S3Mix: Same Category Same Semantics Mixing for Augmenting Fine-grained Images

Data augmentation is a common technique to improve the generalization performance of models for image classification. Although methods such as Mixup and CutMix that mix images randomly are indeed instrumental in general classification, swapping or masking regions not friendly fine-grained since key classification precisely lies discriminative informative regions, it unreasonable generate labels solely consistent with proportion synthesis. Some erasing like Cutout even endanger because by chance. In this paper, we propose Same Category Semantics Mixing method (S3Mix) corresponding characteristics images. Specifically, limit mixture same category semantics. The core two constraints. exchange semantic region ensures discrimination semantics integrity generated image, class avoids problem label generation. At time, homology loss promote relationship between positive pairs. Experiments have been conducted on four datasets results show proposed superior traditional well some data methods.