摘要： 传统卷积神经网络对于特征不明显或歧义性大的图像识别率较低。针对该问题,在卷积神经网络的基础上通过增加局部特征提取层和概率权重综合层,构建基于局部特征的卷积神经网络模型。该模型对输入图像的局部进行识别,得到局部图像的分类概率信息,综合分析所有局部图像的分类概率信息得到最终网络输出。手写字符识别实验结果表明,与经典的卷积神经网络模型相比,该模型识别率较高,尤其是在输入图像特征较为模糊的情况下优势更为明显。

关键词: 深度学习, 卷积神经网络, 局部特征, 手写数字识别, 分类概率

Abstract: The traditional Convolutional Neural Network(CNN) is difficult to obtain high recognition rate when the feature of the input image is not obvious.To solve this problem,this paper conatrusts Convolutional Neural Network model Based on Local Feature(CNN-LF),which is built on CNN by adding Local Feature Extration(LFE) layer and Probability Importance Synthesis(PIS) layer.It firstly recognizes the local feature of the input image and gets the information about classification probability,then makes comprehensive analysis on the information of all the local images to get the final result.In the experiments of handwritten digits recognition,CNN-LF gets higher recognition rate compared with traditional network model,especially when recognizing images which have fuzzy features.

Key words: deep learning, Convolutional Neural Network(CNN), local feature, handwritten digit recognition, classification probability

中图分类号: 

• TP183