Water Meter Reading Recognition Based On Deep Convolution Neural Network

There are still many old-fashioned water meters in society that require manual reading,but relying only on manual reading and testing of water meter readings requires a lot of time and energy.To solve this problem,using computer vision technology to automatically identify the water meter readings can effectively solve this problem.However,in reality,water meters are mostly located in various complex environments,such as dark and humid,so the collected water meter images will have many problems:one is the problem of the water meter dial itself,such as serious pollution of the dial;the second is the manual collection problem,such There will be problems such as camera exposure,various shooting angles,and various dial directions during manual collection.Therefore,if the traditional image recognition method is used,the recognition accuracy of water meter reading recognition will not be high enough and the speed will be low.However,with the development of machine hardware,the use of deep learning methods to detect and identify water meter readings has become an effective solution.The thesis aims to study the target detection algorithm based on deep learning to detect and identify the water meter readings.The main work includes:1.The self-made image data set of water meter readings containing various complex situations under actual conditions.It also sorts out the water meter readings collected under real conditions.Possible problems such as: dial defacement,blurry and unfocused images,presence of light spots,and various dial directions,etc.,can be used to detect and identify model training and improve its robustness;In order to facilitate the detection and identification work,a set of labeling rules is designed for the water meter reading image.2.For the water meter reading pictures taken at any arbitrary angle in various complex environments,this paper proposes a robust automatic recognition method,which is based on a deep residual neural network and uses a self-made water meter reading image data set,The reverse gradient random descent method is used to optimize the weight of the R-FCN model.After repeated iterations of network training,the R-FCN model of the character target detection of the water meter reading is finally obtained,and then the model is used to the water meter reading box in the water meter picture Characters are detected and recognized.3.The characters of the water meter readings detected by the model are sorted by the central sorting algorithm and the specific sorting rules,and a complete water meter reading result can be finally obtained.4.Aiming at the current meter reading operation mode,an image recognition system for water meter readings that can be combined with the current meter reading system is effectively proposed to assist the auditors to review the water meter readings that have been read.Comparative experiments show that the R-FCN algorithm used in this paper has a better effect on small target detection problems such as water meter readings than other detection algorithms,and that the method of this paper can achieve 81.79% recognition accuracy in practical applications,and 0.5293 frames per second recognition rate.Therefore,this method has been verified by experiments and shows that its recognition rate is fast and the accuracy is high,which can meet the actual use requirements.