基于深度学习的字符识别技术研究

Research on Character Recognition Based on Depth Learning

卷积神经网络因为其高效的识别率使得它迅速成为人工神经网络领域最热门的研究方向。本文,把卷积神经网络应用到MNIST数据集的识别上,获得了较好的识别率。在LetNet-5的基础上设计一个6C-2S-12C-2S的CNN-1网络,采用MNIST数据库的一万个样本训练80次后达到最佳的识别效果。同时,分析三种网络迭代次数和错误率之间的曲线图,得出最佳的CNN-1网络,池化降低训练参数提高网络性能。CNN-1网络经过120次迭代时,错误率仅为1.18%,明显低于其他方法。

The convolution neural network has quickly become the hottest research direction in the feld of artifcial neural networks because of its high recognition rate. In this paper, the convolution neural network （convolution neural network） is applied to the recognition of MNIST data sets, and a better recognition rate is obtained. Design a 6C-2S-12C-2S CNN-1 network on the basis of LetNet-5. The ten thousand samples of the MNIST database are trained 80 times to achieve the best recognition effect. At the same time, the graph between the three network iterations and the error rate is analyzed, and the best CNN-1 network is obtained. The pool reduces the training parameters to improve the network performance. After 120 iterations, the error rate of CNN-1 network is only 1.18%, which is signifcantly lower than other methods.