基于改进关系网络的实验数据小样本分类评估方法  

Small sample classification assessment method of experimental data based on improved relation network

作　　者：刘伟发 李晓欢[1] 赵中华[1] 郭庆 唐欣 周文杰 LIU Weifa;LI Xiaohuan;ZHAO Zhonghua;GUO Qing;TANG Xin;ZHOU Wenjie(School of Information and Communication,Guilin University of Electronic Technology,Guilin 541004,China;Gstar Information Technology Co.,Ltd.,Guilin 541004,China)

机构地区：[1]桂林电子科技大学信息与通信学院,广西桂林541004 [2]桂林市国创朝阳信息科技有限公司,广西桂林541004

出　　处：《实验技术与管理》2021年第11期194-199,共6页Experimental Technology and Management

基　　金：国家自然科学基金项目(61762030);广西自然科学基金项目(2019GXNSFFA245007,2018GXNSFDA281013);广西科技计划项目(AA18242021,AB19110050,AA19110044,ZY19183005,AB20238033);桂林市科技计划项目(20190214-3)。

摘　　要：针对实验教学中由于实验内容多样、数据样本严重不足,难以通过传统深度学习方法对学生实验报告中的图像数据进行自动评估问题,提出一种基于改进关系网络的实验数据小样本分类评估方法。该方法将图像分类思想与改进关系网络相结合建立分类评估模型。首先通过注意力残差块构建特征提取模块,用于提取输入实验图像的特征向量;再将标记样本和查询样本的特征向量输入关系模块,通过二者的相似性得分获得查询样本的对应类别,从而实现对实验图像的评估。实验结果表明,在5-way 1-shot、5-way 5-shot条件下,该模型在mini-ImageNet数据集上的分类准确率分别提升了4.5%和1.91%,在CUB数据集上分别提升1.54%和1.03%。该模型基于学校实验室实验数据在5-way 1-shot、5-way 5-shot条件下,分别实现了61.01%和68.91%的分类评估准确率。It is difficult to automatically assess the image data in students’ experiment report through the traditional deep learning method due to the diverse experiments and the serious shortage of data samples in experimental teaching, and a small sample classification assessment method of experimental data based on improved relation network is proposed. This method combines the idea of image classification with the improved relation network to establish a classification assessment model through the attention residual block building feature extraction module to extract the feature vectors of the input experimental images, and then puts the feature vectors of the labeled samples and the query samples into the relation module, and through the similarity scores derive the classification of the query samples, so as to realize the assessment of experimental images. Experiment results show that the proposed model improves the classification accuracy by 4.5% and 1.91% on mini-ImageNet data set, and 1.54%and 1.03% on CUB data set under the conditions of 5-way 1-shot and 5-way 5-shot. Based on the experimental data of the unversity laboratory, the model achieves 61.01% and 68.91% classification and evaluation accuracy respectively under the conditions of 5-way 1-shot and 5-way 5-shot.

关 键 词：小样本学习 关系网络 注意力机制 实验教学 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] G642.0[自动化与计算机技术—计算机科学与技术]