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

人脸表情数据集在收集过程中存在主观的标注差异和客观的条件差异,导致表情识别模型在不同数据集间呈现明显的性能差异。为了提高跨数据集表情识别精度、减少表情识别在实际应用中进行样本打标重训练的过程,本文提出了一种基于表情融合特征的域对抗网络模型,用于跨数据集人脸表情识别。采用残差神经网络提取人脸表情的全局特征与局部特征。利用Encoder模块对全局特征与局部特征进行融合,学习更深层次的表情信息。使用细粒度的域鉴别器进行源数据集与目标数据集对抗,对齐数据集的边缘分布和条件分布,使模型能迁移到无标签的目标数据集中。以RAF-DB为源数据集,以CK+、JAFFE、SFEW2.0、FER2013、Expw分别作为目标数据集进行跨数据集人脸表情识别实验。与其他跨数据集人脸表情识别算法相比,所提方法获得了最高的平均识别率。实验结果表明,所提方法能有效提高跨数据集人脸表情识别的性能。

The expression recognition model shows significant performance differences between datasets due to subjective 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 retraining processes for expression recognition in practical applications.Residual neural networks are used to extract the global 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 crossdataset facial expression recognition algorithms,the proposed method achieves the highest average recognition rate.Experimental results show that the proposed method can effectively improve the performance of cross-dataset facial expression recognition.