Scholarlink

<p>To solve the limitations of traditional classroom assessment methods that are highly subjective and poor in real-time performance, this study constructs a set of automated classroom quality assessment methods that are both high-precision and efficient. This study proposes an improved residual neural network-18 model that integrates Ghost lightweight convolution and squeeze-excitation channel attention mechanism for student face detection. To evaluate the learning status, a multi-scale deep separable pruned capsule network is constructed for facial expression recognition and quantifying classroom concentration. The results showed that the improved residual neural network-18 model had a detection accuracy of 98.05% on the CASIA-WebFace dataset, with only 1.05G of calculations, and its performance and efficiency exceeded the benchmark model. In expression recognition, the AUC-ROC of the multi-scale depth separable pruning capsule network model reached a maximum of 0.965, which was better than the original capsule network. The concentration assessment application could effectively distinguish different teaching scenarios. The average score of the experimental operation class was as high as 0.702, and the concentration was significantly higher than that of the theoretical class. This showed that the face detection and expression recognition model used could achieve accurate and efficient student face detection and learning status assessment. This study provides an innovative technical solution for real-time and objective assessment of smart classroom quality, which has important application value.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Student face detection based on ResNet18 residual network for classroom quality assessment

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月11日

查看全文
  • Xin Li

摘要

To solve the limitations of traditional classroom assessment methods that are highly subjective and poor in real-time performance, this study constructs a set of automated classroom quality assessment methods that are both high-precision and efficient. This study proposes an improved residual neural network-18 model that integrates Ghost lightweight convolution and squeeze-excitation channel attention mechanism for student face detection. To evaluate the learning status, a multi-scale deep separable pruned capsule network is constructed for facial expression recognition and quantifying classroom concentration. The results showed that the improved residual neural network-18 model had a detection accuracy of 98.05% on the CASIA-WebFace dataset, with only 1.05G of calculations, and its performance and efficiency exceeded the benchmark model. In expression recognition, the AUC-ROC of the multi-scale depth separable pruning capsule network model reached a maximum of 0.965, which was better than the original capsule network. The concentration assessment application could effectively distinguish different teaching scenarios. The average score of the experimental operation class was as high as 0.702, and the concentration was significantly higher than that of the theoretical class. This showed that the face detection and expression recognition model used could achieve accurate and efficient student face detection and learning status assessment. This study provides an innovative technical solution for real-time and objective assessment of smart classroom quality, which has important application value.