应对复杂光照下的高精度表情识别方法  

High precision expression recognition method for complex illumination

作　　者：李嘉乾 张雷 LI Jia-qian;ZHANG Lei(Institute of Mechanical Engineering,Jiangsu University of Technology,Changzhou 213001,China)

机构地区：[1]江苏理工学院机械工程学院,江苏常州213001

出　　处：《计算机工程与设计》2023年第8期2489-2497,共9页Computer Engineering and Design

基　　金：国家自然科学基金项目(61701202、61901196)。

摘　　要：针对表情识别存在相似表情识别精度不高和不同光照下识别困难的问题,提出一种改进的双通道残差网络表情识别模型。通过改进局部二值化算子,改善复杂光照下难以提取到鲁棒特征的问题,通过改进注意力机制,改善全局特征提取能力;搭建特征融合网络,通过交叉实验获取对于不同数据集都鲁棒的特征融合系数;将中心损失引入设计联合算法提高相似表情之间的区分度。实验结果表明,该算法提升了相似表情的区分精度,对于光照具有更好的鲁棒性。模型在3个公开数据集上的准确率达98.53%、96.42%、94.24%。To solve the problems that the recognition accuracy of similar expressions is low-precision and the recognition is dif-ficult under the complex illumination,an improved dual-channel residual network was proposed.By improving the local binary patterns,the problem that it is difficult to extract robust feature under complex illumination was improved.By improving attention mechanism,the ability of global feature extraction was improved.By building feature fusion network and cross-cutting experiments,robust feature fusion coefficients in different data sets were extracted.The central loss was introduced into the joint algorithm to improve the discrimination of similar expressions.According to the experiment,the proposed method improves the discrimination accuracy of similar facial expression and is more robust to illumination.The accuracies of the model are 98.53%,96.42%and 94.24%on the three public data sets.

关 键 词：人脸表情识别 复杂光照 改进的局部二值化算子 改进的注意力机制 双通道残差网络 特征融合 中心损失 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]