«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

期数:

2025年02期

页码:

16-23

栏目:

媒体计算

出版日期:

2025-02-10

文章信息/Info

Title:

Infants Expression Recognition in Intelligent Nursing Scene

文章编号:

1673-629X(2025)02-0016-08

作者:

王一评; 王改华

天津科技大学 人工智能学院,天津 300457

Author(s):

WANG Yi-ping; WANG Gai-hua

Institute of Artificial Intelligence,Tianjin University of Science and Technology,Tianjin 300457,China

关键词:

婴幼儿; YOLOv8; 表情识别; 轻量化; 注意力机制; 知识蒸馏

Keywords:

infants; YOLOv8; facial expression recognition; light weight; attention mechanism; knowledge distillation

分类号:

TP391.41;TP183

DOI:

10.20165/j.cnki.ISSN1673-629X.2024.0307

摘要:

基于深度学习的表情识别方法逐渐应用于婴幼儿智能看护场景下,但大多数方法都是基于实验环境下的单人、正面图像的识别,对真实场景下多人表情识别效果欠佳;且深度网络模型存在训练耗时过长、数据处理速度慢的问题,导致表情识别的实时性难以得到保障。 因此,该文提出一种改进的轻量型婴幼儿表情识别方法。 首先,它在 YOLOv8 的基础上,结合深度可分离卷积和注意力机制理论,设计了轻量化注意力机制网络结构,减少了网络的参数并有效提取出显著特征;其次,搭建知识蒸馏框架,以 YOLOv8L 为教师网络模型,利用 Channel-wise Knowledge Distillation(CWD)知识蒸馏的方式,采用改进的损失函数对学生模型进行蒸馏,提高婴幼儿的表情识别准确率及算法的实时性;最后,通过实验验证,该网络在智能看护场景下的婴幼儿(0-3 岁)表情数据集上表现出色,mAP 达到了 73. 4% 。 参数量仅有 2. 6 MB,便于部署在边缘设备上,对比其他表情识别算法,在 CK+数据集上准确率达到 99. 1% ,参数量也较小。

Abstract:

The expression recognition method based on deep learning is gradually applied to the intelligent nursing scene of infants.However,most methods are based on the recognition of single and front images in the experimental environment,and the recognition of multiple expressions in real scenes is poor. Moreover,the deep network model has long training time and slow processing speed,resulting in real-time guarantee of expression recognition. Therefore,we propose an improved method for lightweight expression recognition in infants. Firstly,based on YOLOv8,we introduce a lightweight attention mechanism network structure by integrating deep separable con-volution and attention mechanism. This design effectively reduces network parameters and enhances the extraction of significant features.Then,we establish a knowledge distillation framework utilizing YOLOv8 as the teacher network model. It employs the Channel-wise Knowledge Distillation (CWD) method and an improved loss function to elevate expression recognition accuracy while optimizing real-time performance. The experimental results show that the proposed method performs well on the expression recognition of infants (0-3 years old) in the intelligent care scenario,with the recognition accuracy of 73. 4% . The number of parameters is only 2. 6 MB,making it easily deployable on edge devices. Besides,accuracy reaches 99. 1 % on CK+ data set,and the number of parameters is also small.

常用功能

更新日期/Last Update: 2025-02-10