| 李国强,王天雷,龚宁,王俊妍.基于空间注意力和类协方差度量的小样本学习[J].高技术通讯(中文),2022,32(8):801~810 |
| 基于空间注意力和类协方差度量的小样本学习 |
| Few shot learning based on spatial attention and class covariance metrics |
| |
| DOI：10.3772/j.issn.1002-0470.2022.08.003 |
| 中文关键词: 小样本学习； 元度量学习； 泛化能力； 空间注意力； 类协方差度量 |
| 英文关键词: few-shot learning, metric-based meta learning, generalization capability, spatial attention, class covariance metric |
| 基金项目: |
| 作者 | 单位 | | 李国强 | （燕山大学电气工程学院秦皇岛 071000） | | 王天雷 | （燕山大学电气工程学院秦皇岛 071000） | | 龚宁 | （燕山大学电气工程学院秦皇岛 071000） | | 王俊妍 | （燕山大学电气工程学院秦皇岛 071000） |
|
| 摘要点击次数: 874 |
| 全文下载次数: 748 |
| 中文摘要: |
| 近年来，小样本学习逐渐被学术界广泛研究，其旨在使模型在给定样本较少的情况下完成一系列任务。目前基于度量学习的元学习算法被广泛应用于小样本学习中，本文利用度量学习的思想，对基准元度量学习算法原型网络进行改进，提出了注意力类协方差原型网络。首先，为增加模型的泛化能力，提出IBN-Resnet12作为特征提取器；接着，在特征提取模块后加入了空间注意力模块，有效地增强了局部特征；最后提出类协方差度量作为最终的度量分类器，完成了对图像特征维度间相关性的建模。本文在小样本学习经典数据集上进行实验，证明了模型的有效性；同时还进行了大量消融实验，证明了模型改进中各个部分的有效性。 |
| 英文摘要: |
| In recent years, few-shot learning has gradually been widely studied in academia, which aims to make models accomplish a series of tasks with a small number of given samples. At present, metric-based meta learning algorithms are widely used in few-shot learning. The idea of metric learning is used to improve the prototype network of benchmark metric learning algorithm and the attention class covariance prototypical network (ACCPN) is proposed. First, IBN-Resnet12 is proposed as the feature extraction part to increase the generalization ability of the model; then, the spatial attention module is added after the feature extraction module to effectively enhance the local features; finally, the class covariance metric is used as the final metric classifier to complete the modeling of correlation among image feature dimensions. The experiments are conducted on the few-shot learning classical datasets to prove the effectiveness of the model; also, a large number of ablation experiments are conducted in this paper to prove the effectiveness of each part of the model improvement. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
