Spectral Decomposition and Transformation for Cross-domain Few-shot Learning

Area of Science:

• Machine Learning
• Computer Vision
• Signal Processing

Background:

• Cross-domain few-shot learning (CDFSL) faces challenges due to domain gaps between source and target datasets.
• Limited data in target domains hinders model generalization and fine-tuning.
• Existing methods struggle to bridge the source-target domain gap effectively.

Purpose of the Study:

• To mitigate the domain gap in CDFSL by analyzing frequency-domain characteristics.
• To enhance the generalization capability of models by expanding source domain spectral coverage.
• To propose a novel data augmentation technique for improved CDFSL performance.

Main Methods:

• Frequency-domain analysis to understand and quantify the domain gap.
• Spectral Decomposition and Transformation (SDT) for data augmentation.
• A two-stream network architecture to process original and augmented data.

Main Results:

• The proposed SDT method effectively expands source domain spectra composition.
• Experimental results demonstrate state-of-the-art performance on CDFSL benchmarks.
• The approach successfully reduces the negative impact of domain gaps on model generalization.

Conclusions:

• Frequency-domain analysis offers valuable insights into CDFSL domain gaps.
• SDT is a promising data augmentation strategy for enhancing CDFSL.
• The proposed method significantly improves model performance in cross-domain transfer learning scenarios.