Study On River Flow Measurement Based On CNN And Image Processing

The shortage of water resources has always been an important factor restricting the economic development in China,and the monitoring of river discharge is the key to the quantitative management of water resources.Traditional manual flow measurement has been difficult to meet the requirements of intelligent water resources management due to its heavy workload and inconvenient data transmission and processing.However,in recent years,Doppler flow meter and other instruments used more frequently are not the most convenient and intelligent flow measurement method due to its high cost,large environmental impact and numerous preset parameters.Therefore,it is of great significance to develop and design a fully automated and universal river flow measurement scheme for intelligent flow measurement and quantitative management of water resources.In order to build an intelligent model of river flow measurement and study the flow velocity and water level recognition algorithm,how to ensure real-time measurement and remote control is the key to the model.In recent years,the rapid development of deep learning and image recognition technology has provided ideas for intelligent flow measurement.Because of the corresponding relationship between the river surface velocity and the characteristics of ripples and waves in a stable state,and the water gauge used for water level measurement has a strong presence.This paper proposes to use convolutional neural network to identify the characteristics of the river flow image to determine the surface velocity,use the image recognition morphological transformation to identify the scale of the water gauge to determine the water level,and combine the recognition results of the two models to further calculate the real-time flow of the river.The main research process and results are as follows:(1)Based on the convolutional neural network,the river surface velocity recognition model is constructed.Use adaptive histogram equalization to enhance the features of the water flow surface image;use the data enhancement algorithm to expand the water flow surface image data set,and carry out the structural design of the current three convolutional neural network models with high recognition,and adjust the parameters to Obtain the most suitable model for water surface image recognition.From the five steps of input,convolution,pooling,full connection and output,it explains how to classify and recognize the input water flow image.The design of the convolution kernel and the arrangement of the convolution and pooling layers are the most important.The cohesive transformation between the size of the feature map and the number of channels is the basic principle followed by the research.Using Python and TensorFlow modules to adjust the parameters of the convolutional neural model,the optimal model was determined and the back propagation training of the model and the identification of the surface velocity of the water flow were realized.(2)Based on the image recognition morphology module,the water level recognition model of the water gauge is constructed.Starting from the preprocessing of the original image,the water gauge image is extracted using edge detection and horizontal pixels,and then geometric feature extraction is used to eliminate the influence of numbers and other redundant images to obtain the scale number,and finally calculate the water level value and calibrate the algorithm.The shooting angle and distance of the water gauge image have an impact on the recognition.This study compared images from several shooting angles and adjusted the parameters respectively to achieve better recognition results.(3)Case study and analysis.The flow measurement experiment of the water tank was carried out,and the surface velocity recognition model and the water level recognition model of the water gauge were used to identify the water surface images and water gauge images at different velocities and water levels.Experiments have obtained water flow surface videos at different flow rates,intercepted a large number of water flow images per frame,and constructed a water flow data set after feature enhancement and data enhancement.The trained network was used to identify them with an accuracy rate of more than 90%;for different water levels The recognition result of the water gauge image below shows that the water gauge scale recognition using morphological algorithm is better than the Hough transform,and it has better recognition accuracy for water gauges with different shooting angles,and the maximum error of multiple recognition is 19.05%.Experiments have proved that the surface velocity recognition model established in this study and the water gauge water level recognition model are both effective,and the combination of the two can calculate the real-time flow of the river.In summary,the channel flow measurement scheme based on CNN and image recognition is a brand-new flow measurement method.For different application environments,only the model structure and a few basic parameters can be set to adjust the model itself through training to ensure the algorithm’s performance.Universality and intelligence.At the same time,the river monitoring points are equipped with cameras,which can realize video monitoring and flow measurement at the same time.The management is convenient and the cost is low.With the expansion of the data set,the accuracy and accuracy of flow measurement will be further improved.This research is an important development direction for smart flow measurement.