Automatic segmentation and quantification of global cracks in concrete structures based on deep learning

Automatic detection and quantification of all cracks on the surface of concrete structures are of great importance for evaluating structural safety and damage. Firstly, based on the electric driving platform, a method of closerange scanning and shooting is presented to obtain high-resolution panoramas of the surface of concrete structures. Then, based on convolutional neural networks, an automatic segmentation method suitable for panorama cracks is proposed. Without considering complex scenes, the method achieved crack segmentation at 51.34 frames per second for 512 x 512 images, with 97.79% mean intersection-over-union. Finally, an automatic quantification method suitable for panorama cracks is given through proposed crack matching and property calculation methods. The properties and corresponding distribution of panorama cracks, including branch cracks, are calculated automatically. The average relative error of the maximum crack width value calculated by the proposed crack width calculation method is merely 3.87% (22.57 mu m), indicating high accuracy.

Automated summary

The article introduces a method for automatic detection and quantification of cracks on concrete structure surfaces. The method uses an electric driving platform for close-range scanning and shooting, applies convolutional neural networks for automatic segmentation, and employs crack matching and property calculation methods for automatic quantification. Experimental results demonstrate high accuracy of the method.