基于多尺度特征融合与透视校正的水尺视觉识别模型
A vision-based model for water gauge recognition via multi-scale feature fusion and perspective distortion correction
摘要
提出一种融合多尺度特征与透视校正的水尺图像精确识别模型,旨在实现复杂环境下水位的高精度自动识别.该模型基于改进 YOLOv5 算法的水尺图像水位自动识别模型,对深度学习算法中的骨干网络、特征融合与预测模块进行了优化,并引入基于PANet 的双向特征聚合机制.针对图像中由相机视角引起的透视畸变问题,引入基于样条插值的像素-高程映射函数,并构建了适应不同光照强度与拍摄角度的数据增强策略,提出了 ROI 区域提取与分水岭边缘检测相结合的方法,实现了水位的精准识别.在静态图像与视频流测试中,模型识别误差均在 3 cm 以内,实时视频流图像处理速度最高可达每帧 0.05~0.10 s,具有较高的实时性与稳定性,可为水库大坝安全运行管理以及应急条件下的水位监测提供保障.
This study proposes a high-precision water-gauge image recognition model that integrates multi-scale feature fusion and perspective-distortion correction,aiming to achieve accurate water-level extraction under complex environmental conditions.The framework is developed based on an improved YOLOv5 architecture,in which the backbone network,feature-fusion modules,and prediction head are op-timized,and a PANet-based bidirectional feature aggregation mechanism is introduced.To address geometric distortions caused by oblique camera viewpoints,a spline-interpolation-based pixel-to-elevation mapping function is constructed.Moreover,a comprehensive data-augmenta-tion strategy is designed to accommodate variations in illumination intensity and camera angles.A hybrid method combining region-of-interest(ROI)extraction with watershed edge detection is further proposed to achieve precise quantitative estimation of the water level.Experimental evaluations using both static images and real-time video streams demonstrate that the model maintains a water-level recognition error within 3 cm,while achieving real-time processing speeds of 0.05-0.10 s per frame.The results confirm the model's strong robustness and stability,in-dicating its potential to support reservoir-dam safety management,real-time water-level monitoring under emergency conditions.
年,卷(期):
2026, 48(10)
分类号:
TV22
机标分类号:
TP391.41
TN911.73
TP751
在线出版日期:
2026年10月19日
基金项目:
SS110145B0032025:中国水科院基本科研专项52579122:国家自然科学基金2025MS070:中央高校基本科研业务费专项,