深度有序度量学习改进的课堂参与度识别  

Deep Ordinal Metric Learning for Improved Classroom Engagement Recognition

作　　者：董瑶 徐敏[1] 周丽娟[1] 陈文龙[1] 周修庄 张晓宇 DONG Yao;XU Min;ZHOU Li-juan;CHEN Wen-long;ZHOU Xiu-zhuang;ZHANG Xiao-yu(Information Engineering College,Capital Normal University,Beijing 100048,China;School of Artificial Intelligence,Beijing University of Posts and Telecommunications,Beijing 100876,China;The 47th Institute of China Electronics Technology Group Corporation,Shenyang 110000,China)

机构地区：[1]首都师范大学信息工程学院,北京100048 [2]北京邮电大学人工智能学院,北京100876 [3]中国电子科技集团公司第四十七研究所,沈阳110000

出　　处：《小型微型计算机系统》2023年第4期798-804,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(62177034,61972046)资助;国家重点研发计划项目(2019YFB1406301)资助;北京自然科学基金项目(4202051)资助。

摘　　要：在线课堂已成为当前最重要的教学场景之一.针对当前基于视觉特征学习和距离度量学习的参与度识别模型仍然缺乏足够的判别力和稳定性等问题,本文将有序深度度量学习方法引入课堂参与度识别任务,联合建模视觉特征学习和判别度量学习.首先,提出一个有序度量损失函数建模参与度样本的有序标签结构,使得在学习获得的有序度量空间中,视觉特征与其参与度标签保持有序一致性,提高识别模型的判别力.其次,提出一种四元组困难样本构造策略,对困难正样本的视觉特征相似度进行最小化,同时最大化困难负样本的视觉特征相似度,充分挖掘困难样本,提高模型训练的效率和稳定性.最后在课堂参与度基准数据集DAiSEE上进行测试,验证了算法的有效性.Online classrooms have become one of the most common teaching scenarios today,wherein automatic engagement recognition is an important auxiliary tool to improve the efficiency of online teaching.However,the current engagement recognition models based on feature learning still lack enough discrimination and robustness.To address the issues,we introduce a deep ordinal metric learning approach in this paper,which models visual feature learning and ordinal deep metric learning jointly.Firstly,we utilize an ordinal metric loss to model the engagement samples,so that the visual features of engagement are consistent with the engagement labels in the learned metric space.In addition,we propose a quadruplet hard sample mining strategy,which minimizes the visual feature similarity of hard positive samples,and maximizes the similarity of hard negative samples simultaneously.Experimental results on the widely used DAiSEE dataset demonstrate the effectiveness of theproposed method.

关 键 词：深度学习 课堂参与度识别 视觉特征 有序度量学习 困难样本 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]