Research On Deep Learning And Its Application In The Off-line Handwritten Chinese Character Recognition

Handwritten character recognition is an important part of intelligent interaction and improving the recognition rate of handwritten character recognition has important practical value. Chinese character recognition, because of the numerous categories of Chinese characteristics, which their glyph structure is complex, at the same time there many similar Chinese glyphs, and writing style is various from person to person, etc.that has been a difficult and hot research topic in the field of pattern recognition. Deep learning is a new research focus in the field of machine learning and has achieved good recognition performance in many pattern recognition problems. In this paper, we study the application of depth learning theory and convolution neural network model in off-line handwritten Chinese character recognition, the main work is done as follows:In this paper, the training and recognition of the Chinese character image data by the convolution neural network model is built, and the feasibility of the convolutional neural network in the recognition of off-line handwritten Chinese characters is verified. The factors affecting the recognition of handwritten Chinese characters by adjusting the depth of the network and the network structure parameters are analyzed.The Chinese characters of multiple, complex network model of network parameters are not easy to convergence and slow training speed problem was studied based on transfer learning supervised pre training methods, by first on a small character set training, as large character set training network initialization parameters obtained a better set of network parameters, to speed up the convergence of the network, greatly improve the practicability of the network model.We propose using the sample image transformation method and eight direction gradient feature to optimize the convolutional neural network recognition rate method for handwritten Chinese character recognition, experiment results on CASIA_HWDB database shows that, compared to just training when using the original image as training data, eight-direction gradient features and a combination of the original image data as a network recognition can help to reduce the error rate by 22%.The practical application need to identify the rotation character has nothing to do with the situation, we propose the rotation invariant of hog features and based on radon and Fourier transformation, the rotation invariant feature of the handwritten Chinese character recognition method, the image rotation invariant features and the original image with as a convolutional neural network input data of the direction of the Chinese character recognition. In this paper, the CASIA_HWDB data set is randomly selected from a part of the similar words, and the rotation of the similar words is not an experiment for the recognition of handwritten Chinese characters. The experimental results show that the rotation invariant features are used to train the Chinese characters in the non fixed direction, and the error rate is reduced by 23% compared with the original image.