SaliencyMix plus : Noise-Minimized Image Mixing Method With Saliency Map in Data Augmentation

초록

Data augmentation is vital in deep learning for enhancing model robustness by artificially expanding training datasets. However, advanced methods like CutMix blend images and assign labels based on pixel ratios, often introducing label noise by neglecting the significance of blended regions, and SaliencyMix applies uniform patch generation across a batch, resulting in suboptimal augmentation. This paper introduces SaliencyMix+, a novel data augmentation technique that enhances the performance of deep-learning models using saliency maps for image mixing and label generation. It identifies critical patch coordinates in batch images and refines label generation based on target object proportions, reducing label noise. Experiments on CIFAR-100 and Oxford-IIIT Pet datasets show that SaliencyMix+ consistently outperforms CutMix and SaliencyMix, achieving the lowest Top-1 errors of 24.95% and 34.89%, and Top-5 errors of 7.00% and 12.13% on CIFAR-100 and Oxford-IIIT Pet, respectively. These findings highlight the effectiveness of SaliencyMix+ in boosting model accuracy and robustness across different models and datasets.

키워드