Scholarlink

<p>Detecting small objects in unmanned aerial vehicle (UAV) imagery remains a formidable challenge due to drastic scale variations, occlusion, and limited computational resources. Existing real-time detectors often struggle to capture sufficient global context and fine-grained details simultaneously. To address these issues, we propose MSBN–DFINE, a high-efficiency detector tailored for real-time UAV applications. At the core of our architecture is the multi-scale broadcast neck, which introduces a hub-centric broadcast paradigm to effectively suppress feature dilution. Specifically, we design a TinyStackFusion module that efficiently stacks lightweight <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(3\times 3\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>3</mn> <mo>×</mo> <mn>3</mn> </mrow> </math></EquationSource> </InlineEquation> convolutions to simulate the receptive fields of large kernels. This strategy dynamically aggregates multi-scale contexts while maintaining high efficiency, effectively meeting the requirements of real-time detection. Furthermore, we incorporate our efficient upshift channel block for precise detail recovery, and a lightweight BiMetaFormer to capture long-range dependencies and provide solid global context support. Extensive experiments on three benchmarks, including the VisDrone2019, UAVVaste, and CARPK datasets, demonstrate the superiority of our method. On the challenging VisDrone2019 benchmark, MSBN–DFINE achieves a robust <i>AP</i> of 30.0% and an <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(AP_{50}\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>A</mi> <msub> <mi>P</mi> <mn>50</mn> </msub> </mrow> </math></EquationSource> </InlineEquation> of 49.4%, representing substantial improvements of 2.9% and 3.8%, respectively, and outperforms the baseline D-FINE and other state-of-the-art detectors.</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

MSBN–DFINE: multi-scale broadcast neck with stacked small kernels for real-time UAV detection

  • 期刊论文

  • 发表日期：

    2026年5月2日

查看全文
  • Hongxing Peng,
  • Longbin Shi,
  • Huanai Liu,
  • Yan Chen

摘要

Detecting small objects in unmanned aerial vehicle (UAV) imagery remains a formidable challenge due to drastic scale variations, occlusion, and limited computational resources. Existing real-time detectors often struggle to capture sufficient global context and fine-grained details simultaneously. To address these issues, we propose MSBN–DFINE, a high-efficiency detector tailored for real-time UAV applications. At the core of our architecture is the multi-scale broadcast neck, which introduces a hub-centric broadcast paradigm to effectively suppress feature dilution. Specifically, we design a TinyStackFusion module that efficiently stacks lightweight \(3\times 3\) 3 × 3 convolutions to simulate the receptive fields of large kernels. This strategy dynamically aggregates multi-scale contexts while maintaining high efficiency, effectively meeting the requirements of real-time detection. Furthermore, we incorporate our efficient upshift channel block for precise detail recovery, and a lightweight BiMetaFormer to capture long-range dependencies and provide solid global context support. Extensive experiments on three benchmarks, including the VisDrone2019, UAVVaste, and CARPK datasets, demonstrate the superiority of our method. On the challenging VisDrone2019 benchmark, MSBN–DFINE achieves a robust AP of 30.0% and an \(AP_{50}\) A P 50 of 49.4%, representing substantial improvements of 2.9% and 3.8%, respectively, and outperforms the baseline D-FINE and other state-of-the-art detectors.