Quantitative detection of typical bridge surface damages based on global attention mechanism and YOLOv7 network

Abstract

Surface damages of reinforced concrete and steel bridges, for example, crack and corrosion, are usually regarded as indicators of internal structural defects, hence can be used to assess the structural health condition. Quantitative segmentation of these surface damages via computer vision is important yet challenging due to the limited accuracy of traditional semantic segmentation methods. To overcome this challenge, this study proposes a modified semantic segmentation method that can distinguish multiple surface damages, based on you only look once version 7 (YOLOv7) and global attention mechanism (GAM), namely, YOLOv7-SEG-GAM. Initially, the extended efficient layer aggregation network in the backbone network of YOLOv7 was substituted with GAM, followed by the integration of a segmentation head utilizing the three-scale feature map, thus establishing a segmentation network. Subsequently, graphical examples depicting five types of reinforced concrete and steel bridge surface damages, that is, concrete cracks, steel corrosion, exposed rebar, spalling, and efflorescence, are gathered and meticulously labeled to create a semantic segmentation dataset tailored for training the network. Afterwards, a comparative study is undertaken to analyze the effectiveness of GAM, squeeze-and-excitation networks, and convolutional block attention module in enhancing the network’s performance. Ultimately, a calibration device was developed utilizing a laser rangefinder and a smartphone to enable quantitative assessment of bridge damages in real size. Based on the identical dataset, the evaluated accuracy of YOLOv7-SEG-GAM was compared with DeepLabV3+, BiSeNet, and improved semantic segmentation networks. The results indicate that the mean pixel accuracy and mean intersection over union values achieved by YOLOv7-SEG-GAM were 0.881 and 0.782, respectively, surpassing those of DeepLabV3+ and BiSeNet. This study successfully enables pixel-level segmentation of bridge damages and offers valuable insights for quantitative segmentation.