End-to-End Bank Card Number Detection And Recognition Based On Camera

With the smart camera,mobile phones and other intelligent terminals continuing to popular,automatic text recognition technology based on the camera is being widely used.With card number automatic identification technology,we can quickly,efficiently,automatically and accurately obtain the information of bank card images.Similar to the problem of text detection and recognition in natural scenes,card number detection and recognition are faced with similar problems: influence of light,influence of camera perspective,image quality,color of card number and so on.In addition,the complex background of the card number designed by human also increases the difficulty of detection and recognition.This paper systematically studies the detection and recognition of the end-to-end card number based on camera.The main work and innovation are as follows:In the card number detection,considering the fact that card number is arranged,to narrow the search range of detecting numbers and eliminate the irrelevant background interference,we cleverly use LSD linear detection operator to detect card and fully exploit the significant differences between card number characters and background disturbances at the edge of the structure.Then,the method of card number row detection based on edge boxes and spectral clustering is proposed.Secondly,for the nuances of the card number and the non-card characters in the gradient and texture,the FHOG feature with stronger gradient expression and the improved IRLBP with better description of the fine texture are proposed.Combining with AdaBoost method,we propose a multi-level cascade detector card number digital character detection method,and confirm the effectiveness of the card number detection method by experiments.In the aspect of card number recognition,a GLAC descriptor describing the interspecies fixed difference is proposed for the influence of factors such as complex numbers of complex numbers and the interclass class differences are relatively special and complicated.And the SIFT-BoF descriptor based on the pre-built visual dictionaries is proposed.Then we combine the two descriptors and use the multi-class SVM to construct the character recognizer.Considering the inherent advantages of convolution neural network(CNN)in classification,we propose to decompose the convolution layer toincrease the number of network layers,and finally improve the nonlinear expression ability of the network;introduce the batch normalization technology to avoid over fitting to improve the optimization;and fine-tune the network training process to speed up the convergence rate.Then,the abstract semantic features extracted by CNN network are analyzed by factor analysis,and the style factor and content factor are separated.The character recognition method based on improved CNN is constructed,and the validity of the method is proved by comparison experiment.Finally,we combine the character recognizers to construct an end-to-end card number recognition system,and the adaptability of the detection and recognition methods is proved by comparing experiments.