Few-shot learning based on deep learning: A survey.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning (DL) excels with large datasets but faces limitations in real-world scenarios with scarce data.
• Limited data restricts model performance and generalization capabilities in many practical applications.
• Few-shot learning (FSL) emerges as a solution to train high-performance models using minimal samples.

Purpose of the Study:

• To provide a comprehensive review of DL-based few-shot learning methods for image classification.
• To categorize and introduce classic and advanced FSL techniques.
• To discuss datasets, performance benchmarks, challenges, and future directions in FSL.

Main Methods:

• The review categorizes FSL methods into four main groups: data augmentation, metric learning, meta-learning, and approaches involving auxiliary tasks.
• It systematically introduces established and state-of-the-art FSL algorithms within these categories.
• Performance evaluation on common FSL datasets is presented.

Main Results:

• The study categorizes FSL methods into data augmentation, metric learning, meta-learning, and auxiliary tasks.
• It reviews classic and advanced FSL techniques and their performance on benchmark datasets.
• Key challenges and future research avenues are identified.

Conclusions:

• Few-shot learning is crucial for image classification when large datasets are unavailable.
• The review categorizes FSL methods, discusses their performance, and highlights future research directions.
• FSL techniques offer a promising solution for leveraging deep learning with limited data.