| 摘 要: 领域泛化旨在通过从源域学习可迁移的知识,在源域上训练后,再推广到未见过的目标领域也能良好运行。现有的对比领域泛化方法通过跨领域对齐同一类别的样本来追求领域不变特征,但忽略了类别之间的混淆。因此,学习到的模型可能缺乏区分性和鲁棒性,尤其是在领域之间存在显著差异时。为此提出了判别感知对比网络(DACA),通过增强类别和样本的区分性减少类别之间的混淆,同时保持跨领域的类内一致性。DACA在3个基准数据集上比次优的对比方法准确度平均高3.15%、0.22%、1.76%。验证了DACA在缓解类别混淆、提升特征鲁棒性方面的有效性。 |
| 关键词: 领域泛化 类别混淆 类别区分 样本区分 对比学习 |
|
中图分类号:
文献标识码: A
|
| 基金项目: 国家自然科学基金资助项目(52279069) |
|
| Domain Generalization Learning Based on Discrimination-Aware Contrastive Network |
|
JIN Zemin, SONG Kang, WANG Yunyun
|
(School of Computer Science, Nanjing University of Posts and Telecommunications, Nanjing 210023, China)
1022040814@njupt.edu.cn; 1222045613@njupt.edu.cn; wangyunyun@njupt.edu.cn
|
| Abstract: Domain generalization aims to learn transferable knowledge from source domains,so that the model can perform well on unseen target domains. Existing contrastive domain generalization methods pursue domain-invariant features by aligning samples of the same category across domains but overlook the confusion between categories. Consequently, the learned models may lack discriminability and robustness, especially when significant differences exist between domains. To mitigate this issue, a Discrimination-Aware ContrAstive network (DACA) is proposed, which enhances category and sample discrimination to reduce inte-r class confusion while maintaining intra-class consistency across domains. DACA achieves average accuracy improvements of 3.15% , 0.22% , and 1.76% over the second-best contrastive methods on three benchmark datasets, respectively. The results validate the effectiveness of DACA in alleviating category confusion and enhancing feature robustness.Domain generalization aims to learn transferable knowledge from source domains,so that the model can perform well on unseen target domains. Existing contrastive domain generalization methods pursue domain-invariant features by aligning samples of the same category across domains but overlook the confusion between categories. Consequently, the learned models may lack discriminability and robustness, especially when significant differences exist between domains. To mitigate this issue, a Discrimination-Aware ContrAstive network (DACA) is proposed, which enhances category and sample discrimination to reduce inte-r class confusion while maintaining intra-class consistency across domains. DACA achieves average accuracy improvements of 3.15% , 0.22% , and 1.76% over the second-best contrastive methods on three benchmark datasets, respectively. The results validate the effectiveness of DACA in alleviating category confusion and enhancing feature robustness. |
| Keywords: domain generalization category confusion category discrimination sample discrimination contrastive learning |
