A deep learning framework for real-time multi-task recognition and measurement of concrete cracks

Abstract

This study presents an innovative deep learning framework, YOLO-DL, for automatic multi-task recognition of concrete cracks. The framework integrates the You Only Look Once (YOLO) object detection algorithm with the encoder-decoder architecture of the DeepLabv3 + model, incorporating an attention mechanism and a calibration module, resulting in three distinct branches for crack classification, localization detection, and semantic segmentation. The YOLO-DL model achieves a detection precision of 84.87 %, an [email protected] of 83.55 %, and a mean intersection-over-union (mIoU) of 94.94 % for crack segmentation. The model’s segmentation inference time is significantly shorter than that of the DeepLabv3+, fully convolutional networks (FCN), U-Net, and SegNet models, making it suitable for real-time concrete crack recognition. The model effectively handles classification, detection, and segmentation tasks, demonstrating enhanced performance and robustness, particularly with the inclusion of the attention mechanism. Additionally, a novel crack width measurement method based on the local element grid method is presented, achieving sub-pixel precision. This method provides comprehensive crack width information, including the maximum width of each crack and its corresponding location, with a maximum relative error of less than 10 %. The findings highlight the model’s strong inference performance, robust generalization ability, and promising real-time crack recognition capabilities.