Deep Convolutional Neural Network And Its Application In Handwritten Chinese Character Recognition

The evolution of artificial intelligence(AI) discipline suffers from the early age of “reasoning and search” to the age of “expert system” that make use of the human being's knowledge to solve the engineering problems. And in recent two decades, the method for AI goes through the age of “machine learning” with large training sets and gradually steps into the age of “deep learning” based on the artificial neuron network of today, which is marked by feature representation learning. This is a big leap, and it not only provides a solution for the long-term problem of automatically extracting feature representation in the field of AI, but also guides a new direction of AI. Actually, deep learning which is the typical technology of feature representation learning has been proposed in 2006, it can generate the feature representations automatically by learning from large amounts of data without people's participation and achieve classification with them. Various deep learning architectures such as deep neural network, deep convolutional neural network(DCNN), deep belief network and recurrent neural network have been applied to fields like computer vision, automatic speech recognition, natural language processing and bioinformatics where they have been shown to produce state-of-the-art results on various tasks.On the basis of arranging and summarizing the development and existing studies related to DCNN at home and abroad, I give an insight into the property of receptive field of DCNN. And also I make a comparison of different regularization methods on the experiments of handwritten Chinese character recognition. The main works in the paper list as follow:1) Based on analyzing the properties of receptive field of DCNN thoroughly, I modeled the relationship between the receptive field and the model's parameter number, model recognition exactly.2) Combining with the theory of receptive field and DCNN, I verified the effect of different size of receptive field to the accuracy of model according to different experiments and provided a reasonable reference for the selection of receptive field based on DCNN.3) A particular DCNN model was designed for the offline handwritten Chinese character recognition. Different regularization methods such as Early-Stopping, Noise and Dropout for model generalization were compared, and the results show that the accuracy and efficiency of Dropout method is prior to others.