Real-time detection of concrete cracks via enhanced You Only Look Once Network: Algorithm and software

Abstract

Deep learning algorithms have been employed for real-time concrete crack detection. However, many algorithms are not specifically tailored for this purpose. Moreover, their lightweight iterations are generally optimized at the macro-model level, leaving room for further lightweight enhancements at the block level. Therefore, this study developed an enhanced YOLOv3 (You Only Look Once Network v3) model, named YOLO-Crack. The structural optimization of the model takes into consideration the shapes of concrete cracks in the dataset. Meanwhile, two multiple branch-shaped blocks based on dilated convolutions, convolutions and pooling operations were proposed. The two blocks, incorporating depthwise separable convolutions and attention mechanisms, were used to rebuild the model at the block level. These enhancements significantly reduce the size and improve the detection performance of YOLO-Crack. Furthermore, YOLO-Crack was softwareized for real-time detection of concrete cracks. The software was designed to support parallel computing, allowing for real-time detection of concrete cracks even on laptops with limited computing power. It was utilized to detect cracks on concrete roads at a university in Nanjing, China, enabling real-time detection at a frame rate of 30 frames per second with satisfactory accuracy.