7 lines
1.2 KiB
TeX
7 lines
1.2 KiB
TeX
\section{Conclusion}
|
|
\label{sec:conclusion}
|
|
|
|
We presented DZGCoOp, a zero-shot guided dual-branch prompt learning framework for few-shot image classification. DZGCoOp distills LLM-enriched semantic anchors, zero-shot visual features, probability distributions, and an instance-level confidence from the frozen CLIP; learns an asymmetric strong-weak dual-branch network coupled through shared deep prompt layers; fuses the branches per instance with the zero-shot confidence; and regularizes optimization with feature- and distribution-level guidance losses plus an epoch-weighted averaging strategy. Experiments on base-to-novel, cross-domain, and cross-dataset transfer demonstrate consistent improvements over strong prompt learning baselines, and ablations quantify the contribution of each component. We believe that exploiting a frozen VLM's own predictive confidence to route between conservative and exploratory adaptation paths is a general recipe for knowledge-preserving few-shot learning, and we leave its extension to denser vision-language tasks as future work.
|
|
|
|
\noindent\textbf{Acknowledgments.} This work was supported by the National Natural Science Foundation of China (No. 62306320) and the Natural Science Foundation of Jiangsu Province (No. BK20231063).
|