Research On Handwritten Character Recognition Based On Deep Learning

Handwritten character recognition is a typical image recognition problem.Its implemen-tation will save a lot of manpower and time for processing handwritten information,so it has broad application prospects in accounting,postal,and financial fields.With the improvement of artificial neural networks,deep learning has been extensively and deeply studied,and gradually become a research hotspot in image recognition field.Feature extraction is an important step in image recognition.Deep learning has the ability to extract features automatically,which can avoid the steps of manually designing and selecting image features,so the use of deep learn-ing to automatically complete handwritten character recognition is of great significance.The autoencoder is a feature extraction method commonly used in deep learning,and it can extract features from each layer by unsupervised training manager.This paper studies and improves the autoencoder for handwritten character recognition,and enhances the feature extraction ability and classification ability.The main research contents of this paper are summarized as follows:(1)The forward and back propagation in the network training process are studied,and the advantages of deep learning relative to shallow learning are analyzed.The generative,dif-ferentiated and mixed structures in deep learning are distinguished.Moreover,the principle and training process of the restricted Boltzmann machine,convolutional neural network and autoencoder for constructing above structures are deeply studied.Then,in order to obtain char-acter images that are easy to extract effective features,image graying,image denoising,pixel binarization and size normalization in the preprocessing process are studied.(2)The traditional autoencoder does not use category in the process of extracting features,so there is a problem that the features lack information from the category,whereas the category defines the content of image and contains important information about recognition.Aiming at this problem,this paper proposes a combined autoencoder network to extract main information from the image pixels and the category information from the labels.Finally,information ex-tracted in different ways is mixed to form combined features and used to identify the handwritten character images.By comparing the combined autoencoder network with other single feature extraction algorithms,it is found that using the combined features containing main information and category information to identify the handwritten characters can obtain higher recognition accuracy than others.(3)Due to the disappearance of gradient,the stacked autoencoder can not effectively trans-mit category error information to the bottom layer of network during the fine-tuning process.Parameters in bottom layer can not be fully trained,which weakens the recognition ability of multilayer network and affects the recognition accuracy.Aiming at this problem,this paper proposes a multilayer fine-tuning autoencoder network,which fine-tunes the network param-eters after the unsupervised training,so that the category error information can affect directly the training process of parameters in the pre-training phase.By comparing the multilayer fine-tuning autoencoder network with algorithms that only fine-tunes in the last layer,it is found that the proposed method makes the features extraction process more purposeful,which can reduce the influence of gradient disappearance and improve the recognition accuracy.