Visual measurement method of multi-level water gauge based on deep learning

Vision-based water gauge measurement technology offers advantages over conventional non-contact meters including oblique detection capability, intuitive results, and no temperature drift. In recent years, it has seen increasing application in single-level vertical gauge measurements. However, for short-pile multi-level gauges, commonly used in wide rivers with gentle slopes, several challenges persist: low manual inspection efficiency, complex system calibration, and limited detection accuracy for distant small targets and the water line. This paper proposes a visual measurement method for multi-level gauges that integrates object detection and water line regression. A bank-based online measurement system is constructed using a high-definition fixed-focus camera. For object detection, an enhanced YOLOv5 network is employed, incorporating the CBAM attention mechanism and an additional small object detection head. Furthermore, the SE-HRNetS water line regression model, utilizing multi-scale feature fusion and the SE channel attention mechanism, is adopted to improve water line detection accuracy in low-resolution images. To validate performance, comparative tests were conducted at an experimental site under varying lighting and water flow conditions. Experimental results show that the system achieves an Average Precision of 94.82% on the local dataset. The comprehensive uncertainty of water level measurement is less than 1.72 cm under diurnal and flood scenarios. The research provides an intuitive and efficient solution for visual water level measurement in small to medium-sized rivers with wide, gently sloping cross-sections.