基于二阶表征的条件对抗域适应网络

域对抗学习是一种主流的域适应方法,它通过分类器和域判别器来学习具有可区分性的域不变特征;然而,现有的域对抗方法大多利用一阶特征来学习域不变特征,忽略了具有更强表达能力的二阶特征.提出了一种条件对抗域适应网络,通过联合建模图像的二阶表征以及特征和分类器预测之间的互协方差以便更有效地学习具有区分性的域不变特征;此外,引入了熵条件来平衡分类器预测的不确定性,以保证特征的可迁移性.提出的方法在两个常用的域适应数据库Office-31和ImageCLEF-DA上进行了验证,实验结果表明该方法优于同类方法并获得了领先的性能.

Conditional adversarial domain adaptation networks based on second-order representation

Domain adversarial learning is a mainstream approach of domain adaptation, which learns discriminative domain-invariant feature representation through classifier and domain discriminator. However, existing adversarial domain adaptation methods often use first-order features to learn domain-invariant feature representation, ignoring the second-order features with more powerful representative ability. This paper proposed conditional adversarial domain adaptation networks based on second-order representation, which modeled the second-order moments of features and the cross-covariance between features and classifier predictions for more effectively learning discriminative domain-invariant features. Moreover, it introduced entropy conditioning to guarantee the transferability. The proposed method was evaluated on two commonly used datasets Office-31 and ImageCLEF-DA. Experiments show that the proposed method outperforms its counterpart.