Compact Semantic‐Aware Priors for Large‐Margin Few‐Shot Classification

Few‐shot image classification seeks to recognize novel categories from only a handful of annotated examples, yet current Bayesian‐style methods often depend on coarse priors extracted from base‐class statistics that fail when support samples drift away from class centers. We introduce Compact Semantic‐Aware Priors (CSAP)—a Bayesian framework that leverages lightweight, keyword‐level semantics to build reliable and discriminative priors while explicitly enlarging inter‐class margins. Instead of aligning visual features with high‐dimensional sentence embeddings, CSAP distills a small set of salient word vectors generated directly from class labels, yielding a compact semantic space that robustly estimates cross‐class similarity, selects the most informative base categories, and tightens posterior decision boundaries. Extensive experiments on two real‐world benchmarks show that CSAP surpasses state‐of‐the‐art counterparts by 1.96%–5.29% under both 1‐shot and 5‐shot settings, confirming the efficacy of compact semantic‐aware priors for large‐margin few‐shot classification in the wild. © 2026 Institute of Electrical Engineers of Japan and Wiley Periodicals LLC.