陈朋磊,王江涛,张志伟,何 程.基于特征聚合与多元协同特征交互的航拍图像小目标检测[J].电子测量与仪器学报,2023,37(10):183-192 |
基于特征聚合与多元协同特征交互的航拍图像小目标检测 |
Small object detection in aerial images based on feature aggregation andmultiple cooperative features interaction |
|
DOI: |
中文关键词: 小目标检测 航拍图像 小目标特征聚合网络 多元协同特征交互模块 |
英文关键词:small object detection aerial images small target feature aggregation network multiple collaborative feature
interaction module |
基金项目:国家自然科学基金( 61976101)、安徽省高校自然科学研究重点项目( 2023AH050319)、安徽省高校优秀科研创新团队项目(2023AH010044)资助 |
|
|
摘要点击次数: 764 |
全文下载次数: 590 |
中文摘要: |
针对无人机航拍图像目标尺寸太小、包含的特征信息较少,导致现有的检测算法对小目标检测效果不理想的问题,提出
一种基于特征聚合与多元协同特征交互的无人机航拍图像小目标检测算法。 首先,针对主干网对特征提取不足的问题,采用
Swin Transformer 作为 RetinaNet 主干网络, 以增强算法对全局信息的提取能力。 其次,为提高网络对远处目标即小目标的检测
能力,设计出一种高效的小目标特征聚合网络(SFANet),实现对浅层特征图小目标细节信息的充分整合。 最后,为进一步提高
网络对多尺度目标的检测性能,使低层特征信息流向高层,提出全新的多元协同特征交互模板(MCFIM)。 在公开无人机航拍
数据集 VisDrone2019-DET 上的实验结果表明,所提算法相较于原 RetinaNet 基线网络检测精度提高 7. 6%,对于小目标具有更好
的检测效果。 |
英文摘要: |
Aiming at the problem that the target size of the UAV aerial image is too small and contains less feature information, which
leads to the unsatisfactory detection effect of the existing detection algorithm on small objects, a UAV aerial photography based on feature
aggregation and multi-collaborative feature interaction is proposed. First of all, in view of the insufficient feature extraction of the
backbone network, Swin Transformer is selected as the RetinaNet backbone network to enhance the global information extraction ability of
the algorithm. Secondly, in order to improve the detection ability of remote targets, a small target feature aggregation network is
proposed, which can fully integrate the details of small targets in shallow feature maps. Finally, in order to further improve the detection
performance of multi-scale targets, a new multiple collaborative feature interaction module is proposed to make the low-level feature
information flow to the high-level. Experimental results on VisDrone2019-DET, a public UAV aerial photo data set, show that compared
with the original RetinaNet baseline network detection precision increased by 7. 6%, the proposed algorithm has better detection effect for
small targets. |
查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|