基于因果关系和特征对齐的图像分类域泛化模型

Domain generalization model in image classification based on causality and feature alignment

针对现有域泛化方法性能较差或缺乏理论可解释性的缺点,提出了一种基于因果关系和特征对齐的图像分类域泛化模型,并证明了该模型的可识别性。该模型利用域泛化中的因果关系来学习含有不同信息的特征,将域泛化问题转化为特征相关分布的偏移,再利用特征对齐消除偏移。为提高模型的性能,采用对抗训练进一步优化学到的特征。在公共数据集上的实验结果表明,新提出的模型与目前最优的方法性能相当,表明该模型具有理论可解释性的同时,还有不俗的实际性能表现。

Aiming at the shortcomings of poor performance or lack of theoretical interpretability of existing domain generalization methods,a novel domain generalization model for image classification based on causality and feature alignment is proposed,and the model′s identifiability is demonstrated.The model uses causal relationships in domain generalization to learn features containing different information,transforms the domain generalization problem into the biases of feature-related distributions,and then eliminates the biases by aligning features.To improve the performance of the model,adversarial training is used to further optimize the learned features.Experimental results on public datasets show that the newly proposed model performs comparably with the state-of-the-art models,indicating that the model has theoretical interpretability as well as decent practical performance.