<p>Small object detection remains a fundamental yet challenging task in computer vision, aiming to accurately identify targets with limited pixel representation and minimal semantic information. The inherently low resolution and the interference from complex backgrounds often lead to poor detection performance by conventional methods. To address these challenges, we propose a novel Small Target Detection model based on Swin Transformer and Multiscale Fusion (STD-SM). Specifically, the Swin Transformer is employed to enhance the model’s capacity for global context modeling and long-range dependency capture. To improve feature fusion across different scales, the Bidirectional Feature Pyramid Network (BiFPN) is introduced, which utilizes an adaptive weighting mechanism to emphasize salient features of small objects. Additionally, the Convolutional Block Attention Module (CBAM) is integrated to further refine spatial and channel-wise feature representations, enhancing discriminative capability for small targets. Finally, an Intersection over Union Aware (IoU-Aware) loss function is adopted to optimize bounding box regression accuracy. Extensive experiments on the VisDrone2019 dataset show that the proposed method achieves an average precision (AP) of 33.2, outperforming state-of-the-art approaches by 2.8, 2.5, and 2.4 points in AP, <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11760_2025_4512_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="37" /> </InlineMediaObject> <EquationSource Format="TEX">$AP_{\scriptstyle 50}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>A</mi> <msub> <mi>P</mi> <mstyle displaystyle="false" scriptlevel="1"> <mn>50</mn> </mstyle> </msub> </mrow> </math></EquationSource> </InlineEquation>, and <InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11760_2025_4512_Article_IEq2.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="37" /> </InlineMediaObject> <EquationSource Format="TEX">$AP_{\scriptstyle 75}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>A</mi> <msub> <mi>P</mi> <mstyle displaystyle="false" scriptlevel="1"> <mn>75</mn> </mstyle> </msub> </mrow> </math></EquationSource> </InlineEquation>, respectively. These results clearly demonstrate the effectiveness and superiority of our approach in small object detection.</p>

Small target detection model based on swin transformer and multiscale fusion

期刊论文
发表日期：

2025年7月16日

查看全文

Jinjin Wang,
Meng Li

摘要

Small object detection remains a fundamental yet challenging task in computer vision, aiming to accurately identify targets with limited pixel representation and minimal semantic information. The inherently low resolution and the interference from complex backgrounds often lead to poor detection performance by conventional methods. To address these challenges, we propose a novel Small Target Detection model based on Swin Transformer and Multiscale Fusion (STD-SM). Specifically, the Swin Transformer is employed to enhance the model’s capacity for global context modeling and long-range dependency capture. To improve feature fusion across different scales, the Bidirectional Feature Pyramid Network (BiFPN) is introduced, which utilizes an adaptive weighting mechanism to emphasize salient features of small objects. Additionally, the Convolutional Block Attention Module (CBAM) is integrated to further refine spatial and channel-wise feature representations, enhancing discriminative capability for small targets. Finally, an Intersection over Union Aware (IoU-Aware) loss function is adopted to optimize bounding box regression accuracy. Extensive experiments on the VisDrone2019 dataset show that the proposed method achieves an average precision (AP) of 33.2, outperforming state-of-the-art approaches by 2.8, 2.5, and 2.4 points in AP, $AP_{\scriptstyle 50}$ $A P_{50}$ , and $AP_{\scriptstyle 75}$ $A P_{75}$ , respectively. These results clearly demonstrate the effectiveness and superiority of our approach in small object detection.