基于区域NSBP特征的加权证据融合表情识别  

Weighted evidence fusion expression recognition based on regional NSBP features

作　　者：刘娟 胡敏 黄忠 Liu Juan;Hu Min;Huang Zhong(School of Electronic Engineering and Intelligent Manufacturing,Anqing Normal University,Anqing 246133,China;Anhui Province Key Laboratory of Affective Computing and Advanced Intelligent Machine,School of Computer and Information,Hefei University of Technology,Hefei 230601,China)

机构地区：[1]安庆师范大学电子工程与智能制造学院,安庆246133 [2]合肥工业大学计算机与信息学院情感计算与先进智能机器安徽省重点实验室,合肥230601

出　　处：《电子测量与仪器学报》2020年第11期132-139,共8页Journal of Electronic Measurement and Instrumentation

基　　金：国家自然科学基金(61672202,61702012);安徽省自然科学基金(1908085MF195);安徽省重点实验室开放课题项目(ACAIM180203);中央高校基本科研业务费专项资金(PA2020GDSK0061);安徽省高校协同创新项目(GXXT-2019-030)资助

摘　　要：为了提取鲁棒性强的人脸纹理特征并提高区域特征决策融合的性能,提出一种基于邻近平滑二值模式(neighbor smooth binary pattern,NSBP)特征描述子和加权证据融合(weighted evidence fusion,WEF)的表情识别新方法。首先,提出了一种NSBP描述子,通过判定水平、垂直及对角线方向上的"中心"像素点灰度值是否在各梯度方向上两邻域的灰度值范围内来对图像进行编码;然后基于提取的眉毛、眼睛和嘴巴区域的NSBP纹理特征来构造证据的初始基本概率分配(basic probability assignment,BPA);最后针对登普斯特-谢弗(Dempster-Shafer,D-S)证据理论在证据之间存在冲突时进行融合的不足,提出一种加权证据修正的合成方法,以完成3个区域证据的决策融合。实验结果表明,该方法在CK(Cohn-Kanade)数据库上的平均表情识别率和识别时间分别为95.25%、765 ms,与其他相关方法的比较也验证了其有效性。In order to extract robust facial features and improve the decision-level fusion of multi-regional features,a new expression recognition method based on neighbor smooth binary pattern(NSBP)feature descriptor and weighted evidence fusion(WEF)is proposed.First,a NSBP descriptor is proposed to encode the image by determining whether the gray values of the center pixels in the horizontal,vertical and diagonal directions are within the gray value range of two neighborhoods in each gradient.Then the initial basic probability assignments(BPA)of evidences are constructed based on the extracted NSBP texture features of the eyebrows,eyes,and mouth regions.Finally,aiming at the deficiency of Dempster-Shafer(D-S)evidence theory in conflict evidence fusion,a synthetic method of weighted evidence revision is proposed to realize the decision fusion of three regional evidences.Experimental results show that the recognition rate of this method on the Cohn-Kanade(CK)database is 95.25%,and the average recognition time is 765 ms,compared with other related methods,the effectiveness of this method is also verified.

关 键 词：表情识别 NSBP特征 加权证据融合 DEMPSTER组合规则 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]