization and content errors. MAOL simulates this discrepancy during training through constrained localization perturbations.
- Deployment discrepancy: Detector-predicted ROIs differ from aligned ground-truth regions through offsets, scale variations, truncations, and background contamination.These errors create deployment-time distribution shift for severity grading.
- Perturbation-aware training: With probability p = 0.5, MAOL randomly perturbs ground-truth box boundaries during training to produce realistic imperfect crops.The perturbation is constrained to an IoU range of 0.75–0.90.
- Perturbation-aware training: Perturbed boxes supply image-branch crops, while morphology features are computed from the original instance.This separates noisy appearance input simulation from the original-instance morphology computation.
A. Datasets And Experimental Details
MAOL is evaluated on an industrial defect benchmark using clean and detector-predicted ROIs, with the latter introducing localization-related difficulty. Across both settings, MAOL outperforms the compared grading methods, particularly under predicted ROIs.
- Dataset: The benchmark annotates each defect instance with a category, segmentation mask, and one of four ordered severity levels.The levels are Acceptable, Marginal NG, NG, and Gross NG.
- Dataset: The training split contains 924 images and 1992 instances, while validation contains 230 images and 554 instances.The severity distribution is imbalanced.
- Settings: Clean ROI uses ground-truth crops, whereas Predicted ROI uses crops from a first-stage detector and is more difficult because of localization errors.Morphology descriptors are also computed from detector outputs in the Predicted ROI setting.
- Results: +51.1 Acc., +46.0 Macro Acc., and +61.5 QWK are achieved by MAOL over the baseline under Predicted ROI.MAOL achieves the best results by a large margin in both Clean ROI and Predicted ROI settings.
2) Qualitative Analysis:
Qualitative comparisons show that detector-generated ROI contamination can undermine conventional grading methods, whereas MAOL remains effective on difficult predicted regions. Ablation findings identify morphology as the dominant contributor, with adaptive thresholds and perturbation training providing complementary gains.
- Qualitative comparison: Under predicted ROIs, MAOL correctly predicts Gross NG in an example with larger localization errors and ROI contamination.Most learning-based methods succeed when the predicted box is close to ground truth, while the rule-based method fails in that example.
- Ablation study: The morphology branch yields the largest ablation improvement, about +25 QWK, by adding explicit structural cues beyond ROI appearance features.
- Ablation study: Class embedding provides moderate gains by supplying category-aware context for adaptive threshold modeling.
- Ablation study: Adaptive thresholds further improve performance, while perturbation training reduces the train–test discrepancy between clean and predicted ROIs.
- Ablation study: The components are complementary: morphology provides the dominant gain, adaptive thresholds enhance ordinal modeling, and perturbation training improves robustness under noisy ROI conditions.
2) Analysis of Adaptive Thresholds:
Learned ordinal thresholds vary across defect categories rather than following a single global pattern. Their spacing also differs, indicating category-dependent intervals between severity levels.
- Threshold variation: Adaptive thresholds vary significantly across seven defect categories, unlike the category-agnostic thresholds shown for reference.
- Threshold variation: Relative threshold spacing differs between categories, indicating that ordinal intervals between severity levels are category-dependent.
- Conclusion: MAOL combines category-dependent severity boundaries with ordinal modeling, morphology-aware representation, and perturbation-based robustness training.
- Conclusion: The reported experiments show consistent improvements over rule-based, nominal classification, and standard ordinal baselines, especially for predicted instances.