基于数据增强的小样本图像分类方法  

Few-shot image classification method based on data enhancement

作　　者：尹甜甜 王新[1] 邓亚萍 施国兴[1] YIN Tian-tian;WANG Xin;DENG Ya-ping;SHI Guo-xing(School of Mathematics and Computer Science,Yunnan Minzu University,Kunming 650500,China)

机构地区：[1]云南民族大学数学与计算机科学学院,云南昆明650500

出　　处：《云南民族大学学报（自然科学版）》2023年第1期83-89,123,共8页Journal of Yunnan Minzu University:Natural Sciences Edition

基　　金：国家自然科学基金(61363022);云南民族大学研究生创新基金(SJXY-2021-010).

摘　　要：小样本图像分类训练样本过少,若直接用深度学习的方法对其处理会出现过拟合现象,且存在训练好的模型不能很好的泛化到测试任务上等问题.针对以上问题,提出一种基于数据增强的算法去缓解模型过拟合,并结合深度学习网络wide-ResNet28来提升模型的分类性能.此方法没有引用外部数据对当前任务进行数据扩充,而是借助基类数据的语义先验信息对新类数据的特征进行补充,在形成新的特征分布上进行数据增强.该方法在MiniImageNet和Cub 2个小样本数据集上进行实验,图像特征提取的精确度分别达到83.46%、91.61%,验证了该方法的有效性.There are too few training samples of few-shot image classification,and if it is directly processed by deep learning method,there will be over-fitting phenomenonand problems.For example,the trained model can not be well generalized to test tasks.To solve the above problems,an algorithm based on data enhancement was proposed to alleviate the overfitting of the model,and the deep learning network wide-ResNet28 was combined to improve the classification performance of the model.This method does not refer to external data for data expansion of the current task,but supplements the features of the new class data with semantic prior information of the base class data,and enhances the data in the formation of new feature distribution.The method was tested on MiniImageNet and Cub two small sample data sets,and the accuracy of image feature extraction reached 83.46%and 91.61%,respectively,which verified the effectiveness of the method.

关 键 词：小样本学习 深度学习 数据增强 图像分类 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP18[自动化与计算机技术—计算机科学与技术]