ADLDA: A Method to Reduce the Harm of Data Distribution Shift in Data Augmentation

TL;DR

This work tackles distribution shift introduced by data augmentation by proposing ADLDA, a plug‑in method that partitions augmented data into domain-labeled subdomains and trains a domain classifier to regularize the feature extractor. The training objective combines the standard task loss with a domain loss, yielding $L = L_Y(\hat{Y},Y) - \lambda \sum_i a_i L_D(\mathbb{D}_i, \hat{P_d})$, and ADLDA can be mounted on diverse architectures with negligible inference impact. Empirically, ADLDA improves classification accuracy across multiple datasets (Caltech256, Tiny ImageNet, CIFAR‑10/100) and architectures (ViT, EfficientNet, DenseNet, ResNet, VGG), with more pronounced gains on architectures featuring complex feature extractors, and enhancements in feature localization as shown by GradCAM heatmaps. This suggests that domain‑aware augmentation regularization yields robust improvements in object recognition and potentially segmentation, marking a practical approach to increasing robustness against augmentation‑induced distribution shifts.

Abstract

This study introduces a novel data augmentation technique, ADLDA, aimed at mitigating the negative impact of data distribution shifts caused by the data augmentation process in computer vision task. ADLDA partitions augmented data into distinct subdomains and incorporates domain labels, combined with domain adaptation techniques, to optimize data representation in the model's feature space. Experimental results demonstrate that ADLDA significantly enhances model performance across multiple datasets, particularly in neural network architectures with complex feature extraction layers. Furthermore, ADLDA improves the model's ability to locate and recognize key features, showcasing potential in object recognition and image segmentation tasks. This paper's contribution provides an effective data augmentation regularization method for the field of computer vision aiding in the enhancement of robustness and accuracy in deep learning models.

Figures (4)

• Figure 1: Using different data augmentation methods and dividing the domains
• Figure 2: Shift in image features occurs after data augmentation
• Figure 3: Neural network architecture
• Figure 4: CAM of the trained models under different DARates