Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Granular-ball Guided Masking: Structure-aware Data Augmentation

Shuyin Xia, Fan Chen, Dawei Dai, Meng Yang, Junwei Han, Xinbo Gao, Guoyin Wang

TL;DR

This work tackles robustness under data scarcity and distribution shifts by introducing Granular-ball Guided Masking (GBGM), a structure-aware, hierarchical masking strategy guided by Granular-ball Computing. GBGM preserves semantically informative regions while suppressing redundant background through a coarse-to-fine masking process, enabling effective augmentation for CNNs and Vision Transformers. The approach demonstrates consistent improvements in image classification across CIFAR-10/100 and ImageNet-1K, as well as enhanced masked image reconstruction, with linear-time complexity suitable for large-scale training. The results establish GBGM as a principled augmentation paradigm that leverages structure-aware saliency to improve generalization and efficiency in modern vision models.

Abstract

Deep learning models have achieved remarkable success in computer vision, but they still rely heavily on large-scale labeled data and tend to overfit when data are limited or distributions shift. Data augmentation, particularly mask-based information dropping, can enhance robustness by forcing models to explore complementary cues; however, existing approaches often lack structural awareness and may discard essential semantics. We propose Granular-ball Guided Masking (GBGM), a structure-aware augmentation strategy guided by Granular-ball Computing (GBC). GBGM adaptively preserves semantically rich, structurally important regions while suppressing redundant areas through a coarse-to-fine hierarchical masking process, producing augmentations that are both representative and discriminative. Extensive experiments on multiple benchmarks demonstrate consistent improvements in classification accuracy and masked image reconstruction, confirming the effectiveness and broad applicability of the proposed method. Simple and model-agnostic, it integrates seamlessly into CNNs and Vision Transformers and provides a new paradigm for structure-aware data augmentation.

Granular-ball Guided Masking: Structure-aware Data Augmentation

TL;DR

This work tackles robustness under data scarcity and distribution shifts by introducing Granular-ball Guided Masking (GBGM), a structure-aware, hierarchical masking strategy guided by Granular-ball Computing. GBGM preserves semantically informative regions while suppressing redundant background through a coarse-to-fine masking process, enabling effective augmentation for CNNs and Vision Transformers. The approach demonstrates consistent improvements in image classification across CIFAR-10/100 and ImageNet-1K, as well as enhanced masked image reconstruction, with linear-time complexity suitable for large-scale training. The results establish GBGM as a principled augmentation paradigm that leverages structure-aware saliency to improve generalization and efficiency in modern vision models.

Abstract

Deep learning models have achieved remarkable success in computer vision, but they still rely heavily on large-scale labeled data and tend to overfit when data are limited or distributions shift. Data augmentation, particularly mask-based information dropping, can enhance robustness by forcing models to explore complementary cues; however, existing approaches often lack structural awareness and may discard essential semantics. We propose Granular-ball Guided Masking (GBGM), a structure-aware augmentation strategy guided by Granular-ball Computing (GBC). GBGM adaptively preserves semantically rich, structurally important regions while suppressing redundant areas through a coarse-to-fine hierarchical masking process, producing augmentations that are both representative and discriminative. Extensive experiments on multiple benchmarks demonstrate consistent improvements in classification accuracy and masked image reconstruction, confirming the effectiveness and broad applicability of the proposed method. Simple and model-agnostic, it integrates seamlessly into CNNs and Vision Transformers and provides a new paradigm for structure-aware data augmentation.
Paper Structure (15 sections, 13 equations, 5 figures, 8 tables, 1 algorithm)

This paper contains 15 sections, 13 equations, 5 figures, 8 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Stage 1: Coarse Granularity Masking
  5. Stage 2: Finer Granularity Masking
  6. Stage 3: Importance Mask Generation
  7. Time Complexity Analysis
  8. Experiments
  9. Image Classification
  10. CIFAR-10 with ResNet
  11. CIFAR-100 with EfficientNet-L2
  12. ImageNet-1K on CNN and Transformer Backbones
  13. Image Reconstruction
  14. Empirical Complexity Analysis
  15. Conclusion

Figures (5)

  • Figure 1: Comparison of four augmentation strategies: (a) GridMask, (b) Random Erasing, (c) saliency-based HAS, and (d) our GBGM.
  • Figure 2: Overview of the proposed GBGM pipeline. Granular-ball analysis guides a hierarchical selection process to retain informative regions, producing a high-resolution binary importance mask.
  • Figure 3: (a) CIFAR-10 example and (b) CIFAR-100 example. The first row shows the original input image; the second row presents the partitioning result using the granular-ball representation; the third row illustrates the saliency heatmap guided by granular balls; and the fourth row shows the final importance mask for image masking.
  • Figure 4: Visualization of GBGM on ImageNet-1K. From top to bottom: original images, granular-ball partitioning, saliency heatmaps, and final importance masks. The generated masks highlight semantically informative regions and suppress background, helping ViT focus on structurally important areas and improving classification accuracy.
  • Figure 5: Reconstruction comparison on ImageNet-100. Top: original images. Middle: reconstructions by MAE. Bottom: reconstructions by GBGM-enhanced MAE.