基于支持向量机的动态人脸图像微表情识别  

Dynamic Facial Micro-Expression Recognition Based on Support Vector Machines

作　　者：李昆仑[1] 彭小华 熊婷[1] LI Kun-lun;PENG Xiao-hua;XIONG Ting(College of Science and Technology,Nanchang University,Gongqingcheng Jiangxi 332020,China)

机构地区：[1]南昌大学科学技术学院,江西共青城332020

出　　处：《计算机仿真》2023年第12期331-336,共6页Computer Simulation

基　　金：江西省教育厅科学技术研究项目(GJJ191555);江西省教育厅科学技术研究项目(GJJ151510);江西省高等学校教学改革研究项目(JXJG-22-30-7)。

摘　　要：微表情是反映人内心情感的真实写照,但其变化波动小且留存率低,在模式辨识上具有较大的局限性。为解决受头部姿态变化与环境背景干扰而导致微表情识别准确率低的问题,提出人体骨架分割算法与面部微表情矫正算法。首先采用窗口滑动串联法提取图像轮廓信息,并利用Open Pose算法规划人体骨骼图;然后构建面部微表情分割规则,即以“鼻”骨骼点为原点构建实轴,以“鼻-颈”与实轴角度差,旋转、矫正并分割面部微表情图。最终采用均匀算子提取面部微表情图像的LBP-TOP纹理特征,构建人脸动态微表情识别模型(PAR-SVM模型)。仿真结果表明,经上述算法分割矫正处理后的微表情图像构建的PAR-SVM微表情识别模型较传统SVM微表情识别模型,在SMIC-HS与CASME2数据集上,其准确率分别提高了0.580%与0.575%;与文献[7]相比,F1指标提升了5.6%,召回率提升了8.3%。新构建的PAR-SVM微表情识别模型在解决头部姿态变化与环境背景干扰问题的同时,有效的提高了微表情识别准确率。Micro-expression is a true portrayal of people's inner feelings,but its fluctuation is small and its reten⁃tion rate is small,so it has great limitations in pattern recognition.In order to solve the problem of low accuracy of micro-expression recognition caused by the change of head posture and the interference of environmental background,this paper proposes a human skeleton segmentation algorithm and a facial micro-expression correction algorithm.Firstly,the contour information of the image is extracted by using the window sliding serial method,and the human skeleton map is planned by using the Open Pose algorithm.Then,the facial micro-expression segmentation rule is constructed,that is,the real axis is constructed by taking the"nose"skeleton point as the origin,and the facial mi⁃cro-expression map is rotated,corrected and segmented by taking the angle difference between the"nose-neck"and the real axis.Finally,the uniform operator is used to extract the LBP-TOP texture features of facial micro-expression images,and a dynamic facial micro-expression recognition model(PAR-SVM model)is constructed.Simulation ex⁃periments show that the accuracy of the PAR-SVM micro-expression recognition model constructed by the micro-ex⁃pression images segmented and corrected by the proposed algorithm is improved by 0.580%,and the accuracy is in⁃creased by 0.575%on the CASME2 data set;compared with the literature,the F1 index is improved by 5.6%,and the recall is improved by 8.3%.The PAR-SVM micro-expression recognition model in this paper not only solves the problem of head pose change and environmental background interference,but also effectively improves the accuracy of micro-expression recognition.

关 键 词：支持向量机 表情矫正 微表情识别 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]