Seite - 110 - in Document Image Processing

J. Imaging 2018,4, 43 Figure8.SchemeofNPWonKirschfeatures [28]. 3.3.2.UnsupervisedLearningFeatureandNeuralNetwork With the aimof improving the performance of our proposed feature extractionmethod,we continuedour research on isolated character recognition by implementing the neural network as classifier. In this second step [29], the same combination of feature extractionmethodswasused and sent as the input feature vector to a single-layer neural network character recognizer. In additiontousingonly theneuralnetwork,wealsoappliedanadditionalsub-module for the initial unsupervised learningbasedonK-Meansclustering(Figure9). This schemawas inspiredbythestudy ofCoatesetal. [52,53]. Theunsupervised learningcalculates the initial learningweight for theneural networktrainingphase fromtheclustercentersofall featurevectors. Train Images Features Extraction Unsupervised Learning with K-Means Clustering Feature Centers as Initial Weights for Neural Network Train with Neural Network Final Trained Weight of Network Test Images Features Extraction Recognition Results Instead of using Random Initial Weights Figure 9. Schema of character recognizerwith feature extractionmethod, unsupervised learning feature,andneuralnetwork[29]. 3.3.3.ConvolutionalNeuralNetwork Themultilayerconvolutionalneuralnetworks (CNN)haveprovenveryeffective inareassuch as image recognition and classification. In this evaluation experiment, a vanilla CNN is used. Thearchitectureof theCNN(Figure10) isdescribedasfollows(thisarchitecturehasalsobeenreported inKhmer isolated character recognition baseline in [21]). The grayscale input images of isolated characters are rescaled to48× 48pixels in size andnormalizedbyapplyinghistogramstretching. 110