Source-linked AI summary
Cross-Scenario Industrial Defect Detection via Domain-Robust Training and Metric-Aligned Selection
Dikai Li, Weijia Wang, Yuan-Gen Wang
TL;DR
Cross-scenario industrial inspection must detect defects and screen clean boards under unseen production conditions. The paper combines domain-robust training, small-defect learning, score calibration and threshold selection, and metric-aligned checkpoint averaging. On official challenge test feedback, it reports a Track 1 score of 0.564 and a screening score of 0.696.
Problem
Industrial defect detection must remain reliable under unseen scenario conditions while locating defects and correctly screening clean boards.
Source: §I · p. 1Method
A single-model framework combines appearance-robust training, small-defect preservation, screening-aware calibration, and metric-aligned checkpoint averaging.
Source: §I · pp. 1–2Results
0.564 Track 1 score and 0.696 screening score were reached on official challenge test feedback.
Source: §I · p. 2Takeaways & Limitations
Abstract
from arXiv · showhide
Automated defect detection of industrial boards must remain highly reliable under cross-scenario variations in lighting, surface contaminants, and production line conditions. This makes cross-scenario defect detection a domain generalization problem, wherein models are required to detect, localize, and classify defects across unseen scenarios while minimizing false alarms on clean boards. For the IDA 2026 Track 1 challenge, we propose a cross-scenario defect detection method that includes four complementary modules: cross-scenario robust training, small-defect-aware learning, score calibration and inference threshold selection, and metric-aligned checkpoint selection. These four modules respectively introduce cross-scenario robust normalization and DSU-style statistic perturbation with progressive random convolution, NWD and parameter-free spatial attention, the VFL learning objective, and a custom SWAD variant driven by the Track 1-oriented proxy. Experimental results demonstrate that our model achieves a Track 1 score of 0.564 with a screening score of 0.696 on the test dataset on a single RTX 4090 GPU. The code is available at https://github.com/GZHU-DVL/Dikai.
I. INTRODUCTION
Cross-scenario industrial defect detection is a domain-generalization problem: systems must find defects and correctly screen clean boards under unseen conditions. The paper proposes a four-part framework and reports a Track 1 score of 0.564 with a screening score of 0.696.
I. INTRODUCTION
- Cross-scenario inspection must handle unseen lighting, contamination, and production conditions while locating defects and correctly screening clean boards.Track 1 combines localization, classification, and screening, with screening carrying the largest weight and greatest sensitivity to domain shift.Source: §I · p. 1
- The framework combines appearance-robust training, small-defect preservation, screening-aware calibration, and metric-aligned checkpoint selection.The components use IBN-a and MixRand; small-defect learning; confidence calibration and threshold selection; and a Track-1-oriented proxy with SWAD averaging.Source: §I · pp. 1–2
- Metric-aligned SWAD selection was the clearest measured contributor under a unified local protocol, while composite feature perturbation mainly acted as a stability regularizer.The paper argues that localization-dominated selection can be misaligned with cross-scenario screening reliability.Source: §I · pp. 1–2
- 0.564 Track 1 score and 0.696 screening score were achieved on official challenge test feedback with one RTX 4090 GPU.The reported setup used no ensemble, external data, pseudo-labels, or test-time augmentation.Source: §I · p. 2
II. RELATED WORK
Related work spans industrial anomaly benchmarks, domain-generalization methods, and techniques for small-defect localization and screening. IDA Track 1 differs by evaluating industrial boards under unseen production conditions and explicitly including image-level screening reliability.
Industrial anomaly and defect benchmarks
- IDA Track 1 evaluates industrial-board inspection under unseen production conditions and includes image-level screening reliability, unlike the described anomaly benchmarks.MVTec AD targets unsupervised anomaly detection, while VisA extends industrial anomaly detection to more complex visual categories and detection-oriented evaluation.Source: §II · p. 2
Domain generalization for cross-scenario detection
- Domain-generalization approaches reduce dependence on domain-specific appearance statistics through feature-statistic mixing, uncertainty modeling, random convolutions, normalization, or weight averaging.The cited approaches include MixStyle, DSU, RandConv, IBN, and SWAD.Source: §II · p. 2
Tiny defect detection, attention, and score calibration
- Small-defect methods address localization and feature refinement, while Varifocal Loss targets confidence calibration for screening-oriented inspection.NWD [10] provides a smoother localization signal for tiny objects; SPD preserves spatial detail, and SimAM [12] supplies parameter-free spatial attention.Source: §II · pp. 2–3
III. PROPOSED METHOD
The method combines four modules in a YOLOv11-large segmentation backbone to improve cross-scenario defect detection. Its Cross-Scenario Robust Module uses persistent normalization and training-time stochastic perturbations of feature statistics and local textures.
