Real-time detection algorithm of tunnel cracks based on GRU-CNN

Abstract

Objective Tunnel cracks seriously damage the corresponding life time and traffic safety. However, traditional manual detections cannot efficiently and accurately identify a large number of cracks in long tunnels.This paper proposes a real-time detection algorithm for tunnel surface cracks. Methods It innovatively applies the Gate Recurrent Unit (GRU) model for text learning and signal analysis to image classification, improving detection speed and ensuring detection accuracy of tunnel cracks. To enhance training efficiency, the cracks are preprocessed and converted into the frequency domain to extract the key information of tunnel cracks, and the matrix is reconstructed into one-dimensional vectors. Then, one-dimensional convolutional neural network is used to extract the vector depth feature, and recurrent neural networks can learn corresponding sequential dependencies to realize tunnel cracks detection. Results Test results show that this model can reduce the number of training parameters and hardware configuration requirements. At the same time, the detection accuracy can reach 98.8%, and the detection speed for single image can reach in 2.1 s. Comparing with the mainstream classification detection algorithms, its accuracy remains unchanged, with significantly improvements of both training efficiency and prediction rate respectively. Conclusion Finally, a detection framework is developed for large-scale tunnel cracks to extract corresponding crack information effectively.

Keywords

• tunnel crack
• real-time detection
• classification detection
• frequency domain