基于卷积神经网络的脱机单个手写汉字识别

Offline single handwritten Chinese character recognition based on Convolutional Neural Network

本文在Goog Le Net网络基础上搭建了一个适合手写汉字识别的卷积神经网络。研究建立了新的手写汉字训练集,新训练集综合了现有的训练集并剔除了其中的错误,同时加入印刷体训练集,增加书写风格的多样性。训练神经网络时采用随机梯度下降算法,并加入动量项加速网络参数的收敛,使用正则项防止过度拟合,最终训练出的神经网络在训练集上的正确率为99.56%,在验证集上的正确率达到96%,并具有很好的泛化能力。

This paper builds a Convolution Neural Network suitable for handwritten Chinese character recognition based on GoogLeNet network. The research has set up a new training set of handwritten Chinese characters. The new training set integrates the existing training set and removes the errors in them. At the same time,the research has added a printed training set to increase the variety of handwriting styles. In the training of neural network,a stochastic gradient descent algorithm is used,and the momentum term is added to accelerate the convergence of the network parameters. The regular item is used to prevent over-fitting. The final neural network reaches an accuracy rate of 99. 56% on the training set,and 96% on the validation set,which has a good generalization.