«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

期数:

2025年04期

页码:

29-36

栏目:

媒体计算

出版日期:

2025-04-10

文章信息/Info

Title:

Few-shot Object Detection Method Based on Feature Coupling Attention Mechanism

文章编号:

1673-629X(2025)04-0029-08

作者:

易炟1; 2; 陈东方1; 2; 王晓峰1; 2

1. 武汉科技大学 计算机科学与技术学院,湖北 武汉 430065;

2. 智能信息处理与实时工业系统湖北省重点实验室,湖北 武汉 430065

Author(s):

YI Da1; 2; CHEN Dong-fang1; 2; WANG Xiao-feng1; 2

1. School of Computer Science and Technology,Wuhan University of Science and Technology,Wuhan 430065,China;

2. Hubei Provincial Key Laboratory of Intelligent Information Processing and Real-time Industrial Systems,Wuhan 430065,China

关键词:

小样本学习; 目标检测; 元学习; 特征耦合; 特征融合; 任务解耦

Keywords:

few-shot learning; object detection; meta-learning; feature coupling; feature fusion; task decoupling

分类号:

TP391.4

DOI:

10.20165/j.cnki.ISSN1673-629X.2024.0383

摘要:

在当前基于元学习的小样本目标检测方法中,查询图像和支持图像的特征提取过程往往在进行最终特征融合之前是独立进行的,缺乏有效的信息交互,这导致查询特征的代表性信息不足,尤其在样本极度有限的情况下更为显著。 为解决这一问题,提出了一种基于特征耦合注意力机制的小样本目标检测方法(FC-FSOD)。 FC-FSOD 以 Faster-RCNN 作为基础网络架构,并在此基础上进行了创新。 首先,设计了特征增强耦合模块,该模块增强了支持特征,使其成为更具代表性的支持原型;其次,通过注意力机制将原型与查询特征进行耦合,赋予查询特征以特定于支持特征的感知信息;再次,综合考虑查询图像的 ROI 特征与类级原型之间的差异性和相似性,设计了一种更为鲁棒的特征融合策略;最后,对分类和回归任务进行了解耦处理,消除了全局平均池化对回归预测的负面影响。 在 PASCAL VOC 和 MS COCO 这两个公开数据集上的实验结果表明,该方法在多种小样本场景中的检测精度均有明显提升。

Abstract:

In the current meta-learning-based few-shot object detection methods,the feature extraction processes of query image and support images are often performed independently before the final feature fusion,and there is a lack of effective information interaction,which leads to insufficient representative information of the query features,and it is more significant especially in the case of extremely limited samples. To solve this problem,a few-shot object detection method based on the feature coupling attention mechanism ( FC-FSOD) is proposed. FC-FSOD takes Faster-RCNN as the basic network architecture and makes innovations based on it. Firstly,a feature enhanced coupling module is designed,which enhances the support features to become more representative support prototypes.Secondly,the prototype is coupled with the query feature through the attention mechanism,and the query feature is endowed with the per-ceptual information specific to the supporting feature. Thirdly, a more robust feature fusion strategy is designed by considering the differences and similarities between the ROI features of the query image and the class-level prototypes. Finally,decoupling of the classi-fication and regression tasks removes the negative influence of global average pooling on regression prediction. Experimental results on two publicly available datasets,PASCAL VOC and MS COCO,show a significant improvement in the detection accuracy of the proposed method in a variety of few-shot scenarios.

常用功能

更新日期/Last Update: 2025-04-10