Intelligent Surface Cracks Detection in Bridges Using Deep Neural Network
Xiaobo Zhang, Zhipeng Luo, Jinghao Ji, Yimin Sun, Haihao Tang, Yongle Li
Abstract
Cracking is one of the most common bridge diseases. If bridge cracks are not repaired: in time, they may cause gradual changes to the concrete structure, which can seriously affect its strength. A network called YOLOv5-TS is what we suggest to detect intelligently bridge surface cracks in images. To improve the network performance, we integrate SPPCSPC into the original YOLOv5 network to ensure adaptive image output and obtain receptive fields of various sizes. Meanwhile, transposed convolution is incorporated to improve the capacity of the network for learning weights on its own and reduce characteristic information loss. In response to the diverse morphology of bridge cracks, cracks are identified according to their mechanical causes crack inclination, and divided into four categories: horizontal cracks (0[Formula: see text]–20[Formula: see text]), low-angle cracks (20[Formula: see text]–45[Formula: see text]), vertical cracks (70[Formula: see text]–90[Formula: see text]) and high-angle cracks (45[Formula: see text]–70[Formula: see text]). Experiments on the ZJU SYG crack data set confirm that the proposed YOLOv5-TS has a better crack intelligent identification effect on bridge surface images than other compared baselines. The best performance of YOLOv5-TS is found in [email protected] (0.752), [email protected]:0.95 (0.518), and recall (0.794), thus demonstrating the model’s practical value.