FreqGRL: Suppressing Low-Frequency Bias and Mining High-Frequency Knowledge for Cross-Domain Few-Shot Learning

TL;DR

This work tackles Cross-domain Few-Shot Learning (CD-FSL) under limited target supervision by introducing a frequency-space perspective. It presents FreqGRL, a framework comprising Low-Frequency Replacement (LFR), High-Frequency Enhancement (HFE), and Global Frequency Filter (GFF) that operate in the frequency domain to suppress source-specific low-frequency biases and promote high-frequency, domain-generalizable cues. Through extensive experiments on five CD-FSL benchmarks, FreqGRL achieves state-of-the-art performance in both 1-shot and 5-shot settings, including surpassing 80% average accuracy in the 5-shot regime and setting new baselines for several target domains. The approach provides a practical, frequency-guided strategy for robust cross-domain generalization with scarce target data, with potential extensions to broader frequency-domain representations and tasks.

Abstract

Cross-domain few-shot learning (CD-FSL) aims to recognize novel classes with only a few labeled examples under significant domain shifts. While recent approaches leverage a limited amount of labeled target-domain data to improve performance, the severe imbalance between abundant source data and scarce target data remains a critical challenge for effective representation learning. We present the first frequency-space perspective to analyze this issue and identify two key challenges: (1) models are easily biased toward source-specific knowledge encoded in the low-frequency components of source data, and (2) the sparsity of target data hinders the learning of high-frequency, domain-generalizable features. To address these challenges, we propose \textbf{FreqGRL}, a novel CD-FSL framework that mitigates the impact of data imbalance in the frequency space. Specifically, we introduce a Low-Frequency Replacement (LFR) module that substitutes the low-frequency components of source tasks with those from the target domain to create new source tasks that better align with target characteristics, thus reducing source-specific biases and promoting generalizable representation learning. We further design a High-Frequency Enhancement (HFE) module that filters out low-frequency components and performs learning directly on high-frequency features in the frequency space to improve cross-domain generalization. Additionally, a Global Frequency Filter (GFF) is incorporated to suppress noisy or irrelevant frequencies and emphasize informative ones, mitigating overfitting risks under limited target supervision. Extensive experiments on five standard CD-FSL benchmarks demonstrate that our frequency-guided framework achieves state-of-the-art performance.

Figures (7)

• Figure 1: Illustration of image reconstruction using low-frequency and high-frequency components on both the source dataset (miniImagenet) and target dataset (CUB-2011-200). Original images are first transformed into frequency domain via the Fast Fourier Transformation (FFT), then decomposed into their respective low-frequency and high-frequency components. Subsequently, images containing only low- or high-frequency information are reconstructed by inversely mapping these individual components back into the original spatial domain.
• Figure 2: Accuracy ratios of full-spectrum-trained models evaluated on tasks reconstructed from specific frequency components. Each bar represents the ratio of the model's accuracy on frequency-specific tasks (tasks reconstructed using only low-frequency or high-frequency components) relative to its accuracy on the original tasks. Results are reported for four target datasets using a ResNet-10 backbone, with models trained on a combination of source-domain data and limited labeled target-domain data.
• Figure 3: Overview of the proposed FreqGRL framework, which consists of three key modules: Low-Frequency Replacement (LFR), High-Frequency Enhancement (HFE), and Global Frequency Filter (GFF).
• Figure 4: Illustration of the LFR module. The source and target tasks are transformed into frequency space and decomposed into the low- and high-frequency components. Then, the source high-frequency and target low-frequency are fused and remapped back to the spatial space to reconstruct the novel source task.
• Figure 5: Ablation study of the HFE module. (a) Accuracy under different frequency ranges. (b) Accuracy with different input types.