Water Level Measure By Means Of Machine Vision In Complex Natural Scenes Of Rivers

Hydrological monitoring is an effective way to obtain the condition of rivers and lakes in time.As a major aspect of hydrological monitoring,water level observation refers to the on-site measurement of the water level of rivers and lakes.It can not only directly reflect the condition of rivers,but also help to issue warnings in time.In recent years,non-contact measurement methods based on video image have been applied to water level observation.The task of water level observation based on video image is mainly to detect the water level of natural rivers.Based on the relevant data of hydrological monitoring at a hydrological station in Wuyuan,Jiangxi,the factors that influence the water level detection of natural rivers in the wild are researched in this thesis.The water level detection task includes image preprocessing,water gauge detection and water level value recognition.Finally,the non-contact measurement of water level based on video image is realized.The detailed work of this thesis is as follows:(1)The complex lighting environment in the wild and weather factors are the difficult problems faced by the water level observation in small or middle rivers.Firstly,the image preprocessing denoises the collected hydrological image.Secondly,in order to keep the water level detection work free from the influence of the lighting environment,the common problems such as insufficient sunlight,underexposure,and sun shadows are dealt with.In addition,for the foggy weather that often occurs in small and medium-sized river basins(or mountainous areas),the dehazing algorithm is used for interference removal.(2)Water gauge detection under hydrological monitoring scenarios based on deep learning.The interference of the unknown complex environment background to the feature extraction of hydrological image is the key issue that affects the research of the subject.In this thesis,use saliency detection to suppress complex background interference,combined with grabcut algorithm and water gauge image features based on the HSV color space,the rough detection result of the image interest area of the water gauge can be obtained.The water gauge detection method based on differentiable binarization in hydrological monitoring scenarios is proposed to obtain a better detection effect.The method introduces the idea of text detection in natural scenes,uses the text image features of numbers and characters on the water gauge,and based on the existing pre-training network of public text detection,realizes the detection of the water gauge region through transfer learning.(3)Research on water level value recognition under hydrological monitoring scenarios.Currently,most of the commercially available non-contact water level measurement devices require preset water gauge positions or fixed device positions to meet the needs of algorithm calibration.However,the actual deployment environment and monitoring requirements are complex and changeable.Therefore,continuous monitoring of water level from a non-fixed perspective is an urgent need and difficult problem for current hydrological monitoring tasks.Aiming at this problem,this thesis divides the water level value recognition into two types: the full region recognition and the non-full region recognition after detailed segmentation of the rough extraction water gauge results.Aiming at the full region water gauge,the energy function detection method is used to project and map the water gauge image to obtain the water level value of the water gauge.For non-full region water gauge,the water gauge character database is trained based on tesseract to make the recognition ability better,and finally the water level value of the water gauge is obtained.In this thesis,the applicability of the overall algorithm is enhanced by image preprocessing.In the task of water gauge detection,the complex background interference is suppressed by saliency detection,and the water gauge detection based on differentiable binarization is realized in the absence of public annotated data sets,and the detection accuracy can reach85%.For the water level value recognition,the problem that the position change of the water gauge in the video image affects the water level value recognition is solved.The recognition of the water level value at any position is realized,and the recognition accuracy of the water level value can reach 1cm,which meets the requirements of the water level observation standards.