Water Level Warning Research With Embedded Terminal Based On Semantic Segmentation

Reservoir levels and its early warning are key indicators for the safe operation and maintenance of the reservoir and the prevention of overtopping in the water sector.As human cost increases,it is difficult to maintain the traditional water level sensors.The water level warning system based on cloud computing has the problems of low capacity storage of images and videos and high bandwidth costs of data transmission.Therefore,the use of computer vision in the embedded edge computing terminal to carry out semantic segmentation for water level detection and early warning is pivotal in reducing management costs of the reservoir and improving popularity of intelligent detection for water projects.Semantic segmentation is one of the key research hotspots in the field of computer vision,among which semantic segmentation is more efficient and reliable than traditional segmentation algorithms.However,many complex image segmentation networks are hard to be deployed on the embedded platform,making them difficult to put into practical use.Under the background of reservoir ecology,this paper proposes an embedded terminal water warning system based on semantic segmentation.This system improves the robustness of the model by collecting diversified data with multiple elements in different time periods and weather conditions and by data augmentation.Since the reservoir water area accounts for a large proportion of the entire reservoir image,the reservoir water area is easier to segment and recognize than other segmentation objects.Considering the practicability and versatility of network deployment on the embedded terminal,a simple and effective FCN algorithm is selected to realize the pixel level segmentation of reservoir water area.The FCN model was compared and analyzed on two different common optimizers respectively to select an optimizer suitable for deep learning training.In view of the difference in the proportion of the total number of pixels in each category of reservoir’s image data,a weighted cross-entropy loss function is proposed to balance the data set and train a more effective model based on the FCN algorithm.To solve the problems of poor correlation of spatial information between the context in the output pixels,the FCN model is improved via fusing the multi-scale pooling feature.The FCN model,the improved FCN model of the weighted cross-entropy loss function,and the improved FCN model of the multi-scale pooling and the weighted cross-entropy loss function are compared for experimental analysis.The performance test of the model was carried out on one of the embedded terminal products Jetson TX2 produced by Nvidia.The model with the highest accuracy were finally selected by comparing the test results.An improved FCN model which combines weighted cross entropy loss function and multi-scale pooling features is deployed on the embedded terminal to realize early warning and recognition of reservoir water level.The embedded terminal water level warning based on semantic segmentation proposed in this paper can effectively prevent floods.