Automatic detection and measurement of ground crack propagation using deep learning networks and an image processing technique

• Published in: Measurement : journal of the International Measurement Confederation Vol. 215; p. 112832
• Main Authors: Pham, Minh-Vuong, Ha, Yong-Soo, Kim, Yun-Tae
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 30.06.2023
• Subjects: CNN-based image segmentation models; Ground crack monitoring; Image processing; Otsu method; Ground crack monitoring; CNN-based image segmentation models; Image processing; Otsu method
• ISSN: 0263-2241; 1873-412X
• DOI: 10.1016/j.measurement.2023.112832

•A novel method for automatic detection of ground crack propagation and its dimension measurement.•CNN-based image segmentation models was trained and tested on an image dataset obtained from laboratory slope tests and field crack images.•Models achieved average F1 score of 0.877 to 0.896, and a processing speed of 7.48 to 8.01 FPS.•The accuracy of the calculated dimension results of ground crack ranged between 86.7% and 99.9%. We present a novel approach for automatic identification of ground crack propagation and measurement of their dimensions using deep learning and an image processing technique. The different convolutional neural networks (U-Net, LinkNet, Feature Pyramid Network (FPN), and Deeplabv3) and a traditional image-processing technique based on the Otsu method were employed to identify ground cracks and calculate their lengths and widths on camera views. The deep learning models were trained and tested on an image dataset obtained from laboratory slope experiments and field crack images. Ground crack identified by the convolutional neural network-based image segmentation models afforded average F1 score of 0.877 to 0.896 and processing speed of 7.48 to 8.01 frame per second (FPS), thus better than those of traditional image processing (F1 score of 0.65 and speed of 0.01 FPS). In addition, the accuracies of the calculated crack dimensions on both images and videos ranged from 86.7 to 99.9%.