基于卷积神经网络的MNIST手写数字识别优化研究  

Research on optimization of MNIST handwritten digit recognition based on convolutional neural network

作　　者：牟世桂 MOU Shigui(Jilin Agricultural University,Changchun 130118,China)

机构地区：[1]吉林农业大学,长春130118

出　　处：《计算机应用文摘》2025年第3期81-83,86,共4页

摘　　要：文章旨在优化基于卷积神经网络的MNIST手写数字识别,通过引入残差连接、Dropout层、Batch Normalization层,以及优化算法与学习率调度器来提升模型性能。这些技术的综合应用旨在提高模型在数字识别任务中的精确度。其中,首先采用基于ResNet结构的卷积神经网络,结合Dropout层、Batch Normalization层和残差连接来构建模型。其次,使用SGD优化算法配合学习率调度器和数据增强技术对模型进行训练和优化。研究结果表明,该模型在MNIST测试集上达到了99.5%的准确率,相比传统方法有了显著提升。这些成果不仅在提升数字识别的准确度上取得了显著进展,还证明了在实际应用中综合考虑优化算法、数据增强技术以及网络结构调整对于提升模型性能的重要性,对于推动图像识别技术的发展具有重要的应用价值和实际意义。This article aims to optimize MNIST handwritten digit recognition based on Convolutional Neural Networks by introducing residual connections,Dropout layers,Batch Normalization layers,as well as optimizing algorithms and learning rate schedulers to improve model performance.The comprehensive application of these technologies aims to improve the accuracy of the model in digital recognition tasks.Firstly,a convolutional neural network based on ResNet structure is adopted,combined with Dropout layer,Batch Normalization layer,and residual connections to construct the model.Secondly,the SGD optimization algorithm is used in conjunction with a learning rate scheduler and data augmentation techniques to train and optimize the model.The research results indicate that the model achieved an accuracy of 99.5%on the MNIST test set,which is a significant improvement compared to traditional methods.These achievements have not only made significant progress in improving the accuracy of digital recognition,but also demonstrated the importance of comprehensively considering optimization algorithms,data augmentation techniques,and network structure adjustments to enhance model performance in practical applications.They have important application value and practical significance for promoting the development of image recognition technology.

关 键 词：MNIST CNN ResNet 模型优化 DROPOUT 学习率衰减 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]