摘要
卷积神经网络是图像分类领域效果卓越的深度学习算法,然而训练深度神经网络是一项繁琐且复杂的工作,不仅在结构设计上依赖开发人员丰富的经验,还容易产生过拟合现象。因此,该文提出一种基于模型迁移的图像识别方法,该方法能够在简化设计思路的同时极大地提升卷积神经网络的性能。此外还在三个小型图片集上进行了多次模型训练和对比分析。研究结果表明,经过迁移学习优化的卷积神经网络的测试集准确率均得到显著提升。
Convolutional neural network is an effective deep learning algorithm in the field of image classification.However,training the deep neural network is a tedious and complex work,not only relying on the rich experience of developers in structural design,but also prone to over fitting.Therefore,this paper proposes an image recognition method based on model migration,which can greatly improve the performance of convolutional neural network while simplifying the design idea.In addition,multiple model training and comparative analysis are conducted on three small image sets.The research results show that the test set accuracy of convolutional neural network optimized by Transfer learning is significantly improved.
作者
张文韬
张婷
ZHANG Wentao;ZHANG Ting(School of Mechanical and Automotive Engineering,Shanghai University of Engineering Science,Shanghai 201620,China)
出处
《现代信息科技》
2023年第14期57-60,共4页
Modern Information Technology
基金
国家自然科学基金资助项目(11702168)。
关键词
图像识别
深度学习
卷积神经网络
迁移学习
预训练模型
image recognition
Deep Learning
Convolutional Neural Network
Transfer Learning
Pre-trained Model
作者简介
张文韬(1997-),男,汉族,河南平顶山人,硕士研究生在读,研究方向:深度学习与故障诊断。
