Deep learning-based crack detection in a concrete tunnel structure using multispectral dynamic imaging

摘要

A new computer vision-based method is proposed for concrete crack detection in tunnel structures using multi-spectral dynamic imaging (MSX). The MSX images were collected from a tunnel in the University of Manitoba, Canada. A total of 3600 MSX images (299 × 299 pixels) were used to train the modified deep inception neural network (DINN), and an additional 300 MSX images (299 × 299 pixels) were employed for validation purposes. The MSX images were examined by the trained neural network for concrete crack detection. The main purpose of this research was to examine the potential of the neural network to distinguish between noise and concrete surface cracks in the MSX images. A fully connected layer and a softmax layer were added to the DINN network in the transfer learning section to reduce the network computation cost. The proposed network used green bounding boxes to detect the portions with cracks in the MSX images. A training accuracy of 95.5% and a validation accuracy of 94% were achieved at 1600 iterations. The optimum training steps obtained from the training and validation were used for testing purposes. The robustness of the trained network was evaluated using an additional 96 MSX images (640 × 480 pixels). A maximum testing accuracy of 94% was recorded when the prediction probability was limited to 90%.