Deep Model Based Offline Handwritten Chinese Character Recognition

Handwritten Chinese character recognition has broad application prospects in accessible reading, translation, document entry, dealing with bank notes, sorting postal mails and express parcel and other fields, facilitate users to enter information quickly and improve the efficiency of all works in life. It involves machine learning, pattern recognition, digital signal processing, artificial intelligence, natural language processing, statistics, information theory and many other subjects. Because of different people writing style differences, adhesion between the strokes of Chinese characters and other factors, off-line handwritten Chinese character recognition is one of the most difficult problems in the field of OCR. Therefore, this paper presents an offline handwritten Chinese character recognition method based on deep learning.For the unobstructed images of handwritten Chinese characters, this paper uses a classification algorithm based on deep convolutional neural networks and takes the preprocessed image data as the input of classifier directly. Because the deep convolution neural network has the feature of easy over fitting, this paper proposes several methods to expand the training data set such as sliding window and elastic deformation, improving the robustness and generalization ability of the model. This paper uses several convolution neural networks with different structures to classify handwritten Chinese character images, and then mix them together to further improve the model effects.For text damaged, text blocked by stains and other conditions in ancient books scanning and handwritten manuscript scanning, this paper proposes a method to recognize sheltered handwritten Chinese characters based on deep recurrent neural network and deep belief network. First, using two deep recurrent neural networks to extract high level abstract features in images of unobstructed handwritten Chinese and sheltered handwritten Chinese, and then training deep belief network to convert the sheltered handwritten Chinese characters feature space to the unobstructed handwritten Chinese characters feature space. At last, adding CNN features to improve classification accuracy.Experimental results shows that the unobstructed handwritten Chinese character image recognition algorithm and the sheltered handwritten Chinese character recognition algorithm proposed by this article have good recognition results.