Intelligent Deep Learning Framework for Accurate Crack Detection and Width Measurement

Abstract

The automated crack detection and width measurements in concrete structures are important to structural health monitoring, but the manual inspection of crack is inconsistent, and the traditional image processing methods are difficult to adapt under different illumination environments and conditions. In this study, a novel hybrid deep learning network is designed using the convolutional neural network and transformer networks to effectively identify and segment cracks on complex concrete surfaces under different scenarios. For segmented images, the procedure of skeletonization returns the center of the crack for the whole image allowing measurements of the crack width at multiple positions across the width of the crack that are perpendicular to the centerline. An FOV based calibration step follows, which translates the measurements as obtained from the pixels to accurate real-world metric measurements. Its combination of deep-learning, skeleton extraction and calibration allows for fully-automated quantitative crack characterization. Experimental results are shown with high accuracy and reliability in crack detection, segmentation and width estimation compared with classic techniques which suggest the capability for real-time and automatic inspection of infrastructures through automation provided by CSE.