Research On Deep Learning Based Image Recognition Method For Water Level

Flow head directly determines the power generation capacity of the entire hydropower station,at the same time,the water level in the reservoir area,upstream water level,gross head and other types of hydrological data is considered to maintain the normal work of the entire hydropower station engine group,dam safety monitoring the main focus and monitoring objects,only the water level information in a digital form in a timely manner to the monitoring layer,management,decision-making layer,the upper level of the application Software and decision analysis system can make more scientific judgment.The traditional water level monitoring method through visual inspection of the water ruler or professional water level detection sensor to achieve,and hydropower stations are often located in remote areas,harsh environment,in extreme weather conditions,will threaten the personal safety of the observer,water level monitoring point cable laying most of the self-connection exposed in the field,accelerating the breakage of aging is also easy to be destroyed by animals,engineering construction and post maintenance is relatively difficult.With the popularity of video surveillance technology,the use of digital image processing technology for water level identification slowly become a research hotspot,this non-contact water level detection system has the advantages of high stability,reliability and real-time.Therefore,the use of digital image processing technology for water level monitoring is a very valuable research topic,and its intelligent management of hydropower plants is of great significance.In this thesis,a set of reservoir water level detection method based on digital image processing technology is designed to automatically identify water level values by using image recognition technology to process the captured water scale images and recognize water level characters through deep learning algorithms.The main research contents and innovation points are as follows.(1)In order to better adapt to the positioning of water ruler in foggy weather and low light environment,this thesis explores various image enhancement methods,and finally selects the dark channel a priori algorithm to recover and enhance water ruler details from degraded images.(2)In the pre-processing stage,the captured water ruler images are grayed out,binarized,and edge detected to outline the outline of the water ruler,and then the water ruler is stripped out by combining morphological processing to remove the background according to the characteristics of the water ruler,followed by tilt correction of the water ruler.(3)In the process of water ruler character region segmentation,the stripped water ruler image is projected vertically,segmented into two pieces on the left and right,and the left water ruler image is inverted and cut out horizontally.(4)In the process of water scale character recognition,the cut out character column segmentation,normalization,feature extraction and other operations are fed into BP neural network and convolutional neural network respectively for character recognition,and the water level value is calculated for different water level changes,and the convolutional neural network has better effect after comparison and analysis.The water level acquisition technology based on the image of water ruler is mainly realized by software in addition to the water ruler and camera device,which has the characteristics of high measurement accuracy,simple equipment,easy maintenance and low construction cost,and can be widely used in water level measurement as an effective supplement to the sensor type water level meter.