<p>The capacity for precise target detection in UAV aerial images represents a pivotal prerequisite for the advancement of low-altitude economic activities. The technology for large target detection in images has attained a state of relative maturity, while the identification of small targets remains encumbered by challenges such as indistinct edge information, the absence of comprehensive deep feature information, and the constrained capacity for the expression of detection head information. To address these challenges, the MSD-YOLO11n target detection model has been proposed. Firstly, the P2 layer downsampling convolution is adopted as SPDConv, while the Multiscale Edge Information Selection (MEIS) module is proposed to replace the residual module in C3K2 in order to enhance the edge information feature extraction of small targets. Secondly, the Smalltarget Feature Enhancement Pyramid (SFEP) module has been proposed as a means of achieving the fusion of feature information between the P2 and P3 layers. This is achieved by combining Dysample up sampling with SPDConv down sampling and reconstructing low-quality images to obtain high-quality images using the CSP-OmniKernel module. Subsequently, the DyHead module was utilised to integrate scale, space and task-aware attention. Finally, the proposed MSD-YOLO11n model was evaluated based on the VisDrone2019 dataset through a combination of ablation and comparison experiments. In comparison with YOLO11n, both <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="44443_2025_116_Article_IEq1.gif" Format="GIF" Height="22" Rendition="HTML" Resolution="72" Type="Linedraw" Width="60" /> </InlineMediaObject> <EquationSource Format="TEX">${mAP}_{0.5}^{val}$</EquationSource> <EquationSource Format="MATHML"><math> <msubsup> <mrow> <mi mathvariant="italic">mAP</mi> </mrow> <mrow> <mn>0.5</mn> </mrow> <mrow> <mi mathvariant="italic">val</mi> </mrow> </msubsup> </math></EquationSource> </InlineEquation> and <InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="44443_2025_116_Article_IEq2.gif" Format="GIF" Height="21" Rendition="HTML" Resolution="72" Type="Linedraw" Width="64" /> </InlineMediaObject> <EquationSource Format="TEX">${mAP}_{0.5}^{test}$</EquationSource> <EquationSource Format="MATHML"><math> <msubsup> <mrow> <mi mathvariant="italic">mAP</mi> </mrow> <mrow> <mn>0.5</mn> </mrow> <mrow> <mi mathvariant="italic">test</mi> </mrow> </msubsup> </math></EquationSource> </InlineEquation> demonstrate enhancements of 20.9% and 19.7%, respectively. Furthermore, the parameters <InlineEquation ID="IEq3"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="44443_2025_116_Article_IEq3.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="37" /> </InlineMediaObject> <EquationSource Format="TEX">$A{P}_{vt}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>A</mi> <msub> <mi>P</mi> <mrow> <mi mathvariant="italic">vt</mi> </mrow> </msub> </mrow> </math></EquationSource> </InlineEquation>,<InlineEquation ID="IEq4"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="44443_2025_116_Article_IEq4.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="31" /> </InlineMediaObject> <EquationSource Format="TEX">$A{P}_{t}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>A</mi> <msub> <mi>P</mi> <mi>t</mi> </msub> </mrow> </math></EquationSource> </InlineEquation> and <InlineEquation ID="IEq5"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="44443_2025_116_Article_IEq5.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="32" /> </InlineMediaObject> <EquationSource Format="TEX">$A{P}_{s}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>A</mi> <msub> <mi>P</mi> <mi>s</mi> </msub> </mrow> </math></EquationSource> </InlineEquation> exhibit improvements of 44.4%, 26.9%, and 37.3%, respectively, for small targets. Furthermore, experimental evidence demonstrates that the MSD-YOLO11n model attains superior target detection accuracy compared to the TA-YOLO-n model, despite utilising a significantly reduced parameter count of 3.6 M. The efficacy of the proposed MSD-YOLO11n model in target detection tasks involving UAV aerial images has been substantiated.</p>

MSD-YOLO11n: an improved small target detection model for high precision UAV aerial imagery

期刊论文
出版模式：

开放获取
发表日期：

2025年7月3日

查看全文

Xin Zhang,
Baoyang Du,
Yaoxing Jia,
Kairui Luo,
Li Jiang

摘要

The capacity for precise target detection in UAV aerial images represents a pivotal prerequisite for the advancement of low-altitude economic activities. The technology for large target detection in images has attained a state of relative maturity, while the identification of small targets remains encumbered by challenges such as indistinct edge information, the absence of comprehensive deep feature information, and the constrained capacity for the expression of detection head information. To address these challenges, the MSD-YOLO11n target detection model has been proposed. Firstly, the P2 layer downsampling convolution is adopted as SPDConv, while the Multiscale Edge Information Selection (MEIS) module is proposed to replace the residual module in C3K2 in order to enhance the edge information feature extraction of small targets. Secondly, the Smalltarget Feature Enhancement Pyramid (SFEP) module has been proposed as a means of achieving the fusion of feature information between the P2 and P3 layers. This is achieved by combining Dysample up sampling with SPDConv down sampling and reconstructing low-quality images to obtain high-quality images using the CSP-OmniKernel module. Subsequently, the DyHead module was utilised to integrate scale, space and task-aware attention. Finally, the proposed MSD-YOLO11n model was evaluated based on the VisDrone2019 dataset through a combination of ablation and comparison experiments. In comparison with YOLO11n, both ${mAP}_{0.5}^{val}$ ${mAP}_{0.5}^{val}$ and ${mAP}_{0.5}^{test}$ ${mAP}_{0.5}^{test}$ demonstrate enhancements of 20.9% and 19.7%, respectively. Furthermore, the parameters $A{P}_{vt}$ $A P_{vt}$ , $A{P}_{t}$ $A P_{t}$ and $A{P}_{s}$ $A P_{s}$ exhibit improvements of 44.4%, 26.9%, and 37.3%, respectively, for small targets. Furthermore, experimental evidence demonstrates that the MSD-YOLO11n model attains superior target detection accuracy compared to the TA-YOLO-n model, despite utilising a significantly reduced parameter count of 3.6 M. The efficacy of the proposed MSD-YOLO11n model in target detection tasks involving UAV aerial images has been substantiated.