Abstract:
Mixup Augmentations mitigate overfitting in visual models by generating mixed samples, with the core process consisting of two stages: sample selection and sample mixing. Existing methods typically randomly select pairs of images for interpolation or patch replacement to generate mixed samples, neglecting the relationship between feature space distribution and semantic information, which limits the enhancement effect. To address this issue, a density-guided mixing for image data augmentation method is proposed to leverage feature distribution to guide the augmentation process. In sample selection, the method introduces a density metric to quantify the distribution density of samples and proposes a density difference-based to select image pairs with highly representative features and maximum saliency. In sample mixing, the method incorporates the density difference ofimage pairs and jointly optimizes the mixed sample generation task with the classification task in an end-to-end manner, automatically generating mixing masks to ensure that discriminative semantic regions are preserved in the newly mixed images. Experiments on both standard and fine-grained benchmark datasets demonstrate that the proposed method improves classification accuracy by approximately 1% compared to AutoMix. Furthermore, the proposed image pairing algorithm exhibits strong compatibility and can further enhance the performance of other augmentation strategies.