Research On Fast Recognition Method Of Handwritten Form Digital String Based On Self Learning

The recognition of handwritten form number has become an importantapplication in the field of OCR. A large number of experimental data records, reportform and bill will spend a lot of time if processed artificially, and the process isvery single and boring, also the correct rate can not be guaranteed. This paperintroduces the design process of a automatic recognition software for handwri ttenform number, And applys it to the actual process of experimental data records.The most important part of the recognition of form number is stringrecognition. And the main factors affecting string recognition is the stringsegmentation and single character identification. After analysis the internationalstatus of the recognition of handwritten form number, find that under normalcircumstance, the process of handwritten form number recognition is as follows:firstly, according to the structure of form lines, process the form, get the form part;secondly, locate the cell and extract the digital character image; thirdly, segmentdigital string and classify the results.This paper mainly research on the process of handwritten form numberrecognition, form document image process as follows, preprocessing, locating celland extracting character string, segmentation and character string recognition. Thepreprocessing part includes binaryzation, denoising and tilt correction. Use the formline inclination to slant correction, compare Hough transform and Viterbi algorithmto get the form lines. Using connected components to locate and extract cellcharacter, this method can separate parts which are not connected with each other.When separate character string, use a BP network to guide segmentation, improvingthe correct rate of segmentation. Here this paper uses the confidence of therecognition results to determine the accuracy of the classification results. Use thestring profile to analysis and determine segmentation path. In addition, this paperdesign a neural network for whole recognition, extract whole feature of cellcharacter string to recognize, while the confidence is low, then go to thesegmentation method for identification.The whole form handwritten number recognition system is based on VSplatform developed by C++. And after testing lots of samples, the result is good.