摘要
针对实际应用中交叉数据集无法通过监督学习对预先训练模型进行微调的问题,提出一种基于条件生成对抗网络的跨域面部表情识别框架。该框架分为特征嵌入、对抗性学习和分类三个模块。利用联合学习的嵌入式功能来弥合源和目标数据分布之间的差距,完成从源域到目标域的特征转移;利用无监督生成对抗网络进行优化,根据域自适应方法给出表情分类。实验结果表明,与其他域自适应方法相比,该方法在面部表情识别方面具有极大优势;相对于无自适应的跨域方法,该方法的面部表情识别率有了明显提高。
Aiming at the problem that cross-data sets cannot fine-tune the pre-training model through supervised learning in practical application,this paper proposes a cross-domain facial expression recognition framework based on conditional generative adversarial network.It is divided into three modules:feature embedding,adversarial learning and classification.This algorithm used the embedded function of joint learning to bridge the gap between the distribution of source and target data,and completed the feature transfer from the source domain to target domain.Then,the unsupervised generative adversarial network was used for optimization.Finally,the expression classification was given according to the domain adaptive method.The experimental results show that compared with other domain adaptive methods,our method has great advantages in facial expression recognition.Compared with the non-adaptive cross-domain method,the facial expression recognition rate of our method has been significantly improved.
作者
戴蓉
Dai Rong(Civil Aviation Flight University of China,Guanghan 618307,Sichuan,China)
出处
《计算机应用与软件》
北大核心
2020年第8期166-170,232,共6页
Computer Applications and Software
基金
国防科技重点实验室开放项目(HTL-O-19K01)。
关键词
面部表情识别
条件生成对抗网络
域自适应
跨域数据集
Facial expression recognition
Conditional generative adversarial network
Domain adaptation
Cross-domain datasets