Research on Deep Learning-Based Concrete Bridge Crack Detection

Cracks are prevalent structural defects in concrete bridges. Presently, cracks detection heavily relies on manual inspection, which is inefficient and inaccurate. The industry has been focusing on research into automated and intelligent methods of crack detection. In this paper, we propose an enhanced deep learning technique based on YOLOv5 that addresses the issues of high false detection rates and poor real-time performance in detecting concrete bridge cracks. To begin with, we replace the C3 module in YOLOv5 with the C2f module from YOLOv8, which expands the network's perception range and the capability to represent features. Secondly, we introduce the SE attention mechanism, which adaptively adjusts the channel weights of feature maps, enhancing the focus on important features. Finally, we adopt a DyHead object detection head based on the attention mechanism, which performs excellently in capturing the details and boundary information on bridge cracks. Experiments have demonstrated that this method can rapidly and accurately detect cracks with a low false detection rate. The average precision (mAP, Mean Average Precision) reaches 94.2%, showing a significant improvement of 9.7% compared to the original network. These results showcase the outstanding performance of the proposed method in concrete bridge crack detection tasks.