Extraction And Recognition Of Handwritten Characters In Marking System

The automatic marking system is gradually replacing the traditional manually-marking method because of its high efficiency,easy management and without subjective interference.At present,the commonly used marking system is generally based on the optical marker reading technique,which requires special answer cards and expensive scanning equipment.Meanwhile,this kind of scoring method relies on high quality filling and does not meet the candidate's habits.As a result,misplacement could frequently happen.To solve the problem,a marking system based on optical character recognition technique is designed in this paper.Methods for paper correction character extraction and recognition are studied,and hence a series of algorithms are proposed.The main contribution of this work includes the following three aspects.In the first part of the dissertation,a locator-based test paper correction algorithm is pro-posed for solving the problem of test paper correction.This serves as an important pre-step of handwritten character extraction and recognition,and will directly affects the performance of the marking system.In general,images acquired by scanners have stable quality with s-light noise.However,the situation faced by the digital camera is complex and changeable,and the acquired images could have more serious noise.For that,we correct the test paper images from different sources by detecting the pre-designed locators in the template,thus realizing a locator-based correction algorithm.Experimental results show that the algorithm can be implemented fast and has strong anti-noise performance.In consequence,the test paper image can be corrected stably and accurately.In the second part of the dissertation,for solving the problem of extraction and nor-malization of handwritten characters,an adaptive threshold segmentation algorithm based on gray histogram is proposed.This algorithm obtains an adaptive threshold through the gray histogram of the character region to segment the handwritten characters in the test pa-per image.Since the segmented character images could have different sizes,different gray values and noise,we further design a character normalization method to obtain character images,by which the characters will be easily recognized with a classifier,and the character recognition efficiency is improved.In the third part of the dissertation,for solving the recognition problem of handwrit-ten characters,a 'confusion-aware' convolutional neural network is proposed.In image classification,it is needed to find a clear boundary of each image category.However,the boundaries of some image categories are easy to be confused with each other.To solve this problem,the confusion-aware convolutional neural networks are proposed.At the training stage,a conventional classifier,referred to as the prediction classifier,is trained and its con-fusion matrix is estimated by a cross validation conducted on the training set.Then,with the estimated confusion matrix,a confusion-aware model is established and a set of correction classifiers for those easily-confused image categories are trained.At the classifying stage,the prediction and correction classifiers are collaboratively used via a hierarchical structure,and the confusion-aware model is again used as a connection between them.Experimental results conducted on the Mnist and CIFAR-10 datasets show that the proposed networks perform potentially better than the existing classifiers.Based on the above algorithm,the automatic marking system implemented in this pa-per is able to recognize handwritten characters with high accuracy.On sample papers of thousands of primary and middle school students that we have collected,the stability and accuracy of the system meet the practical application standard.