摘要
针对目前基于卷积神经网络模型(CNN)手写数字辨识算法收敛速度慢、识别率低的问题,设计一种CNN网络模型。在模型训练时,改进模型学习率,使学习率指数可以动态衰减;使用 Dropout正则化方法,提高模型的泛化能力;与批量随机梯度下降法、Momentum算法、Adagrad算法、RMSprop算法、Adam算法等参数优化方法作比较。实验结果表明:基于RMSprop或Adam的优化算法CNN模型在对MNIST数据集进行训练时,算法收敛速度快、测试集识别准确率为99.40%或99.70%。
Aiming at the problems of slow convergence speed and low recognition rate of handwritten numeral recognition algorithm based on convolutional neural network model (CNN),we designed a CNN network model.The learning rate of the model was improved so that the learning rate index can be attenuated dynamically.Dropout regularization method was used to provide the generalization ability of the model.Finally,we compared this method with the batch random gradient descent method,Momentum algorithm,Adagrad algorithm,RMSprop algorithm and Adam algorithm emphatically compared.The experimental results show that when training MNIST data sets based on RMSprop or Adam optimization algorithm CNN model,the algorithm converges fast and the recognition accuracy of test sets is 99.40% or 99.70%.
作者
张烁
张荣
Zhang Shuo;Zhang Rong(Computer Science Department,Shanxi Youth Vocational College,Taiyuan 030024,Shanxi,China;Computer Instruction,Shanxi Medical University,Taiyuan 030001,Shanxi,China)
出处
《计算机应用与软件》
北大核心
2019年第8期172-176,261,共6页
Computer Applications and Software
基金
山西医科大学校级基金项目(XJ2018099)
关键词
CNN
参数优化
手写数字
CNN
Parameter optimization
Handwritten numbers