Optimising Concrete Crack Detection: A Study of Transfer Learning with Application on Nvidia Jetson Nano

Area of Science:

• Civil Engineering
• Computer Science
• Artificial Intelligence

Background:

• Infrastructure inspection relies heavily on manual labor, which is costly, time-consuming, and potentially hazardous.
• Automated defect detection using Artificial Intelligence (AI) offers a promising alternative for improving inspection efficiency and safety.

Purpose of the Study:

• To evaluate the effectiveness of various Convolutional Neural Networks (CNNs) for automated concrete crack detection.
• To assess the performance of these models when deployed on edge computing devices for real-time Structural Health Monitoring (SHM).

Main Methods:

• Trained six CNN models (Resnet18, Resnet50, GoogLeNet, MobileNetV2, MobileNetV3-Small, MobileNetV3-Large) using transfer learning on a dataset of 3000 concrete structure images.
• Augmented the dataset with salt and pepper noise and motion blur to enhance model robustness.
• Deployed the best-performing model on an Nvidia Jetson Nano for real-time inference.

Main Results:

• Resnet50 demonstrated the highest validation accuracy (96%) and F1-score (95%) with the original dataset and a batch size of 16.
• The trained model successfully performed real-time crack detection in both laboratory and field settings.
• Edge AI deployment proved effective for automated crack detection.

Conclusions:

• Transfer learning combined with edge AI devices offers a cost-effective and efficient solution for automated crack detection in concrete structures.
• AI-powered SHM systems can significantly enhance the safety and reduce the cost of infrastructure inspections.