Training-Free Few-Shot Image Classification via Diagnostic Verification and Bounded Correction
Year: 2026
Type: preprint
Abstract: Training-free adaptation of CLIP has become a promising paradigm for few-shot image classification because it avoids gradient optimization and overfitting. Existing methods typically construct visual and textual classifiers independently and fuse them with globally fixed weights, implicitly assuming that all samples share the same optimal fusion weight. This assumption fails in few-shot settings: when the top-2 logits are close, fixed weights cannot resolve ambiguity, while unbounded z-score correction may over-correct ambiguous samples. We propose DiVe, a Diagnostic Verification and Bounded Correction framework. For each sample, DiVe verifies its top-2 candidates, constructs a discriminative direction from fine-grained textual prototypes, projects test features and visual prototypes onto this direction to quantify attribution, and converts verification confidence into a bounded correction signal via a tanh mapping. DiVe is fully training-free and can be plugged into existing training-free CLIP classifiers as post-processing. On 11 datasets with 4 CLIP backbones under the 16-shot setting, DiVe achieves the best average accuracy among training-free methods, reaching 76.43% with ResNet-50, and improves GDA-CLIP by 0.92% and TIMO by 0.51% on six representative datasets. Cross-dataset experiments on ImageNet variants demonstrate robustness under distribution shift, while also revealing boundary conditions under extreme domain gaps.
Source: Research Square
Authors: Lei Chen, Li Duan, Rui Fang, HongPing Zhang
Institution: Naval University of Engineering
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7237220231
