Scholarlink

<p>Object detection holds significant importance in remote sensing applications. However, the view angle and dynamic platform increase complexity compared to traditional tasks, such as limited feature representation, complex backgrounds, and data transmission issues. To address these challenges, ESOD-YOLO, an efficient detector based on YOLOv8 that optimally balances accuracy and efficiency is proposed in this work. ESOD-YOLO integrates three lightweight and plug-and-play modules: Deformable and Global–Local Feature Guidance (DGLFG), Simple Cross-scale and Cross-layer Feature Fusion(SCCFF), and Inverse-residual Shared Parameter based on a Partial convolution (ISPP). These modules improve the model’s capabilities to sense target deformation, local and global correlations in spatial and channel dimensions, and multi-scale features fusion, respectively, avoiding increasing model complexity. Experiments are performed on three typical small-target datasets: VisDrone2019, AI-TOD, and UAVDT. The <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="607_2024_1398_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="52" /> </InlineMediaObject> <EquationSource Format="TEX">\(mAP_{50}\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>m</mi> <mi>A</mi> <msub> <mi>P</mi> <mn>50</mn> </msub> </mrow> </math></EquationSource> </InlineEquation> of ESOD-YOLO reaches 40.6%, 48.3%, 51.7%, meanwhile, results in VisDrone2019 reveal significant improvements over the baseline: the mAP improved by 3.6% and 2.7%, respectively, while the parameter count is reduced by 60%. These results demonstrate the efficiency of the proposed innovative approach.</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

ESOD-YOLO: an enhanced efficient small object detection framework for aerial images

  • 期刊论文

  • 发表日期：

    2025年1月18日

查看全文
  • Xin Xu,
  • Qi Li,
  • Jie Pan,
  • Xingzheng Lu,
  • Hongwei Wei,
  • Mingzheng Sun,
  • Haoze Zhang

摘要

Object detection holds significant importance in remote sensing applications. However, the view angle and dynamic platform increase complexity compared to traditional tasks, such as limited feature representation, complex backgrounds, and data transmission issues. To address these challenges, ESOD-YOLO, an efficient detector based on YOLOv8 that optimally balances accuracy and efficiency is proposed in this work. ESOD-YOLO integrates three lightweight and plug-and-play modules: Deformable and Global–Local Feature Guidance (DGLFG), Simple Cross-scale and Cross-layer Feature Fusion(SCCFF), and Inverse-residual Shared Parameter based on a Partial convolution (ISPP). These modules improve the model’s capabilities to sense target deformation, local and global correlations in spatial and channel dimensions, and multi-scale features fusion, respectively, avoiding increasing model complexity. Experiments are performed on three typical small-target datasets: VisDrone2019, AI-TOD, and UAVDT. The \(mAP_{50}\) m A P 50 of ESOD-YOLO reaches 40.6%, 48.3%, 51.7%, meanwhile, results in VisDrone2019 reveal significant improvements over the baseline: the mAP improved by 3.6% and 2.7%, respectively, while the parameter count is reduced by 60%. These results demonstrate the efficiency of the proposed innovative approach.