Class-relevant patch embedding selection for few-shot image classification

doi:10.1016/j.imavis.2026.106031

Class-relevant patch embedding selection for few-shot image classification

2026-05-22

0

PRE

AI

DOI：10.1016/j.imavis.2026.106031

Original

Original request for help

Save

Abstract

En 中文

• We address background interference in few-shot learning without localization and alignment. • We present CPES, selecting class-relevant patch embeddings by evaluating their similarity to class embeddings. • We explore various methods for patch embedding selection, and adopt top-ranked class-relevant patch embeddings. • Our approach enhances existing patch embedding metrics and boosts their performance. • Visualizations and extensive experiments showcase the superiority of our method.

Keywords:

few-shot learning

patch embedding selection

class-relevant embeddings

image classification

background interference

Class-relevant patch embedding selection for few-shot image classification

Abstract

Journal

Organization