Abstract: Water level monitoring in rivers，reservoirs and other open waters，usually requires special components for water level measurement at the monitoring site，such as water level measuring sticks，pressure sensors and other equipment. It is not possible to monitor the water level by only using a camera. To solve the above problems，a water line detection algorithm based on Convolutional Neural Network （CNN）was proposed. The input of this algorithm was a static image containing water line. The output were the predicted coordinates of the intersection of the water line and the boundary on the left side of the detected image，and the predicted angle between the water line and the horizontal direction. Finally，the predicted water line was drawn according to the intersection coordinates and included angle of the CNN network output. The test results show that the proposed algorithm has a strong adaptability to the detection environment. Even if it rains at night and is only illuminated by infrared light source，the water line can be predicted effectively by this algorithm. The algorithm can realize all-weather non-contact continuous monitoring of water line in open water areas such as rivers and lakes.

Key words: deep learning, Convolutional Neural Network (CNN), Residual Network (ResNet), water line detection; water line identification

摘要： 目前对河道、水库等开放水域的水位进行监控，通常需要在监控地点部署用于测量液位的特殊部件，如水位标尺、压力传感器等设备，无法仅使用摄像头完成水位的监控。为解决以上问题，提出一种基于卷积神经网络（CNN）的水位检测算法，该算法输入为一张包含水位线的静态图像，输出为水位线与检测图像左侧边界交点的预测坐标及水位线与水平方向的夹角预测值，最后根据网络输出的交点坐标及夹角绘制预测水位线。测试结果表明，该方法对检测环境的适应能力极强，即便在夜间下雨，且只由红外光源照明的情况下也能对水位线进行有效预测。使用该方法可以对河道湖泊等开放水域实现全天候非接触式水位连续监控。

关键词: 深度学习, 卷积神经网络, 残差网络, 水位线检测, 水位线识别