File(s) under embargo
Style Augmentation and Domain-Aware Parametric Contrastive Learning for Domain Generalization
conference contributionposted on 2023-10-26, 04:38 authored by M Li, J Zhang, W Zhang, L Gong, Zili ZhangZili Zhang
The distribution shift between training data and test data degrades the performance of deep neural networks (DNNs), and domain generalization (DG) alleviates this problem by extracting domain-invariant features explicitly or implicitly. With limited source domains for training, existing approaches often generate samples of new domains. However, most of these approaches confront the issue of losing class-discriminative information. To this end, we propose a novel domain generalization framework containing style augmentation and Domain-aware Parametric Contrastive Learning (DPCL). Specifically, features are first decomposed into high-frequency and low-frequency components, which contain shape and style information, respectively. Since the shape cues contain class information, the high-frequency components remain unchanged. Then Exact Feature Distribution Mixing (EFDMix) is used for diversifying the low-frequency components, which fully uses each order statistic of the features. Finally, both components are re-merged to generate new features. Additionally, DPCL is proposed, based on supervised contrastive learning, to enhance domain invariance by ignoring negative samples from different domains and introducing a set of parameterized class-learnable centers. The effectiveness of the proposed style augmentation method and DPCL is confirmed by experiments. On the PACS dataset, our method improves the state-of-art average accuracy by 1.74% using ResNet-50 backbone and even achieves excellent performance in the single-source DG task.