The Research And Development Of Embedded Character Recognition Technology

With the increasing degree of automation of production, the automatic recognition of theproduct label or code in production line through machine is needed.However, the complicatedindustrial environments, such as uneven illumination,noisy,etc.,bring greater difficulties toaccurately identify the characters.People also put forward higher requirements of thehardware and the software of character recognition system.In this paper,research is based onembedded intelligent camera hardware platform,and identify object is the pressed characterson bearing.Summarizing the predecessors' experience and work,the author puts forward hisown solution,and achieves the automatic recognition of the pressed characters onbearings.Paper involves the the following study content:1) Bearing image capture and preprocessing.On the basis of the analysis of imagecharacteristics of pressed characters on bearing,image acquisition system which is suitable forbearing is designed.The system is based on embedded intelligent camera platform for imageacquisition,and low-angle LCD ring light source is chosen as auxiliary illuminant.It isconfirmed that the system can get high quality bearing image.In the image pre-processingstage,the structure ofbearing is a circle structure and the position of the character is notfixed,so it is needed to locate bearing character area and correct bearing character forsubsequent processing and recognition.Therefore,after getting bearing binary image by Otsumethod,the center of the bearing should be located firstly. Because of the deficiencies ofRandomized-Hough-Transformation for the circle detection algorithm,an improvedRandomized-Circles-Detection algorithm for the bearing center localization is proposed inthis paper.This method reduces the computational complexity through the simplest decisionsin the process of choosing a candidate circle.The experiment proves the circle centerlocalization algorithm proposed can locate the bearing circle quickly and accurately.Then,usethe center as the origin of polar coordinates and the bearing character area is determined byprojection method.Finally,in the process of correcting bearing character area,there would be alot of burrs in the result image if we use the polar coordinates transform method.Thus,thispaper proposes a character correction method based on affine transformation,the experimentalresults show that the method can transform the image less burr and the speed of thetransformation also can meet the requirements.2) Bearing character feature extraction and classifier.Two commonly used methods offeature extraction are describes in this paper:the direction line element feature extractionmethod and the contour feature extraction method.After analysing their deficiency,twoimproved feature extraction method are proposes in this paper:the direction line elementfeature extraction method based on elastic gird and the contour features extraction methodbased on wavelet analysis.The improved direction line element features make up for thesensitivity of uniform gird division of characters to character deformation which are morerobust.The improved contour features make full use of the characteristics of waveletdecomposition,which have better noise resistance and lower dimension.About the classifier ofbearing character recognition,this paper analyzes BP neural network and support vector machine classifier principle.BP neural network based on momentum term with adaptivelearning rate and SVM based on sequential minimal optimization are designed in thispaper.The effectiveness of the method proposed is confirmed by experiments on real bearingimage,and the method of feature extraction and classifier are determine by experiments results.Feature extraction method and classifier is determined by experiments.Experiments show that the proposed scheme can obtain high quality bearing image,andthe effect of preprocessing methods is better, finally choose the improved contour features anduse support vector machine as classifier.Character recognition accuracy is 96% or more,andthe speed can meet the practical demands. The method proposed in this paper canautomatically recognize bearing characters quickly and accurately.