基于深度可分离卷积的人脸表情识别  

Facial expression recognition based on depthwise separable convolution

作　　者：李春虹 卢宇[1] LI Chun-hong;LU Yu(College of Computer Science,Sichuan Normal University,Chengdu 610101,China)

机构地区：[1]四川师范大学计算机科学学院,四川成都610101

出　　处：《计算机工程与设计》2021年第5期1448-1454,共7页Computer Engineering and Design

摘　　要：为解决在复杂环境下人脸表情识别模型泛化能力不足、识别精度及效率不高的问题,提出一种基于深度可分离卷积的人脸表情识别方法。利用人脸分割网络分割出人脸图像中与表情识别最相关的感兴趣区域,减少非重要因素影响;利用深度可分离卷积构建两个基分类器,实现模型轻量化;采用联合微调方法融合基分类器,提升模型识别率。实验结果表明,在FER-2013数据集上的识别率达到75.15%,较Kaggle表情识别挑战赛冠军提高了3.95%;在CK+和JAFFE数据集上的识别率高达98.98%、97.14%,验证了其有效性。To solve this problem of insufficient generalization ability and low recognition efficiency of facial expression recognition(FER)model in complex environment,a FER method based on depthwise separable convolution was proposed.The face segmentation network was used to segment the area of interest in the face image that was most relevant to expression recognition,reducing the influence of non-important factors.Two base classifiers were constructed using depthwise separable convolution to realize model lightweight.A joint fine-tuning method was used to fuse the base classifier to improve the model recognition rate.Experimental results show that the recognition rate on the FER-2013 dataset reaches 75.15%,which is 3.95% higher than the Kaggle expression recognition challenge champion,the recognition rates of CK+ and JAFFE datasets are as high as 98.98%,97.14%,which verifies the effectiveness of the method.

关 键 词：人脸表情识别 复杂环境 感兴趣区域 深度可分离卷积 联合微调 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]