基于CNN的多尺度特征在手写数字识别中的应用

Application of Multi-scale Features in MNIST Based on CNN

在手写数字识别数据集(MNIST)情景下,为了提高卷积神经网络的识别正确率,提出了一种改进的基于卷积神经网络(CNN)的多尺度特征识别算法.首先,利用卷积操作和池化操作提取图像中的全局特征及局部特征,通过二次卷积与特征融合获得数字图像的多尺度特征.然后,将多尺度特征送入全连接网络和SoftMax分类器,实现手写数字图像识别.最后,通过对不同网络结构的CNN算法进行评估表明,本文提出的算法可以有效提高网络精度,具有较好的泛化能力.

In order to improve the recognition accuracy of convolutional neural networks(CNN)within MINIST handwritten digit recognition data sets,this paper proposes an improved multi-scale features recognition algorithm based on CNN.Firstly,the global and local features are extracted by convolution and pooling operations,and multi-scale features of digital images are obtained by quadratic convolution and feature fusion.Then,multi-scale features are fed into full-connection network and SoftMax classifier to recognize handwritten digital images.Finally,a comprehensive evaluation of CNN algorithms with different network structures shows that the method can effectively improve network accuracy,and has good generalization ability.