Transfer Learning-Based Algorithms for the Detection of Fatigue Crack Initiation Sites: A Comparative Study

Fatigue crack initiation sites (FCISs) identification is routinely performed in the field of engineering failure analyses, which is not only time-consuming but also knowledge-intensive, especially in large components. The emergence of convolutional neural networks (CNNs) has inspired numerous innovative solutions for image analysis problems in the material science field. As an explorative study, we previously trained a model to FCISs from scratch with a small amount of data. However, the results were not ideal for practical applications. In this study, based on the principle of transfer learning, we used three state-of-the-art CNNs, namely VGG-16, ResNet-101, and feature pyramid network (FPN), as feature extractors, and a faster R-CNN as the backbone to establish models for FCIS detection. Compared with our previous study, the results showed that the transfer learning-based models provide significant improvements in terms of the detection performance, with the accuracy and precision reaching up to 83.5 and 95.9%, respectively, at a confidence threshold of 0.6. Among the three models, the ResNet model exhibited the highest accuracy and lowest training cost. The performance of the FPN model closely followed that of the ResNet model with an advantage in terms of the recall. The VGG model outperformed the other models in terms of the detection time and memory requirement.