类别语义信息对齐的无监督领域自适应网络

Category Semantic Information Alignment Network for Unsupervised Domain Adaptation

针对在目标应用场景中缺乏大量有标定训练数据的情况下难以获得有效的深度学习分类模型的问题,结合领域分布差异的方法与对抗学习方法的优势,提出以显式特征对齐与隐式领域对抗及类别对齐为基础的领域自适应框架.对于显式特征对齐模块,考虑到领域知识差异大带来的优化难题,采用渐进式协同优化策略,通过逐层减小不同语义层之间的领域差异,提升领域自适应性能.对于隐式类别对齐模块,为了增强目标特征的判别性,使用自训练方法获得伪标签,克服伪标签存在的标签噪声问题,并通过学习混淆矩阵优化伪标签的准确率,自动构造新的目标领域损失函数,从而在减小领域间差异的同时,提升源领域与目标领域相同类别的特征分布对齐的准确性.基于Office-31数据集的6个跨领域分类任务与基于Office-Home数据集的12组跨领域分类任务的实验结果表明,该方法在迁移学习任务上的平均分类准确率相较于基准方法分别提升11.9%和19.9%,所提出网络对于领域自适应任务是有效的.

To address the problem that it is difficult to obtain effective deep learning classification models in the absence of a large amount of consistent distribution and labeled data in the target application scenario,a domain adaptation framework was proposed based on explicit feature alignment with domain adversarial and implicit category alignment.The advantages of the domain distribution difference approach and the adversarial learning approach were combined in this framework.For the explicit feature alignment module,considering the optimization challenges caused by large differences in domain knowledge,an incremental collaborative optimization strategy was used.The goal of this strategy is to improve the domain adaptation performance by reducing the domain differences between different semantic layers.For the implicit category alignment module,to enhance the discriminability of the target features,a self-training method was adopted to obtain pseudo-labels and overcome the problem of label noises.Besides,the accuracy of the pseudo-labels was optimized by learning the confusion matrix,thus a target domain loss function was automatically constructed.This function is able to reduce the differences between domains,while improving the accurate alignment of the feature distribution of the same categories in both the source and target domains.Experiments based on 6 cross-domain classification tasks on the Office-31 dataset and 12 sets of cross-domain classification tasks on the Office-Home dataset were conducted.Results show that the average classification accuracy of the proposed method on the transfer learning tasks improves by 11.9%and 19.9%,respectively,compared with the benchmark method,demonstrating the effectiveness of the network in domain adaptation tasks.