Abstract
Few-shot image classification (FSIC) addresses image recognition when only a few labeled examples are available for each novel class. This survey reviews five major FSIC method families: meta-learning, transfer learning, data augmentation, attribute-based methods, and vision-language model adaptation. The $N$-way $K$-shot setting is defined, and standard training, theoretical foundations, benchmark datasets, and evaluation protocols are examined. Results for 5-way 1-shot and 5-shot classification are compared across four widely used benchmarks. The survey also examines self-supervised and semi-supervised learning, computational efficiency, and few-shot class-incremental learning. The analysis shows that representation quality strongly affects FSIC performance. Large pre-trained and vision-language models achieve high benchmark accuracy, but their performance is associated with substantially larger pre-training datasets. Direct comparisons across methods remain difficult because backbone capacity, pre-training data, and evaluation settings vary. Cross-domain generalization and practical deployment also remain challenging. Connections with PAC-Bayes theory, information theory, and causal representation learning are discussed. Finally, open research challenges are identified in robustness, long-tail recognition, fine-grained classification, multimodal learning, efficient adaptation, and standardized evaluation.