To solve the problems of rough water-level edge and insufficient detection accuracy in existing algorithms, this paper proposes a high-accuracy water gauge level detection network, called DetSegNet. Unlike traditional methods that use a single network for the water gauge recognition and the water segmentation, DetSegNet innovates in the network structure, loss function, etc., by improving the YOLOv4 and DeepLabv3+ algorithms. It effectively integrates modules for water gauge scale recognition, water area segmentation, waterline detection, and water level calculation, achieving efficient localization and precise segmentation of the interface area between the water gauge and the water body. Experimental results show that DetSegNet surpasses existing algorithms in terms of detection accuracy and speed on water gauge image datasets. Field tests demonstrate that water level detection accuracy of this method is less than 1 cm, meeting the accuracy requirements of hydrological monitoring.