Knowledge driven weights estimation for large-scale few-shot image recognition

doi:10.1016/j.patcog.2023.109668

Knowledge driven weights estimation for large-scale few-shot image recognition

2023-10-01

4

PRE

AI

DOI：10.1016/j.patcog.2023.109668

Original

Original request for help

Save

Abstract

En 中文

We study the topic of large-scale few-shot image recognition with semantic-visual relational knowledge -based transfer learning. Compared with classical few-shot learning, which is defined as a k-way ( k de-notes the number of categories, usually 5/10-way) classification problem, large-scale few-shot recognition contains more categories (100-way +) with few samples per category and is easier to overfit. A promising direction of large-scale few-shot learning is transferring prior relevant semantic/visual knowledge from outside data to accelerate the convergence on limited positives. Inspired by this, we propose a novel Knowledge Driven Weights Estimation framework. Specifically, the framework leverages semantic and vi-sual relations between new few-shot and existed many-shot categories to transfer knowledge trained on many-shot datasets (e.g., ImageNet-10 0 0). We show that the transferred knowledge provides a good ini-tialization for novel few-shot categories leading to faster convergence speed and higher performance than random/imprinting initialization. Experimental results on additional un-seen ImageNet categories (other than the 10 0 0 categories) with few positives show that our method is effective on large-scale few-shot recognition. (c) 2023 Elsevier Ltd. All rights reserved.

Keywords:

Few-shot image

Recognition

Knowledge transfer

Knowledge driven weights estimation for large-scale few-shot image recognition

Abstract

Journal

Organization