Real-Time Monitoring of Bridge Crack Evolution Based on the Lightweight Semantic Segmentation Model DST-Unet

<title>Abstract</title> <p>In bridge operation and maintenance management, crack width is an important indicator reflecting the degree of structural cracking, and achieving real-time, accurate, and automatic monitoring is of great significance. To realize millimeter-level automatic real-time monitoring of crack width and explore the process of crack propagation, this paper proposes a lightweight semantic segmentation model, a neighborhood-analysis-based crack quantitative analysis algorithm, and develops an intelligent millimeter-level crack monitoring instrument, which can monitor crack width changes in real time, improving the accuracy and efficiency of crack quantitative analysis and assessment. By optimizing the Unet semantic segmentation model structure and introducing depthwise separable convolution and adaptive dropout layers, the model parameters are reduced by 79.35%, the computational load is reduced by 81.42%, and FPS is increased by 31.29%, achieving real-time segmentation. This paper combines large-field monitoring of bridge cracks with fine-scale monitoring, enabling both efficient real-time monitoring and detailed monitoring of key structures and critical parts. A millimeter-level crack propagation dataset was collected and created, recording detailed images of crack propagation and performing quantitative analysis. This provides an efficient and practical solution for crack monitoring and analysis in civil engineering and various other fields.</p>