Evolution of Few-Shot Image Classification: A Survey from Metric Learning to Large Model Fine-Tuning

Abstract

While deep learning has achieved remarkable success in image recognition, it relies heavily on massive amounts of labeled data. In real-world scenarios like medical imaging, collecting such large datasets is often expensive or impossible. To overcome this critical bottleneck, Few-Shot Image Classification (FSIC) has emerged as a vital domain, enabling models to learn new categories from extremely limited samples. Based on a comprehensive review of recent literature, this paper systematically divides mainstream FSIC methods into three categories: Metric Learning, Optimization/Meta-Learning, and Transfer Learning with Large Model Fine-Tuning. This survey comprehensively evaluates the underlying mechanisms, inherent strengths, and specific weaknesses of each category. Additionally, the paper provides a clear performance comparison to illustrate the evolutionary impact of these different approaches. Finally, the paper deeply discusses remaining technical challenges, specifically detailing issues like cross-domain generalization and computational dependence, and outlines promising future research directions, such as multimodal knowledge fusion, to further advance the field.