Reasearch on Cross Domain Few-shot Learning Method Based on Local Feature Association

Few-shot classification aims to recognize unseen classes with few labeled samples from each class. Many meta-learn based and metric based methods elaborately design various task-shared meta-knowledge or embeddings to solve such tasks, and have achieved promising peformance. Howere, when there exists...

Full description

Saved in:
Bibliographic Details
Published in2021 6th International Symposium on Computer and Information Processing Technology (ISCIPT) pp. 754 - 759
Main Authors Ding, Yuan, Wang, Ping
Format Conference Proceeding
LanguageEnglish
Published IEEE 01.06.2021
Subjects
Area measurement
attention mechanism
component
Computers and information processing
cross-domain few-shot
Degradation
Design methodology
Learning systems
local feature
Semantics
Task analysis
Online AccessGet full text

Cover

More Information
Summary:Few-shot classification aims to recognize unseen classes with few labeled samples from each class. Many meta-learn based and metric based methods elaborately design various task-shared meta-knowledge or embeddings to solve such tasks, and have achieved promising peformance. Howere, when there exists domain shift between the test tasks and train tasks, the meta-knowledge and embeddings fails to generalize across domain, which causes the performance degradation. To address the problem, this paper employs an attention module upon a local-descriptor based model called DeepEMD to enable interaction between the local features. The experimental results show that proposed method improves 4.63% in 5-way 1-shot and 0.46% in 5-way 5-shot in the cross-domain setting compared to the origin DeepEMD, which verifies the effectiveness of the method.
DOI:10.1109/ISCIPT53667.2021.00159