A common strategy for data augmentation is to apply style transfer on training images to enrich appearance variations. Recent approaches have combined style augmentation with random erasing by selectively replacing image subregions with style-transferred patches. However, existing methods rely on uniformly random region selection or fixed geometric constraints, which may fail to cover informative texture patterns and limit augmentation diversity. To address these limitations, we propose a texture-aware subregion style augmentation framework that prioritizes regions with strong texture responses for style replacement. Furthermore, we extend this framework with a hybrid region strategy that combines a texture-driven region with an auxiliary random region, implicitly enabling more diverse style perturbations. Experiments demonstrate that our approach achieves higher classification accuracy and improves robustness compared with existing style-based augmentation methods.

翻译：暂无翻译