Crack size measurements on fracture surface images using deep neural networks for semantic segmentation

For the safe evaluation of nuclear-relevant safety components, a precise and reliable analysis of the fracture surfaces after the test procedure is required. Within the scope of the studies a framework for automated crack size measurements based on image segmentation has been developed, capable of accelerating the standardized but tedious measurement procedure. Different known image segmentation architectures have been trained and assessed based on a specially created fracture surface image dataset. The fracture surfaces originate from SE(B) specimens made of the German reactor pressure vessel steel 22NiMoCr3-7 and its weld material. The evaluation of the model performances via the mIoU metric show that the investigated architectures are very well suited for the pixel-fine classification of fracture mechanisms. Based on the obtained prediction masks, the initial crack size a0 could be measured using the so-called area average (AA) method. The results have been compared to manual 5-point average (5PA) measurements. The automated crack length measurements show statistically verifiable very high precision, comparability, and economic efficiency.