摘要
在手写阿拉伯数字识别系统中,传统手写数字识别算法准确率不高。基于LeNet-5卷积神经网络算法,提出一种轻量化的卷积神经网络算法。利用全局平均池化算法代替原有的全连接层降低算法训练参数的规模;利用RELU函数代替tanh函数作为激活函数、增加BN层等方法来加速网络模型的收敛速度。利用MNIST手写数据集设计了验证实验,实验表明,改进后的算法不但训练参数规模只有LeNet-5训练参数规模的1/3,而且相较于LeNet-5改进后的算法准确率上升了约0.4%。由此证明该算法具备较高的应用及推广价值。
In the handwritten Arabic numeral recognition system,the accuracy rate of traditional handwritten digit recognition algorithm is not high.To tackle this problem,based on the LeNet-5 convolutional neural network algorithm,a lightweight convolutional neural network algorithm is proposed.Firstly,the global average pooling algorithm is used to replace the original fully connected layer to reduce the scale of the network model training parameters.Then the RELU function is used instead of the tanh function as the activation function,and the BN layer algorithm is added to accelerate the convergence speed of the network model.The verification experiment was designed with the MNIST handwritten data set as the training set.Experiments on MNIST handwritten data set show that the improved algorithm not only has 1/3 of the training parameter scale of LeNet-5,but also has a 0.4 percentage point increase in accuracy compared with LeNet-5,which proves that the improved algorithm has high application and popularization value.
作者
解修亮
徐晓光
丁理
XIO Xiuliang;XU Xiaoguang;DIOG Li(College of Electrical Engineering,Anhui Polytechnic University ,Wuhu 241000,China)
出处
《钦州学院学报》
2019年第7期46-53,共8页
Journal of Qinzhou University
基金
安徽高校省级自然科学研究项目:局部可柔虚杆算法在城市道路交通中的应用与研究(KJ2014A024)
关键词
数字识别
卷积神经网络
参数规模
准确率
digital recognition
convolutional neural network
parameter quantity
accuracy