Research Of Character Recognition Algorithms On Convolution Neural Network

Character recognition technology,which has been put forward in 1960 s, is widely used in many fields like data statistics,intelligent transportation, finance and taxation and so on. Despite the amount of work and extensive applications on character recognition at the past several decades, it is still a considerable challenge nowadays due to numerous categories,multifarious handwriting style and assembler characters etc..This paper investigates and analysis the character recognition algorithm in the domestic and overseas,especially these algorithms about the constructional neural network.Accuse of the hierarchical work mechanism, which is beneficial to avoid the Bojangles prepossessing and complicated feature extraction process, convolution neural network has natural advantages in the field of handwritten character recognition. This paper studies the mainstream character recognition algorithm of constructional neural network,then puts forward two new convolution neural network structure, which achieves a good recognition accuracy on character recognition. The main content of this paper are as follows:1. By test the CNN performance on characters, we do a research based on MNIST handwritten digital data sets and analysis the critical factors which have a important influence on the CNN, such as the depth of the constructional neural network, the scale of slide-window and the pooling method etc.. In order to enhance the generalization ability of the model, we do a further comparison on various nonlinear transformation technology such as activation function and regularization method.2. Experiments and analyses on the feature extraction of CNN are also carried out.Two mufti-stage feature extraction modules have been put forward and replace the place of the original module in the CNN with single structure, achieving a better recognition accuracy than the traditional single-stage CNN. To further analyses, we extract the feature maps of CNN with single-stage structure and mufti-stage structure and visualize them. Features visualization makes feature changes is easy to be observed. The experiments have been down on the offline handwritten Chinese characters CASIA-HWDB1.1 data sets.3.The paper takes CNN structure as a special feature extractor and extract the CNN features, then combines CNN features with the the traditional features such as Gaborfeatures and Gradient features to feature fusion, the traditional support vector machine classifier is used to classification. The CNN+Gabor+Gradient+SVM algorithm proposed in this paper gets a good performance on similar handwritten Chinese characters, which is superior to the ordinary CNN model and the traditional classifier.