基于域特定批量归一化的对抗域适应图像分类

Domain Specific Batch Normalization Based on Adversarial Domain Adaptation Image Classification

摘要: 无监督域适应(UDA)旨在将知识从带有大量标签的源域迁移到没有标签的目标域。目前的研究主要集中在统一两个域的特征分布上。然而，目标域通常具有更为复杂的背景信息，源域和目标域的全局特征分布并不相同，在源域和目标域之间直接共享整个网络强制全局分布对齐会导致性能的下降。针对此问题，提出了一种新的基于域特定批量归一化的对抗域适应模型。首先，采用对抗性学习损失模块，综合考虑域对齐和类别对齐，从对抗学习获得的混淆矩阵中自动构建一个新的损失函数来矫正自训练中的伪标签；其次，在卷积神经网络(CNN)的编码器架构中引入域特定批量归一化模块(DSBN)，通过分离批量归一化层来分别适应源域和目标域。将域特定信息与域不变信息分离，更好地学习域不变特征表示，来获得更好的泛化性能。最后，本文的方法在Office-Home数据集和Office-31数据集的准确率分别达到67.4%和89.4%，验证了模型的有效性。

Abstract: Unsupervised domain adaptation (UDA) aims to transfer knowledge from a source domain with many labels to a target domain without labels. Current research mainly focuses on unifying the feature distributions of the two domains. However, the target domain usually has more complex background information, the global feature distributions of the source and target domains are different, and directly sharing the entire network between the source and target domains to enforce global distribution alignment will lead to performance degradation. In response to this issue, a novel adversarial domain adaptation model is proposed based on domain-specific batch normalization. First, using the adversarial learning loss module, considering domain alignment and class alignment, a new loss function is automatically constructed from the confusion matrix obtained by adversarial learning to correct the pseudo-labels in self-training; second, a domain-specific batch normalization module (DSBN) is introduced in the encoder architecture of a convolutional neural network (CNN), which adapts to the source and target domains separately by separating the batch normalization layers. Separate domain-specific information from domain-invariant information and learn domain-invariant feature representations to achieve better generalization performance. Finally, the accuracy of the method in this paper in the Office-Home dataset and Office-31 dataset reached 67.4% and 89.4%, respectively, which verified the model’s effectiveness.