KeepOriginalAugment: Single Image-based Better Information-Preserving Data Augmentation Approach

TL;DR

KeepOriginalAugment addresses the trade-off between data diversity and information fidelity in image augmentation by detecting salient regions and intelligently integrating them into non-salient areas. It enables augmentation on either region and explores three placement strategies and three augmentation-part strategies, guided by a saliency map with a threshold $\tau$ and region importance $I(R,x,y) = \sum_{(i,j)\in R} s_{(i,j)}(x,y)$. Empirical results on CIFAR-10, CIFAR-100, and TinyImageNet across multiple architectures show systematic gains over SalfMix, KeepAugment, and even HybridMix, with notable improvements such as ~2 percentage-point error-rate reduction on CIFAR-100 with PreActResNet-101. The method achieves higher accuracy and lower error rates while maintaining feature fidelity, suggesting practical impact for improving generalization in computer vision models, with potential for debiasing through future dataset-diversity analyses.

Abstract

Advanced image data augmentation techniques play a pivotal role in enhancing the training of models for diverse computer vision tasks. Notably, SalfMix and KeepAugment have emerged as popular strategies, showcasing their efficacy in boosting model performance. However, SalfMix reliance on duplicating salient features poses a risk of overfitting, potentially compromising the model's generalization capabilities. Conversely, KeepAugment, which selectively preserves salient regions and augments non-salient ones, introduces a domain shift that hinders the exchange of crucial contextual information, impeding overall model understanding. In response to these challenges, we introduce KeepOriginalAugment, a novel data augmentation approach. This method intelligently incorporates the most salient region within the non-salient area, allowing augmentation to be applied to either region. Striking a balance between data diversity and information preservation, KeepOriginalAugment enables models to leverage both diverse salient and non-salient regions, leading to enhanced performance. We explore three strategies for determining the placement of the salient region minimum, maximum, or random and investigate swapping perspective strategies to decide which part (salient or non-salient) undergoes augmentation. Our experimental evaluations, conducted on classification datasets such as CIFAR-10, CIFAR-100, and TinyImageNet, demonstrate the superior performance of KeepOriginalAugment compared to existing state-of-the-art techniques.

Figures (6)

• Figure 1: Overall KeepOriginalAugment process
• Figure 2: Comparison of the relevant data augmentation methods with ours
• Figure 3: Are these augmented examples helpful for training?
• Figure 4: Are these augmented examples helpful for model generalization?
• Figure 5: Strategies: where to place the salient region?