Presenter
Dr. Zhongyue Yan, Nanchang Institute of Technology
Co-author(s)
Miss. Yiwei OU-yang, Nanchang Institute of Technology
Mr. Jiang Deng, Nanchang Institute of Technology
Miss. Xiyu Ni, Nanchang Institute of Technology
Sub-theme
5. Establishing Sustainable Water Infrastructures
Topic
5-4. Intelligent business application system for water regulation and management
Body
Image water level recognizing is a hot topic in water resources management. However, the water level recognizing models still need to be improved in terms of environmental adaptability and accuracy now.So this paper presents a set of new water level recognizing method based on Unet semantic segmentation algorithm and ResNet50 classification algorithm. The method mainly consists of four computational processes. Firstly, the Unet semantic segmentation algorithm is used to generate a new image from the original image above the water shoreline . The new image is geometrically corrected based on the coordinates of the four vertices of the contour line that is identify by polygonal approximation of edge line. Secondly, the corrected image is divided in half vertically, and the coordinates of the area where the right half contains the E character are calculated by the projection method. The coordinates are used to split the area containing the numbers on the corresponding left half as a part. Thirdly, the numbers of each parts are identified from top to bottom by the ResNet50 classification algorithm, and logical sequence checked to determine the minimum number. Finally, the water level value is calculated by combining the ratio of the distance from the water shoreline to the minimum number and the average height of single E character which has a fixed actual length. The algorithm was trained using semantic segmentation datasets including 301 pictures of water gauges annotations from different counties and cities, and classification datasets including 2017 pictures contains 0 to 9 numbers where cut from water gauge. The results showed that the accuracy of Unet semantic segmentation algorithm and ResNet50 classification algorithm were 99.84% and 97.22% respectively. The interpretation accuracy was within 2cm.The method greatly improves the environmental adaptability and accuracy of water level recognizing method, which has a potential and wide application value.