•  
  •  
 

Turkish Journal of Electrical Engineering and Computer Sciences

Author ORCID Identifier

LEILA BOUSSAAD: 0000-0001-5175-5153

Abstract

Few-shot image classification benefits from data augmentation, yet most existing methods operate in pixel space with limited control over spectral semantics. We introduce a lightweight, frequency-guided augmentation strategy based on Variational Mode Decomposition (VMD). Our method constructs an offline, per-class ModeBank by decomposing downsampled luminance patches and retaining midband modes that encode class-specific texture patterns. During episodic training, VMD is never executed online: instead, for each support image, a same-class midband mode is selected and blended using PSNR-targeted scaling with a luminance energy cap, ensuring perceptual consistency. The augmentation is fast, reproducible, class-consistent, and integrates seamlessly into standard metric-based pipelines without modifying the backbone or training procedure. We evaluate our approach on multiple benchmarks, including CIFAR-FS, FC100, Tiny-ImageNet-200, and DTD under 5-way 1-shot and 5-shot settings, using Conv-4 and ResNet-12 backbones. Our results show consistent and statistically significant gains over strong ProtoNet baselines, with effect sizes ranging from small to large and most improvements observed in the challenging 1-shot regime. Ablation studies confirm the importance of targeted frequency selection, query-side consistency, energy/PSNR control, and adaptive mixing schedules. Our method demonstrates that structured, class-aligned frequency perturbations offer a principled and practical complement to spatial augmentations in few-shot learning.

DOI

10.55730/1300-0632.4196

Keywords

Few-shot learning, data augmentation, variational mode decomposition, frequency-domain augmentation, modebank

First Page

661

Last Page

681

Publisher

The Scientific and Technological Research Council of Türkiye (TÜBİTAK)

Creative Commons License

Creative Commons Attribution 4.0 International License
This work is licensed under a Creative Commons Attribution 4.0 International License.

Download

Share

COinS