Design And Implementation Of Handwritten Digit Recognition System For Receipt Based On CNN

In China,while traditional manufacturing enterprises are developing their information technology,the level of informatization and digitization is always uneven.For example,the phenomenon of large number of handwritten receipts in domestic discrete manufacturing enterprises is still very common.In the process of enterprise transformation,the receipts are used in the enterprise,with a large number and a wide range,including important contents.According to requirements for unification,standardization and integration of receipt data when the enterprises are transforming,this thesis studies the handwritten digit recognition technology on receipts,makes the digitization of enterprise information and the standardization of management system,so as to improve the management level of the factory and achieve the purpose of reducing costs and increasing efficiency.According to the characteristics of enterprise-specific manual receipts,this thesis mainly includes:First,the image pre-processing of handwritten receipts and digital area extraction techniques are studied.The basic idea and process are:manual receipts are stored in the office system through the OA client scanning,and Hough transform is used to correct the receipt angle.Based on the receipt characteristics,reference points are defined and the regions of interest are located.The receipt images are denoising in typical regions of interest,and interference lines are removed.Then handwritten numbers are extracted separately.Secondly,the digital recognition method based on convolutional neural network model is studied for handwritten digital regions.The MNIST data set is used in CNN model for training,and the hand-drawn digital part of the enterprise internal ticket image is used to carry out small sample training and recognition through the transfer learning technology.Finally,combined with OA system and technology of receipt processing,the handwritten digit recognition system is designed and implemented.The identification system includes a stand-alone version with multi-algorithm evaluation and a web version of the B/S architecture.The network version system adopts the B/S network service architecture according to the existing optimal processing flow of the stand-alone version,and implements interactive application based on the Web server and client technology and database technology.This paper verifies the effectiveness and feasibility of the pre-processing methods and workflow through a set of receipt processing experiments.BP neural network,SVM,CNN and CNN with transfer learning are compared.The experiment shows that the recognition rate of CNN with transfer learning reaches the best recognition rate.The software which based on B/S architecture is published has been tested in the enterprise,and the digital inquiry and statistics of receipt information have been realized,and the work efficiency has been improved.