基于多重并联图神经网络的小样本图像分类算法
Multi-level parallel graph neural network based few-shot image classification algorithm
吴婕 1张海翔1
作者信息
- 1. 浙江理工大学计算机科学与技术学院305室,浙江 杭州 310018
- 折叠
摘要
小样本学习(few-shot learning,FSL)中,由于样本量过少,导致了特征多样性的降低.为了弥补特征多样性的降低,提出通过提高模型的特征提取能力,以获得更为充分的特征数量.利用多重并联图神经网络来进行多重特征提取,使模型更充分地提取图像特征,从而提升小样本图像分类任务的分类准确率.所提出的多重特征提取方法在5-way 1-shot设置下将基线的分类准确率提高了2.02%,在5-way 5-shot设置下将基线的分类准确率提高了1.98%.
Abstract
FSL leads to the reduction of feature diversity due to the small sample size.To compensate for this,it is proposed to obtain a more sufficient number of features by improving the feature extraction ability of the model.Multiple parallel graph neural networks are utilized for multiple feature extraction,so that the model can extract image features more fully and improve the classification accuracy of few-shot image classification tasks.The proposed multiple feature extraction method improves the classification accuracy of the baseline by 2.02%in 5-way 1-shot setting and by 1.98%in 5-way 5-shot setting.
关键词
小样本学习/小样本图像分类/多重特征提取/图神经网络Key words
FSL(few-shot learning)/few-shot image classification/multiple feature extraction/graph neural network引用本文复制引用
基金项目
国家自然科学基金(61672466)
国家自然科学基金(62011530130)
浙江省自然科学基金联合基金(LSZ19F010001)
浙江省重点研发计划项目(2020C03060)
出版年
2023
NETL
NSTL
维普
万方数据