Deep learning algorithm for real-time automatic crack detection, segmentation, qualification

Abstract

Cracking is one of the typical damages in concrete structures, and it is crucial to detect and quantify cracks in a timely and efficient manner. However, current research primarily focuses on either single-task recognition or dual-task recognition based on multi-step sequential approaches. Less attention has been devoted to the multi-task integration of cracks. To address the challenges of inefficient and multi-step detection in traditional concrete crack detection methods, a novel deep learning-based model, called YOLOv5-IDS, is proposed based on You Only Look Once network v5 (version 6.2) with the combination of bilateral segmentation network while introducing a dilated convolution, pyramid pooling module, and attention refinement module. Moreover, crack parameter measurement algorithms based on the micro-element method are proposed to improve accuracy and efficiency. The method proposed in this study can not only detect and segment cracks with high accuracy and efficiency, but also quickly measure crack parameters, thus developing a complete method for the process from real-time crack detection and segmentation to crack parameter measurement. The experimental results for the YOLOv5-IDS model reveal the following performance metrics. For crack detection, the mean average precision with an intersection of union threshold of 0.5 (mAP@0.5) is 84.33%, and the frames per second (FPS) is 159 f/s. For crack segmentation, the mean intersection over union (mIoU) is 94.78%, and the FPS is 8 f/s, respectively. Compared to existing methods, the proposed approach exhibits improvements in both accuracy and efficiency. Moreover, the calculation of crack parameters proves to be both precise and rapid.