| 摘要: |
| 本文提出一种Res-Swin模型，通过残差网络（ResNet）模块提取课堂行为图片的局部信息特征，提升模型对图像的计算速度.通过Swin-Transformer进一步分析局部特征提取后的图像全局信息，通过窗口移动分析相邻patch的空间特征关系，再通过合并窗口下采样增大感受野，达到提高图像分类准确度的目的.基于公开数据源的课堂行为图像库，将Res-Swin模型与其他基线进行了比较，实验结果验证了该模型在处理课堂行为图像分类问题时兼顾了准确度与计算速度. |
| 关键词: 课堂行为分类 局部与全局特征 计算速度 准确度 |
| DOI：10.20192/j.cnki.JSHNU（NS）.2025.02.016 |
| 分类号:TP301 |
| 基金项目:国家自然科学基金（62201350） |
|
| Res-Swin model for student classroom behaviour image classification based on local and global feature extraction |
|
XIA Zhengyu, GUO Chang
|
|
College of Information, Mechanical and Electrical Engineering, Shanghai Normal University, Shanghai 201418, China
|
| Abstract: |
| A Res-Swin model was proposed in this paper to extracte the local information features of the classroom behaviour images through the residual network （ResNet） module， which improved the computational speed of the model on the images. The image vectors with local information extracted were further analyzed by Swin-Transformer. To analyze their global features and spatial features of adjacent patches， the patches were analyzed by window shifting. Furtherly， the receptive field was increased by merging the window to downsample， which could improve the accuracy of image classification. Based on the classroom behaviour image of public data sources， the Res-Swin model was compared with other baselines. The experimental results verified that the proposed Res-Swin model could reach a balance between the accuracy and computational speed in dealing with classroom behaviour image classification problem. |
| Key words: class behaviour classification local and global feature computational speed accuracy |
