基于改进VGG16网络的失能老人表情识别研究  

Research on Expression Recognition of Disabled Elderly PeopleBased on Improved VGG16 Network

作　　者：何巍[1] 李苏 HE Wei;LI Su(School of Physics and Electronic Engineering,Sichuan Normal University,Chengdu 610101,Sichuan)

机构地区：[1]四川师范大学物理与电子工程学院,四川成都610101

出　　处：《四川师范大学学报(自然科学版)》2025年第3期383-391,共9页Journal of Sichuan Normal University(Natural Science)

基　　金：国家自然科学基金青年基金(62301348)。

摘　　要：为能更好地关注失能老人的情绪状态,采用VGG16作为表情识别的基础模型,并在此基础上进行优化改进.首先,在特征层上将激活函数改用SiLU函数,并添加批归一化层;然后,在分类层上采用自适应平均池化处理图像,同时利用卷积层实现全连接效果,避免参数过多和过拟合问题;最后,通过SENet通道注意力机制迭代式地融合相同通道数的卷积层,实现浅层与深层特征的交互,丰富人脸表情特征提取.实验结果表明,在FER2013和CK+数据集上的识别准确率分别达到72.50%和98.70%,与基础方法对比分别提高8.20%和3.90%,实验表明改进的方法能够提高表情识别率,具有一定的先进性.In order to better monitor the emotional state of elderly individuals with disabilities, this article employs VGG16 as the foundational model for emotion recognition and makes improvements upon it. Firstly, the activation function is replaced with the SiLU function and batch normalization layers are added at the feature-extraction level. Secondly, adaptive average pooling is utilized in the classification layer to process images, while convolutional layers are used to achieve fully connected effects, thereby avoiding issues related to excessive parameters and overfitting. Lastly, through the attention mechanism of the SENet channel, convolutional layers with the same number of channels are iteratively fused to enable interaction between shallow and deep features, enriching the feature extraction of the facial expression. The experimental results indicate that the recognition accuracy on the FER2013 and CK+ datasets reached 72.50% and 98.70%, respectively, which represents an improvement of 8.20% and 3.90% compared to the baseline method. These findings demonstrate that the improved method can enhance emotion-recognition rates and possesses certain advancements.

关 键 词：VGG16 表情识别 自适应平均化 通道注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]