| 冉宁,施高朗,张昊俣,张少康,郝晋渊.基于YOLO11的轻量级遥感图像检测算法[J].电子测量与仪器学报,2026,40(3):124-134 |
| 基于YOLO11的轻量级遥感图像检测算法 |
| Lightweight remote sensing image detection algorithm based on YOLO11 |
| |
| DOI： |
| 中文关键词: 遥感小目标 YOLO11 轻量化 深度学习 |
| 英文关键词:remote sensing small objects YOLO11 lightweight deep learning |
| 基金项目:国家自然科学基金（62373132）、河北省自然科学基金（F2025201023）、石家庄市驻冀高校基础研究项目（241791367A）、河北大学优秀青年科研创新团队建设项目（QNTD202411）、河北大学多学科交叉研究计划项目（DXK202409）资助 |
|
|
| Author | Institution |
| Ran Ning | 1.School of Electronic Informational Engineering, Hebei University, Baoding 071002, China; 2.Laboratory of Energy-Saving
Technology, Hebei University, Baoding 071002, China |
| Shi Gaolang | School of Electronic Informational Engineering, Hebei University, Baoding 071002, China |
| Zhang Haoyu | School of Electronic Informational Engineering, Hebei University, Baoding 071002, China |
| Zhang Shaokang | School of Cyber Security and Computer, Hebei University,
Baoding 071002, China |
| Hao Jinyuan | 4.HBU-UCLAN School of Media, Communication and Creative Industries, Hebei University,
Baoding 071002, China; 5.Laboratory of IoT Technology, Hebei University, Baoding 071002, China |
|
| 摘要点击次数: 384 |
| 全文下载次数: 243 |
| 中文摘要: |
| 针对现有的目标检测算法在面对遥感小目标检测难题时存在网络体积大、模型复杂度高、检测准确率低等问题,提出了一种基于YOLO11的轻量级遥感算法EDB-YOLO11。该算法主要从模型结构和损失函数两个方面进行了改进。首先,采用EMBC模块改进了原本的C3K2模块,有效地提升了网络结构的特征表达能力;其次,采用Downsample下采样模块替换了原本的卷积下采样,在降低模型参数量和计算量的同时提升了主干部分的特征提取能力;然后,采用BiFPN特征融合方式替换了原本的PANet,增强了网络的特征融合能力,实现了更高效的多尺度特征融合;最后,结合Focal机制与WIoU的优势,引入了一种新的损失函数Focal-WIoU,使模型更关注高质量的训练样本,从而降低了低质量训练样本的影响,有效地提升了模型的检测精度。实验结果表明,EDB-YOLO11算法的参数量下降了27.18%,计算量下降了16.43%,同时在VisDrone2019数据集上的mAP@0.5提升了3.4%,mAP@0.5：0.95提升了1.8%,在泛化数据集SIMD和MAR20上,mAP@0.5分别提升了3.8%和0.3%,mAP@0.5：0.95分别提升了3%和0.2%。上述实验结果验证了EDB-YOLO11算法在遥感小目标检测任务中的有效性与优越性。 |
| 英文摘要: |
| To address the limitations of existing object detection algorithms in remote sensing small object detection—such as large model size, high computational complexity, and low detection accuracy—this paper proposes a lightweight remote sensing detection algorithm based on YOLO11, named EDB-YOLO11. The algorithm introduces improvements from two aspects: network architecture and loss function. First, the EMBC module is designed to replace the original C3K2 module, which effectively enhances the feature representation capability of the network. Second, a novel Downsample module is employed instead of the traditional convolution-based downsampling, which reduces the number of parameters and computational cost while improving the feature extraction ability of the backbone. Third, BiFPN is adopted to replace the original PANet for feature fusion, significantly enhancing the network’s multi-scale feature integration efficiency. Finally, a new loss function called Focal-WIoU is proposed by combining the advantages of the Focal mechanism and WIoU loss. This design enables the model to focus more on high-quality training samples and reduce the impact of low-quality samples, thereby improving overall detection accuracy. Experimental results show that EDB-YOLO11 reduces the number of parameters by 27.18% and the computational cost by 16.43%. On the VisDrone2019 dataset, the mAP@0.5 increases by 3.4%, and the mAP@0.5：0.95 improves by 1.8%. On the generalized remote sensing datasets SIMD and MAR20, the mAP@0.5 improves by 3.8% and 0.3%, respectively, while the mAP@0.5：0.95 improves by 3.0% and 0.2%, respectively. These results demonstrate the effectiveness and superiority of the proposed EDB-YOLO11 algorithm in remote sensing small object detection tasks. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
