Scholarlink

<p>The detection of small targets in aerial imagery poses significant challenges due to limited feature representation, complex backgrounds, and substantial scale variations. To address these issues, we propose SFFN-YOLO, a high-precision, lightweight detector specifically designed for small object detection in aerial images. We introduce a Spatial Feature Enhancement Module (SFEM) to enrich feature semantics and spatial information, reducing information loss. Additionally, an Improved Content-Aware ReAssembly of Features module (ICARAFE) is employed to enhance feature map quality through a content-aware upsampling process. Furthermore, a Feature Fusion Network (FFN) is designed to optimize the fusion of multi-scale features by automatically learning the weights of different feature maps. The detection head is modified to utilize higher resolution feature maps, enhancing the model’s sensitivity to tiny targets. Extensive experiments on three aerial image datasets, including Visdrone2019, UAVDT, and VEDAI, demonstrate the effectiveness and superiority of our proposed method, achieving state-of-the-art performance in small object detection tasks.</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

SFFN-YOLO for small object detection in aerial images

  • 期刊论文

  • 发表日期：

    2025年8月21日

查看全文
  • Hongying Zhang,
  • Jiatian Tang

摘要

The detection of small targets in aerial imagery poses significant challenges due to limited feature representation, complex backgrounds, and substantial scale variations. To address these issues, we propose SFFN-YOLO, a high-precision, lightweight detector specifically designed for small object detection in aerial images. We introduce a Spatial Feature Enhancement Module (SFEM) to enrich feature semantics and spatial information, reducing information loss. Additionally, an Improved Content-Aware ReAssembly of Features module (ICARAFE) is employed to enhance feature map quality through a content-aware upsampling process. Furthermore, a Feature Fusion Network (FFN) is designed to optimize the fusion of multi-scale features by automatically learning the weights of different feature maps. The detection head is modified to utilize higher resolution feature maps, enhancing the model’s sensitivity to tiny targets. Extensive experiments on three aerial image datasets, including Visdrone2019, UAVDT, and VEDAI, demonstrate the effectiveness and superiority of our proposed method, achieving state-of-the-art performance in small object detection tasks.