Scholarlink

<p>Remote sensing imagery is used across different fields, but practical applications still face important challenges, including backgrounds with multiple elements, objects that differ in scale, and limited processing resources on airborne and space platforms. These factors pose challenges for developing models that achieve strong performance while meeting limited resource requirements. This work introduces RSRD-Net, a real-time detection approach that operates with limited resources in remote sensing contexts. The model uses a structure built on FasterNet that includes operations that reduce processing to maintain the capacity to represent features. A module termed C3K2_EFAttention provides enhanced extraction across scales with minimal additional parameter cost. For detecting small targets, a component that combines scales, DyHead, reduces the number of floating-point operations and improves feature combination. Experiments on the HIT-UAV dataset show that RSRD-Net achieves 81.8% mAP@0.5, outperforming the YOLOv11s baseline by 3.7% points, while reducing computational cost to 8.2 GFLOPs, representing a 61.5% reduction compared with the baseline. The model maintains strong detection performance while substantially improving processing efficiency, providing a practical approach for real-time detection of remote sensing targets on resource-constrained platforms. The code will be gradually released at <a href="https://github.com/Sylver-Sun/RSRD-Net">https://github.com/Sylver-Sun/RSRD-Net</a> upon acceptance of this paper.</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

RSRD-Net: A Lightweight Model for Real-Time Target Detection in Remote Sensing Images

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月10日

查看全文
  • Yucheng Liu

摘要

Remote sensing imagery is used across different fields, but practical applications still face important challenges, including backgrounds with multiple elements, objects that differ in scale, and limited processing resources on airborne and space platforms. These factors pose challenges for developing models that achieve strong performance while meeting limited resource requirements. This work introduces RSRD-Net, a real-time detection approach that operates with limited resources in remote sensing contexts. The model uses a structure built on FasterNet that includes operations that reduce processing to maintain the capacity to represent features. A module termed C3K2_EFAttention provides enhanced extraction across scales with minimal additional parameter cost. For detecting small targets, a component that combines scales, DyHead, reduces the number of floating-point operations and improves feature combination. Experiments on the HIT-UAV dataset show that RSRD-Net achieves 81.8% mAP@0.5, outperforming the YOLOv11s baseline by 3.7% points, while reducing computational cost to 8.2 GFLOPs, representing a 61.5% reduction compared with the baseline. The model maintains strong detection performance while substantially improving processing efficiency, providing a practical approach for real-time detection of remote sensing targets on resource-constrained platforms. The code will be gradually released at https://github.com/Sylver-Sun/RSRD-Net upon acceptance of this paper.