电子测量与仪器学报

董宝鑫,王江涛.基于重参数化广义金字塔与扩张残差的遥感图像旋转框算法[J].电子测量与仪器学报,2024,38(12):54-61

基于重参数化广义金字塔与扩张残差的遥感图像旋转框算法

Remote sensing image rotation box algorithm leveraging reparameterizedgeneralized pyramid and dilated residual

DOI：

中文关键词: 深度学习遥感图像旋转框目标检测重参数广义金字塔 DRS-YOLO

英文关键词:deep learning remote sensing images rotated bounding box object detection reparameterized generalized pyramid DRS-YOLO

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

作者	单位
董宝鑫	淮北师范大学物理与电子信息学院淮北235000
王江涛	1.淮北师范大学物理与电子信息学院淮北235000；2.智能计算及应用安徽省重点实验室淮北235000

Author	Institution
Dong Baoxin	School of Physics and Electronic Information, Huaibei Normal University, Huaibei 235000, China
Wang Jiangtao	1.School of Physics and Electronic Information, Huaibei Normal University, Huaibei 235000, China; 2.Ahui Province Key Laboratory of Intelligent Computing and Applications, Huaibei 235000,China

摘要点击次数: 278

全文下载次数: 277

中文摘要:

由于遥感图像中目标数量多而密集,且背景信息复杂,导致现有检测算法对于小目标检测精度不够理想,针对该问题,提出了一种基于重参数化广义金字塔与扩张残差的遥感图像小目标旋转框检测算法DRS-YOLO。首先,为克服主干网络对特征提取不足的缺点,以旋转算法YOLOv8OBB为基础,在颈部网络引入扩张式残差模块,以增强遥感目标语义信息。其次,为提高网络对于多尺度目标的检测性能,使底层特征信息流向高层,引入重参数化泛化特征金字塔网络替换颈部网络结构,更高效的融合多尺度特征,易于捕捉高层语义和低层空间细节。最后,为进一步提高网络对于小目标的检测性能,基于SPPF提出SPPFI对目标感受野进行扩展,提升了对遥感目标的检测精度。在公开的DIOR数据集和HRSC2016数据集上相较于原YOLOv8sOBB基线网络的检测精度分别提升了1.5%和9.8%。实验表明改进后的算法显著增强了对遥感图像小目标的检测性能。

英文摘要:

Due to the dense presence of multiple targets and complex background information in remote sensing images, existing detection algorithms often struggle to achieve satisfactory precision in small target detection. To address this issue, we propose a small target rotation box detection algorithm for remote sensing images called DRS-YOLO, based on reparameterized generalized pyramids and dilated residuals. First, to overcome the shortcomings of the backbone network in feature extraction, we enhance semantic information by incorporating a dilated residual module into the neck of the network, building upon the YOLOv8OBB framework. Second, to improve the network's performance in detecting multi-scale targets and facilitate the flow of low-level feature information to high-level features, we replace the neck structure with a reparameterized generalized feature pyramid network for more efficient multi-scale feature fusion, which aids in capturing high-level semantics and low-level spatial details. Finally, to further enhance the network's performance in detecting small targets, we propose the SPPFI module to expand the receptive field, thereby improving detection accuracy for remote sensing targets. Experimental results demonstrate that the improved algorithm achieves an increase in detection precision of 1.5% and 9.8% on the public DIOR and HRSC2016 datasets, respectively, compared to the original YOLOv8sOBB baseline network, indicating a significant enhancement in small target detection performance for remote sensing images.

查看全文查看/发表评论下载PDF阅读器