结合改进RepVGG-A0网络和重新标签的人脸表情识别研究  

Facial expression recognition combined with improved RepVGG-A0network and relabeling

作　　者：李婉婷 罗晓曙[1] 蒙志明 陈吉 LI Wanting;LUO Xiaoshu;MENG Zhiming;CHEN Ji(College of Electronic Engineering,Guangxi Normal University,Guilin 541000,China;College of Innovation and Entrepreneurship,Guangxi Normal University,Guilin 541000,China)

机构地区：[1]广西师范大学电子工程学院,广西桂林541000 [2]广西师范大学创新创业学院,广西桂林541000

出　　处：《现代电子技术》2022年第20期69-74,共6页Modern Electronics Technique

基　　金：广西科技重大专项(桂科AA18118004);广西人文社会科学发展研究中心科学研究工程·创新创业专项(重大委托项目)(ZDCXCY01)。

摘　　要：目前大规模人脸表情识别的主要问题在于不确定性,这些不确定性来源于模棱两可的面部表情、低质量的脸部图像和标注者的主观性。为此,文中提出一种基于RepVGG-A0改进后的网络模型。该模型引入有效通道注意力机制,即在卷积层和ReLU激活函数之间插入ECA通道注意力模块,在特征提取后引入加权模块来预计样本的权重,对于不确定的样本给予的权重较小,并采用重新标签的方法对低权重的人脸表情图片重新给予伪标签,目的是使修改过的样本在下一次训练中获得高权重,从而提高人脸表情识别率。最后,在RAF-DB和FER-2013数据集上进行实验验证。结果表明,文中改进模型的人脸表情识别率分别达到88.90%和75.61%,说明该方法对人脸表情识别具有有效性。The existing main challenge of large-scale facial expression recognition is the uncertainty,which derives from the ambiguous facial expression,low-quality facial image and the subjectivity of the tagger. On this basis,an improved network model based on RepVGG-A0 is proposed. In this model,the effective channel attention mechanism is introduced,that is,the ECA channel attention module is inserted between the ReLU activation function and the convolution layer,and the weighting module is introduced after feature extraction to predict the sample weight. The weight given to the uncertain samples is small,and the method of relabeling is used to give a pseudo label to the low weight facial expression pictures,so as to make the modified samples obtain high weight in the next training, and then improve the facial expression recognition rate. The experimental verification was conducted on RAF-DB and FER-2013 datasets. The results show that the facial expression recognition rate of the model improved in this paper can reach 88.90% and 75.61% respectively,which proves that the method is effective for facial expression recognition.

关 键 词：人脸表情识别 RepVGG-A0 重新标签 网络模型 特征提取 数据预处理 ReLU 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TP391[电子电信—信息与通信工程]