From Synthetic Cracks to Real Inspection: Data- and Model-Level Control of Domain Shift

Synthetic data offers a practical solution to data scarcity in visual inspection, but models trained on synthetic images often suffer from severe performance degradation when deployed on real data due to domain shift. This paper addresses synthetic-to-real generalization for crack segmentation by jointly controlling domain shift at the data and model levels. At the data level, cracks are synthesized and refined using unpaired appearance adaptation while explicitly preserving defect geometry to suppress refinement-induced distorted appearance. At the model level, a stability-divergence-guided early stopping strategy is introduced to select checkpoints based on representation convergence rather than validation performance. Experiments on real crack images demonstrate that neither refinement nor training control alone is sufficient, whereas their combination yields consistent and robust improvements in real-domain performance. The results highlight the importance of coordinated data- and model-level strategies for reliable synthetic-to-real deployment in industrial inspection.