Fine-Grained Image Classification

Fine-grained image classification studies how to tell apart categories that share an overall appearance but differ in subtle details, such as bird species, vehicle models, or disease grades. Approaches include pretrained CNNs and vision transformers, part- and attention-based features, metric and contrastive learning, and methods that use class hierarchies or vision-language models; few-shot, semi-supervised, and active learning address limited labels. Benchmarks span natural species, vehicles, food, and medical images, with evaluation increasingly considering hierarchy, calibration, open-set recognition, and deployment costs alongside accuracy. A central challenge is ensuring models rely on diagnostic visual evidence and remain reliable under shifts in lighting, background, domain, or available evidence.