Industrial defect detection models often fail when lighting, surface contaminants, or production line conditions change between deployment sites. The authors frame this as a domain generalization problem, where a model must detect, localize, and classify defects on unseen scenarios while keeping false alarms on clean boards low.

Their proposed method for the IDA 2026 Track 1 challenge consists of four complementary modules: cross-scenario robust training, small-defect-aware learning, score calibration with inference threshold selection, and metric-aligned checkpoint selection. The training module introduces cross-scenario robust normalization and DSU-style statistic perturbation with progressive random convolution. For small defects, the method employs NWD and parameter-free spatial attention, along with the VFL learning objective.

The final module uses a custom SWAD variant driven by a Track 1-oriented proxy to select checkpoints. On the test dataset, the model achieved a Track 1 score of 0.564 and a screening score of 0.696, running on a single RTX 4090 GPU. The paper, accepted for IEEE ICME 2026, includes code, but this summary relies solely on the arXiv source and does not compare with other approaches.