VAT Invoice Character Recognition Method Research Based On Improved Adaptive GA-BP

As the reform of the business tax with a value added tax is promoted to everyindustry in the whole nation, value added tax invoice checking will become more andmore popular and important. The input value added tax invoices which come from thesupplier can be the valid vouchers to deduct tax, directly related to the purchasingenterprise benefits. In order to improve the efficiency of the value added tax invoicecertification, the research using the computer to automatically identify the keyinformation of value added tax invoice has important and practical significance.According to the characteristics of value added tax invoices and thetechnological requirement to identify it, the present study puts forward the generaltechnological framework of character recognition for value added tax invoices,illustrates the character recognition process and main technology. By studying therelevant technical papers, it deeply analyzes and induces the development of OCR andrelated techniques.By analyzing several commonly used gray image binarization method, this paperselects the local dynamic threshold Bernsen algorithm which is widely adaptive forimage binarization; it also uses the Hough transform to detect the lines in binaryinvoice image to get the skew angle and correct the image skew; it designs a particlenoise filtering method based on mathematical morphology; it uses histogramprojection method for segmentation of digital character; it extracts the character gridfeature and horizontal and vertical through feature and forms a40dimensional featurevector.In order to meet the requirements of high recognition rate for value added taxinvoices, this paper designs an improved genetic BP neural network digital characterrecognition algorithm. It studies the basic theory of genetic algorithm and BP neuralnetwork, analyzes the advantages and disadvantages of the two algorithms. By usinggenetic algorithm to optimize the BP neural network, it can well combine theexcellent global search ability of genetic algorithm with the good local searchingability of BP neural network, so as to improve the defect that BP neural network islikely to fall into local optimum and to improve the robustness of the algorithm. Thereare some drawbacks of standard genetic algorithm: it may easily fall into a localoptimum because of premature convergence and the search has no direction. These disadvantages are directly related to the standard genetic algorithm of using thecrossover and mutation probability which does not change with populationenvironment. In order to improve these shortcomings, this paper proposes animproved adaptive genetic algorithm. Based on the definition of convergencecoefficient, this paper proposes a new adaptive crossover and mutation probabilityadjustment formula, so that the probability of crossover and mutation will adaptivelychange with the convergence of population, can improve the defect of prematureconvergence of standard genetic algorithm, and improve the global optimizationcapability of genetic algorithm. The research designs a BP neural network, combinesthe improved adaptive genetic algorithm and BP neural network, uses the adaptivegenetic algorithm to search for a set of suboptimal connection weights and thresholdsof BP neural network, and then use the BP network to do the subsequent training tomake the sample global error less than the predetermined error limit value, thiscombined algorithm can improve the defect of BP neural network that it may fall intolocal optimum. In VS2010, combined with OpenCV, using C language codeimplements the algorithm in this paper, experiments show that there are certainadvantages of the improved adaptive genetic algorithm, the improved adaptive geneticBP neural network algorithm is robust and effective.