作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2022, Vol. 48 ›› Issue (4): 262-268. doi: 10.19678/j.issn.1000-3428.0061107

• 开发研究与工程应用 • 上一篇    下一篇

基于注意力机制与特征融合的课堂抬头率检测算法

倪童, 桑庆兵   

  1. 江南大学 人工智能与计算机学院, 江苏 无锡 214122
  • 收稿日期:2021-03-12 修回日期:2021-05-02 发布日期:2021-05-12
  • 作者简介:倪童(1995—),男,硕士研究生,主研方向为图像处理、计算机视觉;桑庆兵,副教授、博士。
  • 基金资助:
    江苏省自然科学基金(BK20171142)。

Class Head up Rate Detection Algorithm Based on Attention Mechanism and Feature Fusion

NI Tong, SANG Qingbing   

  1. School of Artificial Intelligence and Computer, Jiangnan University, Wuxi, Jiangsu 214122, China
  • Received:2021-03-12 Revised:2021-05-02 Published:2021-05-12

摘要: 课堂教学是整个教育任务中的重要环节,教育信息化的发展为提升教学管理水平提供了更多方案。为加强教学情况正反馈,提高课堂抬头率检测的准确性,提出一种结合注意力机制和特征融合的新型检测算法。将原图及视觉特征RGB difference作为网络输入,令其经过特征提取网络后得到信息更丰富的深层特征。在此基础上,提出一种改进的注意力模型(ICBAM)并加载至特征提取网络上,ICBAM使用通道注意力模块和空间注意力模块并行的双流结构,提升网络的特征提取能力。在通道注意力和空间注意力中加入空洞卷积以过滤输入特征中的冗余特征,减少网络对背景等无用特征的关注。此外,设计精炼模块优化预测结果,并在所提算法的基础上实现课堂行为分析软件的开发与应用。实验结果表明,该算法在抬头率检测数据集RDS上的平均抬头率误差为15.648%,相比于SolvePnP等主流检测算法具有更低的误差率。

关键词: 抬头率, 课堂视频, 注意力机制, 特征融合, 空洞卷积

Abstract: Classroom teaching is an important part of the educational task, and the development of educational informatization provides more schemes for improving the level of teaching management.To strengthen the positive feedback of teaching situations and improve the accuracy of classroom head-up rate detection, a detection algorithm for the classroom head-up rate combined with the attention mechanism and feature fusion is proposed.In addition to the original image, the network input uses another visual feature:the Red-Green-Blue (RGB) difference.Following the feature extraction network, the two kinds of inputs are fused to obtain more-abundant deep features.An improved attention model, the Improved Convolutional Block Attention Module (ICBAM), is proposed to be loaded on the feature extraction network.ICBAM uses the parallel dual-flow structure of the channel attention and spatial attention modules, which can improve the feature extraction ability of the network.Hole convolution is added to channel attention and spatial attention to filter the redundant features in the input features.In addition, the refining module is designed to optimize the prediction results further, and the development and application of classroom behavior analysis software are realized based on the proposed algorithm.The experimental results show that the average head-up rate error of this algorithm is 15.648% on the header rate detection data set RDS, which is lower than that of SolvePnP and other mainstream detection algorithms.

Key words: head up rate, classroom video, attention mechanism, feature fusion, dilated convolution

中图分类号: