부산대학교 도서관

Highlights •We propose a feature extractor with attention mechanism to obtain the feature representations from the source and target domains respectively.•A domain adaptation module based on feature dictionaries is proposed to preserve class-specific feature structures for the source domain and target domain, respectively.•A pre-training and fine-tuning structure based on Efficientnet-B0 is employed to enable stable classification.•Visualized representations illustrate the effectiveness of the proposed domain adaptation modules in reducing the difference in feature representations the two domains.•Ablation and comparative experimental results demonstrate the effectiveness of the proposed method for obtaining emotionally discriminative and domain-invariant representations in cross-corpus SER.