基于Swin Transformer的无监督域自适应图像分类

Unsupervised Domain Adaptation Image Classification Based on Swin Transformer

摘要: 大多数当前的无监督域自适应(UDA)技术从域级别或类级别学习域不变的特征表示。基于域级别的主流方法是对抗学习，对抗学习通常不考虑目标数据的固有判别信息。基于类别级别的UDA方法通常是为目标域样本生成伪标签，由于这些伪标签通常噪声太大，这不可避免地会影响UDA性能；其次，现有方法没有明确地强制区分不同类别的特征。为了解决以上问题，我们提出了基于Swin Transformer的无监督域自适应(SwinUDA)。首先，对于域对齐，将Swin Transformer与对抗性自适应相结合，提高模型对噪声输入的鲁棒性，其次，对于类别对齐，使用正交投影损失(OPL)直接在特征空间中实施约束。此外，正交投影损失对标签噪声干扰的影响更有鲁棒性。最后，引入了互信息最大化损失(IML)来保留目标域的可区分特征。本文提出的SwinUDA模型可以同时学习可迁移和可区分的特征。在Office-Home、Office-31和VisDA-2017三个公开数据集上进行实验，SwinUDA都展现了最佳的性能。

Abstract: Most current unsupervised domain adaptation (UDA) techniques learn domain invariant feature representations from the domain-level or class-level. Adversarial learning is the dominating strat-egy based on the domain-level. It tries to align the global feature distributions of the two domains without considering the target data’s innate discriminative information. Class-level-based ap-proaches typically generate pseudo-labels for data in the target domain. These pseudo-labels im-pact UDA’s performance because they are generally overly noisy. In addition, existing methods do not explicitly enforce a good separation of different classes of features. To solve the above problems, we propose the Unsupervised Domain Adaptation Using Swin Transformer (SwinUDA). First, for domain alignment, the Swin Transformer is combined with adversarial adaptation to improve the robustness of the model to noisy inputs. The experimental results show that using the transformer as a feature extractor has higher transferability. Second, constraints are directly enforced in the feature space for class alignment using Orthogonal Projection Loss (OPL). Samples from the same class (whether from the source or target domain) are pulled closer, while samples from different classes are pushed away. In addition, the orthogonal projection loss is more robust to the influence of label noise interference. To preserve the discriminative information of the target domain, a mu-tual information maximization loss (IML) is introduced to protect the discriminating features of the target domain. The SwinUDA model proposed in this paper can simultaneously learn transferable and differentiable features. Experiments were performed on the three public datasets Office-Home, Office-31, and VisDA-2017. SwinUDA showed the best performance.