基于改进卷积神经网络的面部表情识别  

Facial expression recognition based on improved convolutional neural network

作　　者：王亮亮 薛明明 王晓玲 Wang Liangliang;Xue Mingming;Wang Xiaoling(China Construction Eighth Engineering Division Corp.Ltd.,Shanghai 200131,China)

机构地区：[1]中国建筑第八工程局有限公司,上海201131

出　　处：《计算机时代》2025年第3期16-20,共5页Computer Era

摘　　要：为提高基于图像特征点的面部表情识别算法精度,解决深度卷积神经网络体积过大、参数过多而导致运行缓慢等问题,提出了一种基于残差学习的卷积神经网络算法。该算法将深度可分离卷积和对特征提取进行解耦,使用全局平均池化代替全连接层,从而进一步减少参数,并将改进轻量化模型最终架构权重存储在855 kB的文件中。最后,对改进的网络模型在FER2013数据集中进行评估。测试结果表明,改进轻量化模型的识别准确率为68%,有效解决了原有算法匹配精度低的问题。In order to solve the problems of slow operation due to the excessive size of deep convolutional neural networks and too many parameters,this paper proposes a convolutional neural network model based on residual learning.This model decouples deep separable convolution from feature extraction and uses global average pooling instead of a fully connected layer to further reduce the parameters.The final architecture weights of the improved lightweight model are stored in an 855 kB file.The improved network model is also evaluated on the FER2013 dataset.According to the training results,the improved lightweight model shows good recognition performance,and its recognition accuracy is 68%.

关 键 词：表情识别 卷积神经网络 轻量化模型 残差学习 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]