<p>Synthetic aperture radar (SAR) plays a pivotal role in critical civil applications due to its all-weather imaging capabilities. However, target detection in SAR images remains challenging, particularly for small vessels obscured by speckle noise and ambiguous feature characterization. To address these challenges, we propose SIE-YOLO11, a SAR-specific detection model based on the YOLO11 framework. SIE-YOLO11 employs the space-to-depth block module to preserve detailed features during downsampling and integrates the inverted efficient multiscale attention mechanism to enhance feature expression and contextual modeling. Furthermore, we optimize the detection head for small targets, which significantly improves the recognition capability. Experiments on the high resolution SAR images dataset (HRSID) and SAR ship detection dataset (SSDD) demonstrate that SIE-YOLO11 achieves 93.9% <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="371_2025_4148_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="50" /> </InlineMediaObject> <EquationSource Format="TEX">$\textit{mAP}_{50}$</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mi mathvariant="italic">mAP</mi> <mn>50</mn> </msub> </math></EquationSource> </InlineEquation> on HRSID and 98.4% <InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="371_2025_4148_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="50" /> </InlineMediaObject> <EquationSource Format="TEX">$\textit{mAP}_{50}$</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mi mathvariant="italic">mAP</mi> <mn>50</mn> </msub> </math></EquationSource> </InlineEquation> on SSDD, outperforming YOLO11n (91.6% on HRSID and 97.7% on SSDD) by 2.3% and 0.7% in <InlineEquation ID="IEq3"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="371_2025_4148_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="50" /> </InlineMediaObject> <EquationSource Format="TEX">$\textit{mAP}_{50}$</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mi mathvariant="italic">mAP</mi> <mn>50</mn> </msub> </math></EquationSource> </InlineEquation>, respectively. By COCO standards, on HRSID, it shows remarkable improvements compared with YOLO11n: <InlineEquation ID="IEq4"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="371_2025_4148_Article_IEq4.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="37" /> </InlineMediaObject> <EquationSource Format="TEX">$AP_{75}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>A</mi> <msub> <mi>P</mi> <mn>75</mn> </msub> </mrow> </math></EquationSource> </InlineEquation> is increased by 4.8%, and <InlineEquation ID="IEq5"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="371_2025_4148_Article_IEq5.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="34" /> </InlineMediaObject> <EquationSource Format="TEX">$AP_{S}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>A</mi> <msub> <mi>P</mi> <mi>S</mi> </msub> </mrow> </math></EquationSource> </InlineEquation> is enhanced by 6.1%. These results validate SIE-YOLO11’s superiority in SAR ship detection, providing a new technological approach for detecting similar small targets. The code is publicly available at <a href="https://github.com/wangjihang0239/SIE-YOLO11.git">GitHub repository</a>.</p>

Enhanced small-target detection in SAR images via SIE-YOLO11: a deep learning approach

期刊论文
发表日期：

2025年8月31日

查看全文

Jihang Wang,
Dezhi Han,
Xiang Shen,
Bing Han,
Zhongdai Wu

摘要

Synthetic aperture radar (SAR) plays a pivotal role in critical civil applications due to its all-weather imaging capabilities. However, target detection in SAR images remains challenging, particularly for small vessels obscured by speckle noise and ambiguous feature characterization. To address these challenges, we propose SIE-YOLO11, a SAR-specific detection model based on the YOLO11 framework. SIE-YOLO11 employs the space-to-depth block module to preserve detailed features during downsampling and integrates the inverted efficient multiscale attention mechanism to enhance feature expression and contextual modeling. Furthermore, we optimize the detection head for small targets, which significantly improves the recognition capability. Experiments on the high resolution SAR images dataset (HRSID) and SAR ship detection dataset (SSDD) demonstrate that SIE-YOLO11 achieves 93.9% $\textit{mAP}_{50}$ ${mAP}_{50}$ on HRSID and 98.4% $\textit{mAP}_{50}$ ${mAP}_{50}$ on SSDD, outperforming YOLO11n (91.6% on HRSID and 97.7% on SSDD) by 2.3% and 0.7% in $\textit{mAP}_{50}$ ${mAP}_{50}$ , respectively. By COCO standards, on HRSID, it shows remarkable improvements compared with YOLO11n: $AP_{75}$ $A P_{75}$ is increased by 4.8%, and $AP_{S}$ $A P_{S}$ is enhanced by 6.1%. These results validate SIE-YOLO11’s superiority in SAR ship detection, providing a new technological approach for detecting similar small targets. The code is publicly available at GitHub repository.