Transfer learning strategies comparisons of pre-trained neural model for tunnel flaws detection

Innovative ground-breaking data-driven and machine learning (ML) based solutions are currently revolutionizing the civil engineering field. Regarding the structural health monitoring (SHM) area, they may boost the reliability, capabilities, and productivity of field surveys and indirect inspections. Therefore, a deeper understanding of ML-based engineering applications is still relevant and essential to achieving these goals soon. The indirect inspection of road tunnels with ground penetrating radar (GPR) procedure for an automatic flaw classification is discussed in the current study. Specifically, two convolutional neural networks (CNN) based models (i.e. the ResNet-50 and the EfficientNet-b0) are compared using their pre-trained version and implementing three different transfer learning (TL) strategies. As expected, a proper fine-tuning method, by definition, would slightly reduce the classification performances because it relies on general-purpose features and activation maps. Nonetheless, it should provide the advantage of a lighter computational effort during the training process. On the other hand, a complete domain-adaptation strategy demonstrated that it might not be the ideal alternative for the best accuracy, anyway delivering the highest training computational costs among transfer learning approaches, sometimes comparable with the model’s training from scratch. In conclusion, as is often the case in general in engineering, the best compromise when dealing with pre-trained neural models for tunnel flaw classification could be adopting an effective training scheme for achieving the best trade-off for both training computational burden and classification accuracy, directly impacting on the survey of safety levels of the tunnels.