Toward High-Precision Crack Detection in Concrete Bridges Using Deep Learning

Concrete cracks that cause durability problems and bearing capacity failure are critical for evaluating the in-service performance of concrete bridges. It is still challenging for deep-learning methods to accurately detect and quantify concrete cracks. To improve detection precision, a computer-vision (CV)-based crack detection framework is presented in this study. The proposed framework introduces redesigned deep convolutional generative adversarial networks (DCGANs) and extends the dataset of concrete crack images by generating synthetic examples from collected crack images. Based on the enlarged dataset and the you-only-look-once version 5s (YOLOv5s) algorithm, model training and crack detection are conducted through backbone, bottleneck, and prediction part. Subsequently, two algorithms, including the Ostu method and the medial axis algorithm, are combined to calculate the length and width of concrete cracks. Experimental tests compared the YOLOv5s model with YOLO series algorithms, region-based fast convolutional neural network (faster R-CNN), and single shot multibox detector (SSD). The dataset augmentation by redesigned DCGANs increased mean average precision (mAP) by 3.7%, and YOLOv5s outperformed in detection speed with 43.5 frames/s. The crack size measurement of concrete members in the laboratory demonstrates that the calculation of crack size is achieved at pixel level. The proposed concrete crack detection framework can meet precision requirements and provide a promising measure for patrol inspection.