Transfer learning fails in the face of metal cracks; this case shows that production-level ML cannot rely solely on ImageNet pretraining, as the domain gap truly exists.

AIMPACT reports that aerospace manufacturing plants face cold start issues with data scarcity during quality inspection: only three photos of turbine blade microcracks are needed to achieve 99% accuracy. Even with ResNet-50 pretrained on ImageNet, small samples struggle to identify microcracks on metal surfaces, revealing the few-shot dilemma in production-level machine learning: transfer learning often fails, potentially leading to batch defects shipped, material waste, and downstream risks.