全文下载:
202602019.pdf
文章编号:1672-6987(2026)02-0131-08;DOI:10. 16351/j. 1672-6987. 2026. 02. 019
吴沭运,彭天文,刘国柱,梁宏涛* ,崔文静 (青岛科技大学 信息科学技术学院,山东 青岛 266061)
摘 要:由于遥感图像具有背景复杂、噪声严重的特点,因此遥感目标的检测是一项具有挑战 性的任务。为提高遥感目标的检测精度,提出一种基于双向采样特征金字塔和具有注意力机 制的解耦头的遥感图像目标检测算法。双向采样特征金字塔旨在将深层次的特征图与浅层 网络特征图聚合,以获取更为丰富的空间信息和语义信息,此外,通过新增路径聚合通道的方 式,增强金字塔网络的特征融合能力。具有注意力的解耦头通过将目标类别、坐标与置信度 预测解耦,减少不同层次特征的冲突,并且在检测支路中分别引入改进的空间注意力和改进 的通道注意力,进一步增大网络对待检目标特征的权重。实验表明,改进后的算法在 DIOR 数 据集的检测精度达到 78. 3%,比 YOLOX 高 2. 6%;在 RSOD 数据集的检测精度达到 95. 7%, 比 YOLOX 高 5. 6%。
关键词:遥感图像;目标检测;特征金字塔;注意力机制
中图分类号:TP 391. 4 文献标志码:A
引用格式:吴沭运,彭天文,刘国柱,等 . 基于改进特征金字塔和注意力的遥感目标检测算法 [J]. 青岛科技大学学报(自然科学版),2026,47(2):131-138.
WU Shuyun, PENG Tianwen, LIU Guozhu, et al. Remote sensing object detection algorithm based on improved feature pyramid and attention[J]. Journal of Qingdao University of Science and Technology(Natural Science Edition),2026,47(2):131-138.
Remote Sensing Object Detection Algorithm Based on Improved Feature Pyramid and Attention
WU Shuyun,PENG Tianwen,LIU Guozhu,LIANG Hongtao,CUI Wenjing (College of Information Science and Technology, Qingdao University of Science and Technology, Qingdao 266061,China)
Abstract:Due to the complex background and severe noise in remote sensing images, remote sensing object detection is a challenging task. In order to improve the detection precision of remote sensing targets, a remote sensing object detection algorithm based on a bidirectional sampling feature pyramid and a decoupled head with attention mechanism is proposed. The bidirectional sampling feature pyramid aims to aggregate deep feature maps with shallow net⁃ work feature maps to obtain more abundant spatial and semantic information, and the feature fusion ability of the pyramid network is further enhanced by adding path aggregation channels. The decoupled head with attention mechanism reduces the conflict of different level features by decoupling the prediction of target category, coordinate, and confidence, and introduces improved spatial attention and channel attention separately in the detection branch, further increasing the weight of the network's target feature. Experiments show that the improved algorithm achieves a detection precision of 78. 3% on the DIOR dataset, which is 2. 6% higher than that of YOLOX, and a detection precision of 95. 7% on the RSOD dataset, which is 5. 6% higher than that of YOLOX.
Key words:remote sensing images; object detection; feature pyramid; attention mechanism
收稿日期:2025-05-28
基金项目:国家自然科学基金项目(61973180,62172249);山东省产教融合研究生联合培养示范基地项目(2020-19).
作者简介:吴沭运(1998—),男,硕士研究生 . * 通信联系人 .