An Advanced Real-Time Semantic Segmentation Algorithm for Water Level Detection

For achieving automated water level detection based on image recognition, emphasizing precision and real-time capability, we introduce a water level detection algorithm founded on real-time semantic segmentation. Firstly, we introduce a shared feature extraction layer to effectively reduce channel and parameter redundancy in the BiSeNetV2 network. Subsequently, the feature extraction backbone is redesigned using partial convolution and FasterNet block, to achieve a balance between precision and inference speed. Additionally, we introduce an auxiliary edge loss function to improve the precision of water gauge segmentation edges. We achieved an mIoU of 97.3% at 132.9 FPS on the water gauge dataset, with a water level reading error of less than 1 cm under practical measurement conditions. Compared to other real-time semantic segmentation algorithms, our approach achieves an excellent balance between precision and real-time performance. Our method has been implemented at hydrological stations in two other regions and continues to maintain high accuracy in practical applications, demonstrating its robustness and generalization.