Cross-Scenario Industrial Defect Detection via Domain-Robust Training and Metric-Aligned Selection
Industrial board inspection must remain reliable across unseen lighting, surface contamination, and production-line conditions, yet training data is typically drawn from limited source scenarios. This paper treats the IDA 2026 Track 1 benchmark as a domain generalization problem and argues that standard mAP-driven checkpoint selection misaligns with the screening-heavy deployment objective. The authors propose a single-model YOLOv11-based framework with four modules—appearance-robust training (MixRand and IBN-a), small-defect preservation (SPD, SimAM, and NWD), confidence calibration (VFL and threshold search), and metric-aligned checkpoint averaging (a custom Track 1 proxy plus SWAD)—to improve cross-scenario defect detection and clean-board screening without ensembles or external data.
The paper presents a pragmatic competition-oriented system whose strongest contribution is the analysis and remediation of checkpoint-selection misalignment. The proposed Track 1 proxy and SWAD-based averaging yield the clearest measured gains, while several other modules are motivated by prior work but not independently validated here. The authors are commendably transparent about the limitations of their local ablation protocol and the training budget constraints that prevent full component-wise verification.
The central argument—that default mAP-based fitness is poorly correlated with the screening-aware Track 1 objective—is well supported by training dynamics and ablations. Specifically, replacing the default detector fitness with the custom proxy $\hat{S}_{\text{Track1}}$ and averaging checkpoints via SWAD yields the largest measured improvement under the unified local protocol.
A major limitation is the lack of independent ablations for most architectural additions. The authors explicitly state that "the training budget allocated for this paper does not permit including all of them in the main table," so IBN-a, SPD, SimAM, NWD, and VFL are treated as fixed recipe components rather than independently validated sources of gain. Similarly, MixRand contributes only a marginal composite improvement ($0.632$ versus $0.630$) on the local split. A deeper concern is the severe local-to-test distribution shift: $S_{\text{loc}}$ collapses from $0.826$ locally to $0.489$ on the challenge test, which undermines confidence in the local protocol for comparing modules.
The evidence supports the checkpoint-averaging claim most clearly, but the paper reports only its own test score without any direct competitor or baseline comparison, making it impossible to judge relative standing on the leaderboard. Related work coverage is adequate for the brief format, and the authors appropriately caution that local validation is an imperfect proxy for official challenge test performance, as shown by the large local-to-test discrepancy in Table I.
Reproducibility is partially strong: the authors release code, report key hyperparameters, and use a standard YOLOv11-large backbone. However, the full recipe contains many coarsely validated moving parts, the 2000-epoch schedule has no matched ablation, and the severe local-test discrepancy means independent researchers cannot reliably reproduce the test score without access to the hidden test set.
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.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.