Wheel Water Meter Reading Recognition Based On Improved Convolutional Neural Network

With the increase of data volume and the rapid improvement of computer computing power,the drawbacks of the traditional manual meter reading method are increasingly obvious.Most of the existing water meter reading recognition technologies use shallow learning models,but their generalization ability for complex data sets will be limited,and the selection of sample features needs to be based on prior knowledge,and the operation is relatively difficult.In view of the above problems,in the context of deep learning,guided by convolutional neural network theory and support vector machine,the root mean square error,misclassification rate,precision rate,recall rate and time required for network training are used as evaluation indexes,the recognition results of wheel water meter images are compared and analyzed.The simulation results show that the algorithm not only improves the recognition accuracy but also reduces the time required for network training,which proves that the paper algorithm is effective for character recognition of character on wheel water meter images.The main contributions of the paper include the following three aspects.Firstly,the wheel water meter images is preprocessed based on digital image processing technology.For the problem that the character wheel water meter images contains a lot of noise due to damage,environment and light,which affects the recognition accuracy.Based on the digital image space processing theory,the mean value method is used to perform gradation processing on the character wheel water meter images.The horizontal projection method and the vertical projection method are used to extract the character region of the character wheel water meter images,and the single character segmentation of the character region of interest is performed by the equalization method.The simulation results show that the paper algorithm is effective for a series of pre-processing operations on the wheel water meter images,which lays a foundation for the subsequent recognition of character readings of wheel water meter images.Secondly,based on the data set theory of handwritten character recognition at Stanford University,a data model is established.For the problem of insufficient sample data size,based on the handwritten character recognition data set theory created by Stanford University,the character wheel water meter images is divided into single characters,the data set containing 2000 wheel water meter character images,which is expanded to a scale containing 10000 character data sets,7500 of them are used as training sample sets,and 2500 of them are used as test sample sets.The simulation results show that the paper has high test accuracy for the expansion and division of the image dataset of the wheel water meter,which proves that it is effective.Thirdly,the CNN structure was designed and optimized to improve the recognition rate of character readings on the wheel water meter images.Aiming at the problem of insufficient recognition performance and slow convergence speed of existing algorithms,based on the CNN theory,the CNN is used to automatically extract and learn the characteristics of the input dataset,the SVM is used to character recognition of the wheel water meter images,and the dropout function is used to accelerate the network convergence.The simulation experiment shows that,compared with the traditional shallow learning algorithm and the classical CNN model,the accuracy of the reading recognition algorithm of the paper is increased by 0.89% and 0.07% respectively,and the network training time is also relatively reduced,which proves that the reading recognition algorithm of the paper is effective.