结合改进卷积神经网络与自编码器的表情识别  

Expression Recognition Based on Improved Convolutional Neural Networks and Self-encoder

作　　者：魏赟[1] 李栋[1] WEI Yun;LI Dong(School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China)

机构地区：[1]上海理工大学光电信息与计算机工程学院,上海200093

出　　处：《小型微型计算机系统》2022年第2期387-392,共6页Journal of Chinese Computer Systems

基　　金：外高桥项目上海市科学技术委员会科研计划项目(19511105103)资助。

摘　　要：为了克服单一神经网络模型提取表情特征困难,以及堆叠深层网络结构会造成训练过程复杂、参数冗余等问题,本文提出了一种引入注意力机制的轻量级CNN通道和卷积自编码器预训练通道的双通道模型.在轻量级CNN通道中以具有残差思想的深度可分离卷积结构进行深层次特征提取并且减少了模型参数量,还引入了通道域注意力机制使得该通道能够学习到更有用的特征;同时使用卷积自编码器对输入人脸表情图像进行无监督预处理,使得模型提取的特征更加多样化.实验结果表明,在FER2013和CK+表情数据集上分别取得了72.70%和97.50%的识别率.通过与相关方法对比,表明了本文模型在保证较少参数量的同时也具有较高的识别率.In order to overcome the difficulty of extracting expression features from single neural network,and the complexity of training process and redundancy of parameters caused by stacking deep structure,a dual channel model of lightweight CNN and convolutional self-encoder pretraining model is proposed.In the lightweight CNN channel,the structure with residual block and depth separable convolution is used for feature extraction,and the amount of model parameters is reduced.The channel of attention mechanism is used to enable the channel to learn more useful features;at the same time,the convolution self-encoder is used for unsupervised preprocessing of the facial expression image input,which makes the features extracted from the model more diversified.The experimental results show that the recognition rate is 72.70%and 97.50%respectively on fer 2013 and CK+expression data sets.Compared with the related methods,the proposed model has a higher recognition rate while ensuring less parameters.

关 键 词：表情识别 深度可分离卷积 卷积自编码器 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]