中圖分類號： TP37
文獻(xiàn)標(biāo)識碼： A
DOI： 10.19358/j.issn.2096-5133.2021.11.005
引用格式： 梁廣. 基于歐氏距離解纏的多角度跨庫人臉表情識別[J].信息技術(shù)與網(wǎng)絡(luò)安全，2021,，40(11)：29-36.

Euclidean distance disentanglement adaptive method for multi-view cross-database facial expression recognition

Abstract： At present, the accuracy of face expression recognition is affected by database biases and posture variations. To solve this problem, this paper proposes a novel cross-database unsupervised learning model based on feature disentanglement. There are three strategies in this model：multiple attribute features disentanglement, contrastive learning within face expression feature, adversarial domain adaptation. The feature disentanglement can get expression-related, pose-related and database-related features, and try to separate the features of different attributes. Contrastive learning can make expression features of the same label closer, and expression features of the different labels as far apart as possible. Adversarial domain adaptive learning makes the unseen sample′s features closer to sample′s features of the training database. Our experiments are carried out on 8 databases and achieved state of the art results.

Key words : facial expression recognition；feature disentanglement,；cross database,；contrastive learning