Paper Currency Number Recognition Based On OCR And Hardware Simulation

Paper currency number recognition technology is the essential of the sorter.A lot ofautomatic teller machines also have this function.But now the recognition technologyrequirements of paper money is relatively new,no stains or otherwise the error on coin.The main reason is that the machine with built-in recognition algorithms can not processthe pollution. The content of this paper is trying to put forward some solutions to solvethese problems. Through analysising the error samples of a sorter disigning company ingShanghai, there are several recognition problems: large stains plaque, the lineardisturbance number region, low resolution characters, recognition algorithm. The causesof these problems and solutions are discussed in this paper.First of all,the plaque or particulate is one of the pollution.We usually use the way ofdiscovering to smooth picture,but the effect is not very good, because of the some bigarea pollution can not being smoothed away. To solve this problem, this paper uses theconnected domain area deletion based eight-approach.By comparing the results withmedian filtering, the paper shows its good performance on such issues.Secondly, there is the sample containing linear noise characteristic,which is theimportant reason to affect the character recognition. On line processing,compared with thetwo value of traditional chart algorithm, this paper uses linear algorithm based on graylevel domain, has obvious advantages, it also remove the line completely left thecharacter information.Due to low resolution of the image sensor that the machine uses now, low resolutionfor character recognition is one of the difficulties in this article. For this problem, this articlestarts from the character normalization and improved recognition algorithm to resolve thetwo aspects. Low resolution means information content of one pixel contains more,particularly, important character details. In this paper,double edge detection algorithmbased on unified character strokes is used.Unifying The maximum stroke width,the binaryimage can contain more character details reserved. Accordingly,the overall recognitionrate has been7%increase.Recognition algorithm contians a multi-layer BP neural network architecture whichbased on a decision tree,significantly improving the character recognition rate.By testingto identify1000Paper currency numbers which gain from the sorter machine, therecognition rate can run up to98.3%.this algorithm compared to the original93%machinerate, improve.Finally, in order to test and verify the time effectiveness of the algorithm, the algorithmis transplanted to the DSK6416platform to process hardware simulation operation.Verifing the actual results, the average recognition time of the algorithm is of15.6milliseconds per piece of image,achieving the goal of20milliseconds recognition speed.