Enhancing Image Classification via PPSEAUG: A Plug-and-Play Segmentation-Guided Augmentation Framework

Image classification performance is heavily reliant on the quality and diversity of training data. Traditional methods often fall short due to the high cost and labour intensity of obtaining high-quality segmentation annotations. This paper introduces PPSEAUG, a plug-and-play segmentation-guided augmentation framework that leverages lightweight segmentation models to enhance classification accuracy. By combining uncertainty estimation and information gain optimization, PPSEAUG effectively mitigates the impact of low-quality samples. Evaluations on Mini-ImageNet and Caltech-101 datasets demonstrate that PPSEAUG consistently outperforms traditional augmentation methods, improving Top-1 accuracy by up to 10.25\% on ResNet-50 and by 5.25\% on RepViT on the Caltech-101 dataset. These results confirm the practicality and effectiveness of PPSEAUG in enhancing image classification models,Our code is available at https://github.com/SEGAUG/PPSEAUG.