智能系统学报2023，Vol.18Issue(6)：1205-1212,8.DOI:10.11992/tis.202212030
融合全局与局部特征的跨数据集表情识别方法
Cross-dataset facial expression recognition method fusing global and local features
摘要
Abstract
The expression recognition model shows significant performance differences between datasets due to subject-ive annotation and objective condition differences in the collection of facial expression datasets.A domain adversarial network model based on expression fusion features is proposed for cross-dataset facial expression recognition.This model aims to improve the accuracy of cross-dataset expression recognition and reduce the sample marking and retrain-ing processes for expression recognition in practical applications.Residual neural networks are used to extract the glob-al and local features of facial expressions.An encoder module is then employed to fuse global and local features to learn deep expression information.A fine-grained domain discriminator is adopted to antagonize the source dataset against the target dataset,aligning the edge and conditional distributions of the dataset and facilitating the migration of the model to the unlabeled target dataset.RAF-DB is used as the source dataset,and CK+,JAFFE,SFEW2.0,FER2013,and Expw are used as the target datasets for cross-dataset facial expression recognition experiments.Compared with other cross-dataset facial expression recognition algorithms,the proposed method achieves the highest average recognition rate.Ex-perimental results show that the proposed method can effectively improve the performance of cross-dataset facial ex-pression recognition.关键词
跨数据集/人脸表情识别/领域自适应/特征融合/自注意力机制/迁移学习/细粒度域鉴别器/残差网络Key words
cross-dataset/facial expression recognition/domain adaptation/feature fusion/self-attention mechanism/transfer learning/fine-grained domain discriminator/residual network分类
信息技术与安全科学引用本文复制引用
梁艳,温兴,潘家辉..融合全局与局部特征的跨数据集表情识别方法[J].智能系统学报,2023,18(6):1205-1212,8.基金项目
国家科技创新2030重点项目(2022ZD0208900) （2022ZD0208900）
国家自然科学基金项目(62076103). （62076103）
