The single domain generalization(SDG) based on meta-learning has emerged as an effective technique for solving the domain-shift problem. However, the inconsistent semantic information between source and augmented domains and difficult separation of domain-invariant features from domain-related features make SDG model hard to achieve great generalization. To address the above problems, this paper proposed a novel meta-learning method based on domain enhancement and feature alignment (MetaDefa) to improve the model generalization performance. This method utilizes background replacement and visual damage techniques to generate diverse and effective augmented images for each image, ensuring the consistency of semantic information between the source domain and the enhanced domains. The multi-channel feature alignment module fully mines image information by focusing on similar target regions between the source and enhanced domains feature spaces and suppressing feature representations of non-target areas, thereby effectively extracting sufficient transferable knowledge. Through experimental evaluation, MetaDefa achieved 88.87%, 73.06% and 57.06% accuracy on office-Caltech-10, office31 and PACS datasets, respectively. The results show that the MetaDefa method successfully achieves semantic consistency between the source and augmented images and adequate extraction of domain-invariant features, which significantly improves the generalization performance of single domain generalization models.

[2024-04-08] Accepted Paper /doc/2023.11.0585/v1