Scholarlink

<p>The parameter count of existing models for small object detection limits their applicability on resource-constrained devices. To address these challenges, this paper introduces a detail-enhanced lightweight network (DDCNet) for small object detection. DDCNet incorporates a detail feature compensation module at downsampling layers to recover lost detail information, a detail feature enhancement module between the backbone and neck to enhance detail features, and a cross-scale detail feature fusion module to merge multi-scale information. Experimental results demonstrate that DDCNet achieves state-of-the-art performance with a parameter count of only 1.75&#xa0;M, obtaining mAP50 scores of 48.9% and 55.4% on the VisDrone and AI-TOD datasets, respectively. Our work presents a novel solution to the issue of feature loss due to downsampling in small object detection, offering a balance between detection performance and model efficiency. The code for DDCNet can be found at: <a href="https://github.com/wxz0426/DDCNet">https://github.com/wxz0426/DDCNet</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

Detail-enhanced lightweight network for small object detection in aerial images

  • 期刊论文

  • 发表日期：

    2025年3月30日

查看全文
  • Xiongzhen Wang,
  • Chuan Lin,
  • Yongcai Pan

摘要

The parameter count of existing models for small object detection limits their applicability on resource-constrained devices. To address these challenges, this paper introduces a detail-enhanced lightweight network (DDCNet) for small object detection. DDCNet incorporates a detail feature compensation module at downsampling layers to recover lost detail information, a detail feature enhancement module between the backbone and neck to enhance detail features, and a cross-scale detail feature fusion module to merge multi-scale information. Experimental results demonstrate that DDCNet achieves state-of-the-art performance with a parameter count of only 1.75 M, obtaining mAP50 scores of 48.9% and 55.4% on the VisDrone and AI-TOD datasets, respectively. Our work presents a novel solution to the issue of feature loss due to downsampling in small object detection, offering a balance between detection performance and model efficiency. The code for DDCNet can be found at: https://github.com/wxz0426/DDCNet.