Scholarlink

<p>In UAV aerial imagery, small objects often occupy minimal pixel regions and are prone to being obscured by noisy backgrounds, posing significant detection challenges. Existing methods primarily rely on multi-scale information fusion to address these issues, yet their overall detection accuracy for small objects remains limited. To tackle this, we propose ASOD-YOLOX, an aerial small object detection algorithm based on YOLOX, which integrates advancements in multi-scale fusion, contextual information, and attention mechanisms to enhance feature representation and discrimination. Specifically, the backbone network incorporates the CrissCross Attention module to improve multi-scale feature representation, while the Bidirectional Adaptive Layered Weight Fusion Network (BALWF-Net) employs bidirectional information flow and a Dynamic Adaptive Feature Fusion (DASF) module to adaptively adjust fusion weights, addressing information loss in traditional unidirectional flows. Additionally, the Multi-Scale Self-Attention Network (MSA-NET) leverages dilated convolutions and self-attention mechanisms to enrich contextual information, enhance the separation of targets from the background, and reduce false positives, leading to improved small object detection performance. Results obtained from experiments using VisDrone-2021, UAVDT, and MS COCO datasets indicate that the <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11227_2025_7243_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="34" /> </InlineMediaObject> <EquationSource Format="TEX">\(\text {AP}_{S}\)</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mtext>AP</mtext> <mi>S</mi> </msub> </math></EquationSource> </InlineEquation> values of ASOD-YOLOX are 31.1%, 27.2%, and 23.1%, respectively. These findings surpass existing advanced methods, demonstrating their capability in detecting small targets within aerial images.</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

ASOD-YOLOX: a study on small object detection in aerial images based on YOLOX

  • 期刊论文

  • 发表日期：

    2025年4月16日

查看全文
  • Hongying Zhang,
  • Wentao Liu,
  • Enyao Chen

摘要

In UAV aerial imagery, small objects often occupy minimal pixel regions and are prone to being obscured by noisy backgrounds, posing significant detection challenges. Existing methods primarily rely on multi-scale information fusion to address these issues, yet their overall detection accuracy for small objects remains limited. To tackle this, we propose ASOD-YOLOX, an aerial small object detection algorithm based on YOLOX, which integrates advancements in multi-scale fusion, contextual information, and attention mechanisms to enhance feature representation and discrimination. Specifically, the backbone network incorporates the CrissCross Attention module to improve multi-scale feature representation, while the Bidirectional Adaptive Layered Weight Fusion Network (BALWF-Net) employs bidirectional information flow and a Dynamic Adaptive Feature Fusion (DASF) module to adaptively adjust fusion weights, addressing information loss in traditional unidirectional flows. Additionally, the Multi-Scale Self-Attention Network (MSA-NET) leverages dilated convolutions and self-attention mechanisms to enrich contextual information, enhance the separation of targets from the background, and reduce false positives, leading to improved small object detection performance. Results obtained from experiments using VisDrone-2021, UAVDT, and MS COCO datasets indicate that the \(\text {AP}_{S}\) AP S values of ASOD-YOLOX are 31.1%, 27.2%, and 23.1%, respectively. These findings surpass existing advanced methods, demonstrating their capability in detecting small targets within aerial images.