摘要 深度学习模型应用于小样本图像分类时,存在训练时间过长和过拟合的问题。鉴于此,提出了一种基于迁移学习的小样本图像分类方法。首先,将MobileNet-V2,Ineption-V3,Xception 3种深度卷积神经网络放在大型数据集中进行预训练,然后保留并冻结在源网络预训练过程中的基本参数,用数据增强的方法强化小样本数据后,再对小样本的目标数据集进行特征提取训练,最后对预训练的网络模型进行微调,并解冻部分层次,用于调整网络权重,并再次训练目标数据集。实验结果表明,迁移学习在小样本图像的应用中是有效的,可以构造出泛化性能很高的模型,大大减少了原深度模型训练时产生的过拟合问题。 When deep learning model is applied to small sample images classification,there would be such problems as the long training time and over fitting.A small sample images classification method based on transfer learning is proposed.Firstly,three kinds of deep convolution neural networks,MobileNet-V2,Ineption-V3 and Xception,are pre-trained in large data sets.Then,the basic parameters in the process of source network pre-training are preserved and frozen.After strengthening the small sample data with the method of data enhancement,the target data sets of small sample are trained by feature extraction.Finally,the pre-trained network model is fine tuned.And part of the hierarchy is unfrozened to adjust the network weight and the target data set is trained again.The experimental results show that transfer learning is effective in the application of small sample images,which can construct a model with strong performance and greatly reduce the over fitting problem of the original depth model training.
机构地区 安徽理工大学计算机科学与工程学院
出处 《湖北理工学院学报》 2021年第2期27-32,共6页 Journal of Hubei Polytechnic University
关键词 深度学习 迁移学习 深度卷积神经网络 数据增强 deep learning transfer learning deep convolutional neural network data enhancement
