Research And Realization Of Character Recognition Algorithm Based On Convolution Neural Network

With the rapid development of machine vision technology,various kinds of industrial applications began to join the machine vision technology,make the test in the industry to generate identification speed rapid ascension,such as the character recognition is a widely used machine vision technology field.But machine vision are widespread some common shortcomings,such as machine vision algorithm of character recognition,must be in a stable environment,light type single sample cases,can achieve good recognition effect.This disadvantage comes from the fact that the machine vision algorithm works essentially by using manual setting characteristics for similarity comparison,which makes it difficult to deal with the changeable situation.With the deep learning algorithm in recent years,it is a feasible method to improve machine vision based on deep learning technology.In the case of the convolutional neural network in deep learning,it has the ability to independently learn image characteristics and further abstract,which makes the image recognition in the complex environment feasible.In this paper,the traditional machine vision algorithm is difficult to deal with the problem of multi-character recognition in complex environments,and the neural network of three recognition characters is designed based on the convolution neural network.Specific work is: first of all,this paper introduces the convolution of the neural network basic ideas,basic network architecture and training process,and then briefly expounds the selection of data sets collected and preprocessing,and then began to design a neural network.Started design is shallow neural network,it is through the LeNet5 modified,which improved the layer such as convolution convolution kernels,convolution layer number,pooling layer algorithm,etc.,reducing convolution kernels,deepened the convolution number a little,this paper proposes a maximum value of average pooling method,and the design of network has carried on the experiment,the result confirmed the feasibility of the design of shallow network.And then design the second kind of network is a double row of neural network,based on deeper network tend to have better results,after learning convolution piece of ideas from VGGNet network,the nerve network layer increased,also had better effect through the experiment,after in order to further enhance the accurate of the recognition,this paper proposes a weighted value of double row neural networks,experiments confirmed that the network can further deepen the accuracy.Residual network,finally we design a depth accuracy cannot be given ordinary convolution neural network after deepen network was further promoted,based on the residual ResNet network design a suitable for character recognition,the depth of the residual network solves the problem of network accuracy have not been able to go up before.At the end of the article,the experimental results of the three networks are compared and analyzed,and the original design goal is realized.