Few-Shot Learning With Enhancements to Data Augmentation and Feature Extraction.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Few-shot image classification aims to identify novel classes using minimal labeled data.
• Current methods often struggle with limited data and feature degradation during extraction.
• Improving data utilization and feature extraction is crucial for robust few-shot learning.

Purpose of the Study:

• To propose novel methods for enhancing few-shot image classification performance.
• To address the limitations of data scarcity and feature degradation in existing models.
• To improve the model's ability to accumulate knowledge effectively from limited training data.

Main Methods:

• Introduced self-mixup (SM), a data augmentation technique that assembles augmented instances of the same image.
• Developed calibration-adaptive downsampling (CADS) to calibrate and utilize feature characteristics, preventing degradation.
• Integrated SM and CADS to improve both data utilization and feature extraction.

Main Results:

• The proposed methods demonstrated superior performance on four widely adopted few-shot classification datasets.
• Self-mixup (SM) effectively enhances knowledge accumulation from limited training data.
• Calibration-adaptive downsampling (CADS) facilitates robust feature extraction, benefiting classification accuracy.

Conclusions:

• The combined approach of improved data utilization and feature extraction significantly advances few-shot image classification.
• The proposed self-mixup and calibration-adaptive downsampling methods offer a promising direction for future research in low-data regimes.
• This work contributes to developing more robust and efficient models for scenarios with limited labeled data.