III. PROPOSED METHOD
- Four modules target appearance robustness, small-defect recognition, score calibration and screening, and objective-aligned checkpoint selection in a YOLOv11-large segmentation backbone.Stochastic perturbations operate during training; inference uses one exported checkpoint without ensembling or test-time augmentation.Source: §III · p. 3
A. Cross-Scenario Robust Module
- IBN-a uses a 0.6 ratio in the stem and stage 1 and remains active during both training and inference as a standing domain-invariance prior.The Cross-Scenario Robust Module combines IBN-a with MixRand, which incorporates DSU and progressively random convolution.Source: A. Cross-Scenario Robust Module · p. 3
- MixRand combines DSU feature-statistic perturbation with random depthwise convolutions whose probability and strength increase over the training schedule.The perturbation strengths are attenuated to remain within a stable regularization range, with a tail decay during the final close_mosaic window.Source: A. Cross-Scenario Robust Module · p. 3
B. Small-Defect-Aware Module
The small-defect module preserves fine spatial cues and adds a denser localization signal for tiny boxes. The calibration and threshold-selection module targets false alarms and chooses inference settings against the composite Track 1 metric.
B. Small-Defect-Aware Module
- SPD adapters preserve fine spatial detail, SimAM adds parameter-free spatial attention, and NWD combined with CIoU provides a dense localization signal for small boxes.The module aims to preserve defects that downsampling and sparse IoU gradients can obscure, without modifying the detection head.Source: B. Small-Defect-Aware Module · p. 3
C. Score Calibration and Inference Threshold Selection
- VFL replaces BCE for score calibration and defect classification to suppress high-confidence false alarms on clean boards.The authors report that BCE spills probability mass onto benign textures, penalizing Sscreen.Source: C. Score Calibration and Inference… · p. 3
- Inference settings are selected by scanning confidence, IoU, and maximum-detection thresholds with the official evaluator and maximizing local-split STrack1.The queried long-schedule checkpoint used conf = 0.08, iou = 0.30, and max det = 50.Source: C. Score Calibration and Inference… · p. 3
D. Metric-Aligned Checkpoint Selection
The section replaces trainer fitness with a Track 1-oriented proxy for checkpoint ranking and SWAD selection. The proxy combines box and mask metrics, with localization emphasizing mask mAP50-95, classification using mAP50, and screening weighting recall slightly more than precision.
D. Metric-Aligned Checkpoint Selection
- The authors define an approximate detection score ˆSTrack1 as an alternative to the official detection score; Ultralytics fitness, used for checkpoint selection, is dominated by mAP50-95 and correlates more strongly with defect localization than screening efficiency SScreen.Source: D. Metric-Aligned Checkpoint Selection · pp. 3–4
- Precision and recall are averaged across box and mask branches as a metric-level heuristic, rather than by instance-level logical OR fusion.The section defines these as branch-averaged precision and recall.Source: D. Metric-Aligned Checkpoint Selection · p. 4
- The proxy emphasizes strict localization with mask-dominant mAP50-95, uses the looser mAP50 criterion for classification, and gives recall slightly more weight than precision for screening.The stated rationale is that false negatives carry higher practical cost than false alarms in industrial defect detection.Source: D. Metric-Aligned Checkpoint Selection · p. 4
- The ˆSTrack1 proxy replaces default trainer fitness for epoch ranking and best-checkpoint selection, and guides a stable SWAD window whose checkpoints are averaged.The SWAD driver selects a flat region around the smoothed proxy peak (window w=5, tolerance ratio 0.20, at least two converged epochs).Source: D. Metric-Aligned Checkpoint Selection · p. 4
A. Settings
Experiments use only the released IDA Track 1 data at native 512×512 resolution, with a long-schedule queried checkpoint and a common official-style local evaluation protocol. The section distinguishes held-out validation scores from official challenge test feedback.
