Research And Implementation On Recognition Algorithms For The Catering Invoices Of Local Tax

Automatic bill image processing system has been widely used in the fields of taxtreatment, financial, insurance and so on. Our country make paper invoice as transactioncertificate, and changing the paper invoice into data information is one of the basic elementsof computerized accounting. Tax including national tax and local tax in China,VAT(value-added tax) invoice or state tax had been implemented electronically in2003. Thereare many different kinds of local tax invoice, but there is no electronic management in anall-round way till now. In order to improve the work efficiency and realize the automatizationof local tax invoice processing, we make local catering tax invoice as study object, andresearch bill automatic processing technology and system implementation.In this paper, based on the in-depth analysis and summary of the layout characteristicsfor catering local tax invoice and the needs of the computerized accounting, we coulddetermine the main content of the invoice recognition, and then adopt a series of methods toachieve the invoice image preprocessing, image segmentation, feature extraction andcharacter recognition.Using Otsu(the largest class difference method) algorithm to binary the invoices imagefirst, and then we proposed a denoising method, which using connected domain that couldeliminate discrete noise points effectively in the denoising process. Last we used the Houghtransform to correct the catering local tax invoice slope according to the layout features of theinvoice image.After analyzing the prior knowledge such as the characteristic of numbers, the width ofthe numbers, we proposed the vertical projection of the character segmentation algorithm andit could solve the problem of blocking character better.After studying the structural characteristics,we proposed a feature extraction methodwhich combine grid characteristics and features of strokes throughing based on the grid. Thismethod combines the advantages of structural features which could reflect character shapesfully, statistical characteristics with strong robustness and anti-interference ability effectively. Combining the prior knowledge of invoice information, we used genetic algorithm tooptimize the BP neural network recognition method, it could overcomes the disadvantage ofthe BP neural network, such as convergence in local minima and slow training speed.According to the above methods and research, we developed an automatic processingsoftware for invoice, and used208invoices as experimental object, the average recognitionrate is95%, it meets the basic requirements of automatic processing invoices.