Transfer Learning for Few-Shot Image Classification

Abstract

Image classification technology has numerous vital applications across various industries and research fields. However, in many practical cases, it is often challenging to obtain sufficiently high-quality labelled data to train reliable models. As a classic and widely used paradigm of few-shot learning, transfer learning effectively mitigates the serious problem of over-fitting caused by limited data samples. It significantly enhances model generalisation ability by transferring knowledge learned from large-scale source domains (e.g., ImageNet) to few-shot target domain tasks. This paper systematically and comprehensively investigates transfer learning-based few-shot image classification methods, categorising them into three main types: feature-representation-based transfer, parameter-based transfer, and relational-knowledge-based transfer. It has also been clearly found that transfer learning methods, such as Adversarial Discriminative Domain Adaptation (ADDA) and knowledge distillation, serve as practical and efficient tools for addressing few-shot image classification problems. However, several critical challenges, such as negative transfer, multimodal heterogeneous transfer, and domain shift, still urgently require careful consideration and effective resolution in the future.