| 摘要: |
| 目的 针对现有遥感图像目标检测算法在目标尺度变化大以及存在复杂背景信息干扰的场景中检测精度不
佳的问题,提出一种基于多尺度下采样的遥感图像目标检测算法 MSD-YOLO( Multi-Scale Downsampling-YOLO) 。
方法 首先设计一种多尺度下采样( Multi-Scale Downsampling, MSD) 模块,通过 3 个并行下采样分支在进行下采样
的同时实现多尺度特征提取,避免了现有模型在经过多次下采样后,小目标特征信息丢失严重的问题,并引入具有
自适应激活特性的 ACON( Activate or Not) 激活函数,提升模型泛化能力;其次对三重注意力( Triplet Attention) 机制
进行改进,提出 ITA( Improved Triplet Attention) 机制,通过捕获跨维度交互并强调空间注意力自适应地调整特征图
的权重分配,提升模型在复杂背景信息干扰场景中的检测性能。 结果 实验结果表明:MSD-YOLO 在 NWPU VHR-
10 及 RSOD 数据集上的平均精度( mAP) 分别达到 94. 9%及 96. 8%,相较于基线网络 YOLOv7,均提升了 1. 5%,并
且精度优于其他经典网络模型。 结论 提出的 MSD-YOLO 算法可以有效提升在尺度变化大以及存在复杂背景信息
干扰等场景中的检测精度,在遥感图像目标检测场景中有着一定的应用价值。 |
| 关键词: 计算机视觉 遥感图像目标检测 MSD-YOLO YOLOv7 Triplet Attention |
| DOI： |
| 分类号: |
| 基金项目: |
|
| Remote Sensing Image Object Detection Algorithm Based on Multi-scale Downsampling |
|
ZHOU Huaping LIU Weidong
|
|
School of Computer Science and Engineering Anhui University of Science and Technology Anhui Huainan 232001
China
|
| Abstract: |
| Objective Aiming at the problem of poor detection accuracy of existing remote sensing image object detection
algorithms in scenes with large variations of object?? s scales and complex background information interference a remote
sensing image object detection algorithm MSD-YOLO Multi-Scale Downsampling-YOLO based on multi-scale
downsampling was proposed. Methods Firstly a multi-scale downsampling MSD module was designed which
employed three parallel downsampling branches to extract multi-scale features simultaneously during downsampling. This
approach avoided the problem of severe loss of small target feature information after many times of downsampling in
existing models. Additionally an ACON Activate or Not activation function with adaptive activation characteristics was
introduced to enhance the model?? s generalization ability. Secondly the Triplet Attention mechanism was improved and
the Improved Triplet Attention ITA mechanism was proposed to adaptively adjust the weight allocation of feature maps by
capturing cross-dimensional interactions and emphasizing spatial attention. This mechanism improved the detection
performance of the model in scenes with complex background information interference. Results The experimental results showed that the mean average precisions mAP of MSD-YOLO on NWPU VHR-10 and RSOD datasets reached 94. 9%
and 96. 8% respectively which were both improved by 1. 5% compared with the baseline network YOLOv7 and the
precisions outperformed those of other classical network models. Conclusion The proposed MSD-YOLO algorithm
effectively improves detection accuracy in scenarios with significant scale variations and complex background interference
and has valuable applications in remote sensing image object detection scenarios. |
| Key words: computer vision remote sensing image object detection MSD-YOLO YOLOv7 Triplet Attention |
