基于面部多个局部特征的人脸表情识别算法  

Facial expression recognition technology based on multiple local facial features

作　　者：贾茜伟 闫娟 杨慧斌 刘向前 JIA Qianwei;YAN Juan;YANG Huibin;LIU Xiangqian(School of Mechanical and Automotive Engineering,Shanghai University of Engineering Science,Shanghai 201620,China)

机构地区：[1]上海工程技术大学机械与汽车工程学院,上海201620

出　　处：《智能计算机与应用》2022年第10期144-149,共6页Intelligent Computer and Applications

摘　　要：人脸面部表情通常来说可以显露出人的内心活动变化,目前现有的表情识别方法一般依靠面部的整体特征进行处理,没有考虑面部的一些局部特征,导致面部表情识别的准确度不理想。人的面部表情进行变化时,面部局部肌肉会随之变化,基于此,提出一种基于面部多个局部特征的人脸表情识别算法。本文首先对检测到的人脸进行面部分区,分为23个子区域,再将分好的区域输入到卷积神经网络中进行局部特征的提取。最终使用AM-softmax函数将表情分为中性、愤怒、厌恶、惊讶、高兴、悲伤和恐惧七类。评估实验在CK+和JAFFE数据集上对本文算法进行验证,得到的平均准确率分别是99.87%和96.72%,均超过S-Patches算法,该结果表明本文算法对表情识别性能有所提高。Facial expressions can usually reveal changes in human inner activities. Current facial expression recognition methods generally rely on the overall features of the face for processing, and do not consider some local features of the face, resulting in a decrease in the accuracy of facial expression recognition. When a person’s facial expression changes, local muscles will change accordingly. Based on this, a facial expression recognition algorithm based on multiple local features of the face is proposed. This paper first partitions the detected face into 23 sub-regions, and then inputs the divided regions into the convolutional neural network to extract local features. Finally, the AM-softmax function is used to divide the expressions into seven categories: neutral, angry, disgusted, surprised, happy, sad and fearful. The evaluation experiment verifies the average accuracy of the algorithm in this paper on the CK+ and JAFFE data sets, are 99.87% and 96.72%, respectively, which both exceed the S-Patches algorithm. The results show that the algorithm has improved the performance of facial expression recognition.

关 键 词：表情识别 面部分区 卷积神经网络 AM-softmax函数 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]