A. Settings
- The model is trained only on released IDA Track 1 data, without external samples, pseudo-labels, ensembles, or test-time augmentation, using native 512×512 images.Training uses AdamW with initial learning rate 3×10−4 and conservative color, flip, scale, translation, and copy-paste augmentation.Source: A. Settings · p. 4
- The queried long-schedule checkpoint trains for 2000 epochs; targeted ablations use a unified 600-epoch budget, with local results evaluated through the official-style scoring pipeline.Unless stated otherwise, inference uses swad_best.pt with conf=0.08, IoU=0.30, and max_det=50.Source: A. Settings · p. 4
- Local scores come from a held-out validation split, whereas official challenge test scores are feedback from submitted checkpoints.The reported score set includes localization, classification, screening, and the composite Track 1 score; Table I covers local re-evaluation and official test feedback for the queried checkpoint.Source: A. Settings · p. 4, Table I · p. 4
B. Main Result and Local-Hidden Score Analysis
The queried checkpoint’s local and official composite scores are close, but their component balances differ; under the unified local ablation protocol, removing SWAD causes the largest measured drop. MixRand changes are small, and several recipe components were not independently validated.
B. Main Result and Local-Hidden Score Analysis
- 0.596 and 0.564 are the close local and official composite scores, but local results favor localization while official feedback shows a stronger image-level screening signal.The official test feedback was used only as test-time evidence; component conclusions came from independently trained short-schedule variants under a unified local protocol.Source: B. Main Result and Local-Hidden Score… · p. 4
C. Ablation Study
- In the comparable 600-epoch ablation set, disabling MixRand changed composite score only marginally (0.632 vs. 0.630), while unattenuated MixRand achieved a similar 0.631 result.Source: C. Ablation Study · p. 4
- 0.632 to 0.614 is the composite-score drop when SWAD averaging is removed, driven mainly by lower Sscreen and Scls.The no-SWAD variant shows the largest component-level degradation in the comparable 600-epoch ablations.Source: C. Ablation Study · p. 5, Figure 2 · p. 5
- The training budget precluded main-table ablations for IBN-a, SPD, SimAM, NWD, and VFL, so their independent performance gains were not validated.The 2000-epoch queried checkpoint is a local re-evaluation reference, not a valid single-module comparison with the 600-epoch variants.Source: C. Ablation Study · pp. 4–5, Table II · p. 5
D. MixRand Decomposition
On the current local split, MixRand variants differ little from the full recipe, so the authors treat composite feature perturbation mainly as a stability regularizer. The training-dynamics figure contrasts the Track 1 proxy, its screening component, and default detector fitness, with SWAD’s window selected from the screening-weighted proxy.
D. MixRand Decomposition
- Because observed differences are small on the current local split, the authors interpret MixRand primarily as a stability-oriented regularizer that reduces sensitivity to source-domain appearance statistics, rather than as a large independent score gain.Source: D. MixRand Decomposition · p. 5
- The training-dynamics figure compares the smoothed Track 1 proxy, Pscreen, and default mAP-based fitness, and marks the SWAD window selected from the screening-weighted proxy.The shaded averaging window is not selected from default detector fitness.Source: Figure 3 · p. 5
- The DSU-only and RandConv-only variants were omitted from the main ablation table because of training-budget constraints, so their separate gains were not verified on this split.The composite design is motivated by prior MixStyle, DSU, and RandConv studies.Source: D. MixRand Decomposition · p. 5
E. Training-Schedule Effect
The Track 1 proxy, screening component, and default mAP-based fitness peak at different epochs, so the authors use a Track 1-oriented proxy to identify a flat region for checkpoint averaging. The no-SWAD control shows lower composite and screening scores under the same local protocol.
E. Training-Schedule Effect
- 0.632 to 0.614 in STrack1 and 0.629 to 0.602 in Sscreen quantify the no-SWAD drop under the same 600-epoch local protocol.The proxy, Pscreen, and default mAP-based fitness peak at different epochs; SWAD averages checkpoints from a flat, high-proxy plateau.Source: E. Training-Schedule Effect · p. 5
F. Qualitative Comparison
Figure 4 illustrates selected local-validation cases comparing the full recipe with a no-MixRand ablation. The conclusion reports official test feedback and notes limits to attributing component effects.
F. Qualitative Comparison
- The full recipe tends to suppress spurious detections on clean regions while maintaining responses to small defects in selected local-validation examples.The examples are selective, and quantitative claims rely on Table II component scores.Source: F. Qualitative Comparison · p. 5, Figure 4 · p. 6
V. CONCLUSION
- STrack1=0.564 with Sscreen=0.696 on official challenge test feedback, using one model without external data, an ensemble, or test-time augmentation.The result was obtained on one RTX 4090 GPU.Source: §V · p. 6
- Feature-perturbation gains are modest on the current split, fixed minor components lack independent ablations, and broader official feedback is needed to separate their effects.Source: §V · p. 6