Design And Implementation Of Code-character Recognition System Of Gas Meter Based On Convolutional Neural Network

Gas metering information automation has become an indispensable part of information automation management.However,there are still a large number of wheel-type gas meters that cannot collect information automatically.In order to enable them to be used in an information-automated way,it is more cost-saving and time-saving to achieve the project of installing an industrial camera that captures code-character images on the wheel-type gas meter,and using code-character image recognized instead of the scheme that replaces directly all wheel-type gas meters with electronic gas meters read automatically.In this paper,a gas meter code-character recognition system based on convolutional neural network is designed and implemented.The main work in this dissertation is to collect and construct the gas meter code-character data set.At the same time,utilizing the gas meter code-character data set and training two machine learning models to recognize the gas meter complete code-character and incomplete code-character respectively are also the goals of this paper.The specific contents are as follows:(1)Based on the design and construction of the hardware and software environment of the gas meter code-character recognition system based on convolutional neural network,the hardware system consisting of front-end image acquisition module,intermediate communication module based on RS485 and computer is introduced.(2)Collected the gas meter code-character data and the data is denoised and normalized.The gas meter complete code-character data set and the gas meter incomplete code-character data set are compiled and constructed.According to the characteristics of gas meter codecharacter data,a convolutional neural network model will be constructed which includes a convolutional layer,two max-pooling layers and an InceptionV2 structure.The CNN model is trained on the gas meter complete code-character training set.The network model eventually reaches the goal of accurately recognizing the complete code-character of the gas meter.(3)In order to overcome the problem of unbalanced data in the gas meter code-character data set,namely insufficient data of incomplete code-character.Three process of training and fine tuning finally completes an incomplete code-character recognition model training based on convolutional neural network,using mask code-character assisted training set generated,gas meter complete code-character training set and incomplete code-character training set combining with the loss function Grid Loss of the dominant parameter optimization direction proposed in this paper,finally completed the training of the incomplete code-character recognition model based on the convolutional neural network and achieved the goal of accurately identifying gas meter incomplete code-character.In this paper,two kinds of gas meter code-character recognition models were tested for gas meter code-character to get a final recognition.The accuracy of the recognition of the complete code-character of the gas meter reached 99.97%,and the recognition accuracy of the incomplete code-character reached 99.19%.What is more,the results show that the convolutional neural network containing InceptionV2 is a stable,efficient,and highlyfeasible pattern recognition algorithm for solving the problem of recognition gas meter codecharacter.