Research On Character Recognition With BP Neural Network And Convolution Neural Network

Character recognition has been applied widely for many aspects in our daily life,many systems and classification algorithms that can be able to achieve the recognition ofthe text have been proposed. But the recognition rate of handwritten characters and thelarge character set of text is still not high enough. In recent years, the artificial neuralnetworks become a powerful tool to simulate the artificial intelligence, and it shows astrong capability in the field of pattern recognition. So, how to conduct research in thefield of character recognition with the neural network becomes very important.BP neural network is a typical method of pattern recognition for mapping therelationship of given input and output data pairs, and the relationship can be obtainedthrough training/learning processes, nay more, BP neural network is still able to makevery good predictions even the input data have not studied. These properties make it apowerful method for pattern recognition and prediction in an extremely broad field.Comparing to BP neural network, the structure of convolution neural network is morecomplex and has more network layers. However, because of its properties of sparseconnectivity, shared weights and other characteristics, it does not increase the difficultyof training and learning too much, on the contrary, its ability to identify can be improvedgreatly. In addition, the parallel connection of convolution neural network makes itespecially suitable for image recognition. Using both of the two neural networks forcharacter recognition has a great practical and research value.Usually we can use the image as a medium of text to achieve the characterrecognition. Under this premise, the image processing will affect the success ofrecognition directly. It should be noted that there are many random noises during theimage processing in the text which should be removed or smoothed, apart from that, theposition of the text must be found before recognition, and the size of picture and picturecolor channel should be normalized. Only after these pretreatments the picture then canbe composed into a training set for the neural network. Because BP neural networkcannot use the image as an input directly, it also needs us to extract characteristicparameters of the text from the picture. In this article we use the word complexity index and moment invariants as the character’s feature, and the characteristic parameterextraction method has been improved.This paper describes the derivation and computing process of the network structureof BP neural network and convolution neural network. On this basis, we use both of thesetwo networks to identify lower case letters and Chinese characters which includedprinted and handwritten text. Through analysis the results of the two networks forrecognition, we point out these two networks’ advantages and disadvantages. At last, wecompared the recognition rate of convolution neural network and the function of AdobeAcrobat’s OCR, and it shows that the convolution neural network has more advantagescompared with BP neural network and the function of Adobe Acrobat’s OCR on thecharacter recognition.