Scholarlink

<p>Object detection in remote sensing images is an important research topic in the field of computer vision. Since the object sizes in remote sensing images varies greatly and there are more complex background objects included, existing detectors often struggle with detection performance or incur heavy computational costs. Therefore, a single-stage detector more suitable for remote sensing images is proposed based on the YOLOv8 model in this paper. Firstly, we propose a lightweight convolution and improve the C2f module by using the proposed convolution as the basic unit, which reduces the number of parameters while maintaining its feature extraction capability. Secondly, to enable model to better meet the receptive field requirements of multi-scale objects, we embed the LSK modules before the PDC2f layers in the backbone network to form cascade structures, a step that enhances multi-scale feature fusion capability of the model. In addition, aiming at lightweighting the neck network, the GSConv and VoVGSCSP modules are used to construct slim-neck. Finally, a grouped spatial attention mechanism is designed to fuse the difference information between global and local features and enable the model to better capture key features. To validate the effectiveness of the proposed model, we conduct experiments on the DIOR and TGRS-HRRSD datasets. The mAP of the improved model is boosted by 1.4% and 2.2% with respect to the baseline model YOLOv8s, while the number of parameters and computational overhead of the model are reduced by 14.4% and 20.2%. In summary, our model strikes a good balance between detection performance and model complexity. The code will be available at <a href="https://github.com/YongliLiu/BA-YOLO.">https://github.com/YongliLiu/BA-YOLO.</a></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

BA-YOLO: a balanced object detector based on lightweight convolution and attention mechanism for remote sensing images

  • 期刊论文

  • 发表日期：

    2026年2月16日

查看全文
  • Yongli Liu,
  • Degang Yang,
  • Tingting Song,
  • Yichen Ye,
  • Xin Zhang

摘要

Object detection in remote sensing images is an important research topic in the field of computer vision. Since the object sizes in remote sensing images varies greatly and there are more complex background objects included, existing detectors often struggle with detection performance or incur heavy computational costs. Therefore, a single-stage detector more suitable for remote sensing images is proposed based on the YOLOv8 model in this paper. Firstly, we propose a lightweight convolution and improve the C2f module by using the proposed convolution as the basic unit, which reduces the number of parameters while maintaining its feature extraction capability. Secondly, to enable model to better meet the receptive field requirements of multi-scale objects, we embed the LSK modules before the PDC2f layers in the backbone network to form cascade structures, a step that enhances multi-scale feature fusion capability of the model. In addition, aiming at lightweighting the neck network, the GSConv and VoVGSCSP modules are used to construct slim-neck. Finally, a grouped spatial attention mechanism is designed to fuse the difference information between global and local features and enable the model to better capture key features. To validate the effectiveness of the proposed model, we conduct experiments on the DIOR and TGRS-HRRSD datasets. The mAP of the improved model is boosted by 1.4% and 2.2% with respect to the baseline model YOLOv8s, while the number of parameters and computational overhead of the model are reduced by 14.4% and 20.2%. In summary, our model strikes a good balance between detection performance and model complexity. The code will be available at https://github.com/YongliLiu/BA-YOLO.