基于长短期记忆神经网络的手写数字识别

Handwritten Number Recognition Based on Long Short-term Memory Neural Network

手写笔迹识别是模式识别的一个重要研究领域。因为每个人的书写习惯有所不同,导致手写的字体有一定的差异。传统的Softmax模型在手写数字的识别结果上并没有达到人们的期望。目前,深度神经网络框架是模式识别领域的主流方法。长短期记忆神经网络(long-short term memory network,LSTM)是一种特殊的循环神经网络,它由输入门、遗忘门、输出门以及神经元组成。长短期记忆神经网络对于长序列问题有很好的处理。文中提出采用双向长短期记忆神经网络进行手写数字识别。采用MNIST数据集,分别使用传统的Softmax方法和双向长短期记忆神经网络方法对MNIST数据集里的图片进行识别。实验结果表明,传统的Softmax模型的正确率为92%左右,而LSTM模型的正确率达到了96.3%,提升4.3%。

Handwriting recognition is an important research area of pattern recognition.Because each person’s writing habits are different,there are some differences in handwritten font.The traditional Softmax model does not live up to the expectations in the recognition of handwritten numbers.At present,the deep neural network framework is the main method in pattern recognition.The long-short term memory network(LSTM)is a special cyclic neural network consisting of input gates,forgetting gates,output gates and neurons,which can solve the long sequence problems better.In this study,we propose a two-way long-term memory neural network for handwritten digit recognition.The MNIST dataset is applied.The images in the MNIST dataset are identified respectively by the traditional Softmax method and the two-way long-term memory neural network method.The experiment shows that the accuracy of the traditional Softmax model is about 92%,while the correct rate of the LSTM model is 96.3%,an increase of 4.3%.