Scholarlink

<p>Optical remote sensing image target detection, particularly for small targets, poses significant challenges due to dense distributions, variable scales, and insufficient feature information. To address these issues, this paper proposes an Enhanced Receptive Field Multi-scale Feature Fusion Network (ERF-MNet). We introduce a Receptive Field Enhancement Aggregation Module (REAM) into the backbone network to expand its receptive field, enabling the extraction of richer multi-scale features. Secondly, the proposed Feature Pyramid Network based on Feature Refinement (FR-FPN), by virtue of multi-scale feature fusion and refined processing, demonstrates advantages in leveraging multi-scale features and accurately capturing small target details in optical remote sensing target detection, thus enhancing the detection accuracy and reliability. Furthermore, a Multilayer Feature Extraction Module (MFEM) is implemented to incorporate coordinate information, enhancing the network’s spatial perception and ability to capture intricate details of small targets. Experimental results on the DOTA and NWPU NHR-10 datasets show that ERF-MNet achieves average detection accuracies of 78.3% and 94.5%, respectively, with a computational complexity of 98.6&#xa0;G and a model size of 33.9&#xa0;M, and inference speeds of 36.5 FPS on DOTA and 108.2 FPS on NWPU VHR-10. These findings demonstrate that ERF-MNet significantly improves the detection precision of small targets while maintaining favorable real-time performance, thereby enhancing the overall effectiveness of optical remote sensing image target detection. And, these performance metrics underscore the substantial computational load and real-time processing requirements of the proposed model, making it a suitable candidate for deployment on high-performance computing (HPC) platforms to handle large-scale optical remote sensing data efficiently.</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

Enhanced receptive field multi-scale feature fusion for optical remote sensing target detection

  • 期刊论文

  • 发表日期：

    2026年3月14日

查看全文
  • Huilin Shan,
  • Shuoyang Wang,
  • Xin Chen,
  • Yanhao Zhang,
  • Xuewei Wang,
  • Yinsheng Zhang

摘要

Optical remote sensing image target detection, particularly for small targets, poses significant challenges due to dense distributions, variable scales, and insufficient feature information. To address these issues, this paper proposes an Enhanced Receptive Field Multi-scale Feature Fusion Network (ERF-MNet). We introduce a Receptive Field Enhancement Aggregation Module (REAM) into the backbone network to expand its receptive field, enabling the extraction of richer multi-scale features. Secondly, the proposed Feature Pyramid Network based on Feature Refinement (FR-FPN), by virtue of multi-scale feature fusion and refined processing, demonstrates advantages in leveraging multi-scale features and accurately capturing small target details in optical remote sensing target detection, thus enhancing the detection accuracy and reliability. Furthermore, a Multilayer Feature Extraction Module (MFEM) is implemented to incorporate coordinate information, enhancing the network’s spatial perception and ability to capture intricate details of small targets. Experimental results on the DOTA and NWPU NHR-10 datasets show that ERF-MNet achieves average detection accuracies of 78.3% and 94.5%, respectively, with a computational complexity of 98.6 G and a model size of 33.9 M, and inference speeds of 36.5 FPS on DOTA and 108.2 FPS on NWPU VHR-10. These findings demonstrate that ERF-MNet significantly improves the detection precision of small targets while maintaining favorable real-time performance, thereby enhancing the overall effectiveness of optical remote sensing image target detection. And, these performance metrics underscore the substantial computational load and real-time processing requirements of the proposed model, making it a suitable candidate for deployment on high-performance computing (HPC) platforms to handle large-scale optical remote sensing data efficiently.