基于转移学习的小样本数据深度学习研究  

Research on Deep Learning Based on the Small Sample Data of Transfer Learning

作　　者：赵颖 ZHAO Ying(Qinghai Radio and Television University,Xining 810008,China)

机构地区：[1]青海广播电视大学继续教育学院

出　　处：《长江工程职业技术学院学报》2019年第3期14-17,共4页Journal of Changjiang Institute of Technology

摘　　要：卷积神经网络的深度学习在图像识别领域取得了巨大的成功,但是训练一个深度学习网络需要大量的数据样本。在实际工作中,很难得到大量的训练样本,在数据集有限的情况下,容易过度拟合。针对这一问题,设计了一种基于转移学习的深度卷积神经网络来解决小样本数据集的问题。采用数据扩充的方法来扩大样本数据集的数量,利用转移学习将训练好的网络(CNN)从大样本数据集中转移到的小样本数据集中进行二次训练,使用全局平均池而不是全连接层来训练网络,并利用Soft max进行分类。该方法解决了深度学习中样本数据集小的问题,提高了操作效率。实验结果表明,该方法对小样本数据集的分类具有较高的识别率。Deep learning of convolutional neural network has achieved great success in the field of image recognition,but training a deep learning network requires a large number of data samples.In practical work,it is difficult to get a large number of training samples,and it is easy to over-fit when the data set is limited.To solve this problem,a deep convolutional neural network based on transfer learning is designed to solve the problem of small sample data sets.The data expansion method is adopted to expand the number of sample data sets,the trained network(CNN)is transferred from the large sample data sets to the small sample data sets by transfer learning for secondary training,Global Average Pond is used to train the network instead of the full connection layer,and Soft Max is used for classification.This method solves the problem of small sample data sets in deep learning and improves operational efficiency.Experimental results show that this method has high recognition rate for the classification of small sample data sets.

关 键 词：转移学习 小样本 深度学习 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]