Research On Segmentation And Recognition Algorithm Of Bank Card Number

With the rapid development of the Internet,Internet Finance grows quickly,and the electronic payment is becoming increasingly ubiquitous.The technology of automatic recognition of bank card number can improve the efficiency greatly in bank card processing associated to Internet financial industry and electronic payment,achieve the effective management of bank cards and related services as well as improve user experience.Therefore,the bank card recognition system has significant application prospect.Currently,the bank card number printed on actual card is composed of 16 or 19 digits.Although the sizes of bank cards are basically similar,their backgrounds are various and the printing methods of card number including flat printing and emboss printing are different.Besides,there are varying degrees of wear and tear on bank cards during daily use.Above all,automatic recognition of bank card number based on image has a certain challenge.In this thesis,in order to realize the automatic recognition of bank card number,the main works are as follows:(1)With the process of converting to grayscale image,size normalization,denoising and edge extraction successively,the area containing card number is precisely positioned and the edge information of characters is obtained.In view of the issue that the number of the bank card printing method including black printing and embossing,a method of judging printing type according to the proportion of black pixels in the card images is proposed,which lays the foundation for the subsequent segmentation and recognition.(2)Considering the background interference in flat printing card,a method of transferring black pixels to white ones is proposed to remove the interference.Considering the background interference in emboss printing card,based on the YUV color space conversion method by the relevant literature,some optimizations about the projection segmentation are implemented to reduce the error of segmentation.Filling in the vertical projection followed by coarse and fine segmentation,which improves the segmentation speed and accuracy.(3)The recognition of bank card number is realized by the method of deep belief network and convolution neural network respectively.The Deep Belief Networks is constructed by 3 Restricted Boltzmann Machines and 1 BP neural network.The Convolutional Neural Network uses LeNet-5 network model which has a compromise between model size and recognition effect.For testing the performance of algorithm,2700 bank card samples with different background and printing method are collected.There are total 51300 character images after segmentation,of which 39900 images are selected to train and 11400 images to test.The accuracy rate of Deep Belief Networks recognition is 94.4%,and the accuracy rate of Convolutional Neural Network recognition is 99.19%.Experimental results shows the effectiveness of our proposed algorithm.