The Research On Convolutional Neural Networks Based Isolated Handwritten Chinese Character Recognition

The handwritten Chinese character recognition has a wide range of applications, such as automated character processing and intelligent character input. However, handwritten Chinese characters have the characteristics of large vocabulary, complex structure, lots of similar characters and shape diversification, which makes it as a difficulty and hotspot in the field of character recognition.This thesis mainly focuses on the recognition of isolated offline and online handwritten Chinese characters. We take a preliminary research and conduct a lot of experiment based on the previous work and achievement in this field, apply the deep learning techniques to handwritten Chinese character recognition, and get some useful conclusions. The primary contents of this thesis are as follows:First, in this thesis, we utilize deep convolutional neural networks for isolated handwritten Chinese character recognition. Through a series of comparative experiments, we verify the feasibility of using CNNs in this task and analyze its performance. Experimental results show that, benefited from the stronge feature representation ability of CNNs, we can use rather complex deep architectures for handwritten Chinese character modeling and its intrinsic "end to end" characteristic also simplifies the procedure of Chinese character recognition system.Secondly, we take a further study on the traditional methods of Chinese character preprocessing and feature extraction and combined them with CNNs. Compared with using character samples for feature learning directly, we obtained higher accuracies by means of first extracting artificial features (e.g. directional gradient feature, directional element feature), and then use CNNs for feature learning. Experimental results show that the combination of traditional feature extraction method and CNNs’ feature learning can produce better recognition result.Thirdly, due to the large vocabulary of Chinese characters, the network’s parameters are not easily to convergence and the training speed is slow. Aimed at this problem, we apply a supervised pre-training scheme based on the concept of transfer learning. We first train the network based on a small-set data (e.g. handwritten digits in MNIST) in advance to get a better network parameters initialization. Then we can train a new network with the pre-trained model for large-set handwritten Chinese characters. Experimental results show that, this method can make a rapid convergence of parameters for CNN models, as well as keep the recognition accuracy without a fall.Finally, we test the proposed method on the CASIA-HWDB for the task of offline handwritten recognition and CASIA-OLHWDB for the task of online handwritten recognition. The experimental results show that, compared with other methods, the proposed method based on the convolutional neural networks greatly improves the recognition rate for isolated handwritten Chinese characters.