Research On Deep Learning And Its Application In The Handwritten Chinese Character Recognition

Handwritten Chinese character recognition is an important part forhuman-computer interaction, so that solving the key difficult problem has veryimportant practical value. Chinese character recognition, because of the numerouscategories of Chinese characteristics, which their glyph structure is complex, at thesame time there are many similar Chinese character glyph, and writing style is variousfrom person to person, etc., that has been a difficult and hot research topic in the fieldof pattern recognition. Deep learning mainly includes Deep Belief Networks (DBNs),the Convolutional Neural Networks (CNNs), that the two model can automaticallycapture samples’ probability distribution or learning sample characteristics’ features.With the two advantage, deep learning can avoid the problem of handwritten Chinesecharacter glyph feature extracting manually, although there is the difficult problemswhile training, so in this paper we studies the deep learning especially theconvolutional neural networks, and used it in the problem of handwritten Chinesecharacter recognition, and made the improvement and try to solve the training probl em.Through the research about bottleneck of the recognition rate of the off-linehandwritten Chinese character, recognition rate of the handwritten Chinese characterrecognition is low due to the difficulty to extract the distinct feature from the similarhandwritten Chinese character glyphs. CNNs model for10categories of similarhandwritten Chinese character recognition are builded, and random elastic deformationof the Chinese characters for sample expanding is proposed, which improved thegeneralization ability of CNNs model. Through comparison with the traditional methodfor the similar handwritten Chinese character recognition, we proves that the CNNsmodel can gain high recognition rate, avoiding the problems of feature extracting; Themethod of sample expanding with random elastic deformation can improve the CNNsmodel’s generalization capacity for handwritten Chinese character recognition.By testing CNNs model’s performance for different number of categories, wemapped the CNNs model’s performance curve, descriping the performance to differentnumber of categories about handwritten Chinese character recognition problem, andpointed out that increasing the size of the CNNs model to solve to the problem ofhandwritten Chinese character recognition. According to the problem that a CNNs’ sizeincreasing lead to the learn difficulty increasing, we proposed a training method calledaccumulating category, namely, first preliminary training CNNs model by a few categories of samples, then the model has been learnt is reserved as the initial modelfor training more categories of characters, then gradually increase the categories of themodel for training. This way can make CNNs model rapidly convergence. For500categories of handwritten Chinese character recognition experiment results show thatwith this method the CNNs model’s training times decreased to25%of the originalmethod, at the same time recognition error rate reduced by0.5%to1.1%relatively.Then we conclude that the proposed method is effective.