Recognition of Tunnel Cracks Based on Deep Convolutional Neural Network Classifier.

Min Yang,Qing Song,Xueshi Xin,Lu Yang

DOI: https://doi.org/10.1007/978-981-13-2203-7_55

2018-01-01

Abstract:It is of vital importance to obtain timely identification and treatment of tunnel cracks in ensuring railway traffic safety. As the amount of tunnel images collected by high-speed cameras equipped on the train is extremely large, manual method is unable to meet the actual needs in recent years while traditional image processing methods become powerless to achieve a satisfying speed and accuracy. In this paper, we propose a novel algorithm based on deep learning to solve this problem. We first use Simple Linear Iterative Clustering super-pixel algorithm to segment the original tunnel data and then artificially annotate the images. And eventually establish a dataset called CLS-CRACK for training and testing the deep convolutional neural network classification model. We design our classification network on the basis of ResNet18 and finish the experiments with one NVIDIA GeForce GTX Titan X on a framework named caffe. Finally, results demonstrate that the new algorithm performs well in the validation set with the result of a 94.0% accuracy and an 83.0% recall and time cost for each original tunnel image is within 0.025 s.