A comprehensive review for Generalized Few-Shot Image Classification

In the era of deep learning, Few-Shot Learning (FSL) is a promising direction to bridge the gap between application of Artificial Intelligence (AI) and human intelligence. The specific implemented AI application in the image classification field of few-shot image classification has achieved tremendous developments in the past decade, thereby gradually evolving to be the Generalized Few-Shot image Classification (GFSC) with several groups. This paper comprehensively reviews recent advances in GFSC and highlights possible future work. Firstly, we formulate the formal definition for each group of GFSC, and elaborate on the difference between GFSC and related machine learning tasks to avoid conceptual confusion. Based on the formulation, we provide a comprehensive review of GFSC in vanilla few-shot classification, semi-supervised few-shot classification, unsupervised few-sot classifications, cross-domain few-shot classification, few-shot open-set recognition, and robust few-shot classification, covering various aspects of each sub-group ranging from algorithm comparison, detailed analysis to heuristic remarks. Next, performance comparisons of the representative reviewed methods on benchmark datasets are provided and discussed. Finally, we conclude this survey by highlighting several promising directions that should be further addressed in the future.