Scholarlink

<p>To address critical challenges in UAV infrared object detection, such as low contrast between targets and backgrounds, the scarcity of salient features, and the difficulty of detecting small targets, this paper proposes a lightweight and efficient detection method named CFPM-DETR. First, to enhance feature extraction capabilities while reducing model complexity, we introduce the cross-stage adaptive mixer module, which efficiently captures multi-scale features and lowers computational overhead. Subsequently, we design the frequency fusion feature pyramid network to enhance cross-scale feature fusion from a frequency domain perspective, effectively sharpening target boundaries. Additionally, the polarity spatial enhancement block was introduced to enhance the model’s ability to perceive local details. Finally, we design the MAL-WIoU loss function to synergistically optimize the localization and classification of difficult small targets. Compared to the baseline RT-DETR-r18 with an input resolution of 640<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\times \)</EquationSource> <EquationSource Format="MATHML"><math> <mo>×</mo> </math></EquationSource> </InlineEquation>640, CFPM-DETR reduces its parameters and computational cost by 25.0% and 22.8% respectively, while increasing FPS by 29.9%. On the HIT-UAV dataset, mAP50 and mAP50:95 improved by 4.3% and 2.5% respectively, and on the NCHU-A2G-SIRST dataset, mAP50 increased by 1.5%. The experimental results demonstrate that the proposed model achieves an excellent balance between detection accuracy and computational efficiency.</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

CFPM-DETR: a lightweight model for UAV infrared small target detection

  • 期刊论文

  • 发表日期：

    2026年1月28日

查看全文
  • Shaoqiang Nan,
  • Sheng Bin,
  • Gengxin Sun

摘要

To address critical challenges in UAV infrared object detection, such as low contrast between targets and backgrounds, the scarcity of salient features, and the difficulty of detecting small targets, this paper proposes a lightweight and efficient detection method named CFPM-DETR. First, to enhance feature extraction capabilities while reducing model complexity, we introduce the cross-stage adaptive mixer module, which efficiently captures multi-scale features and lowers computational overhead. Subsequently, we design the frequency fusion feature pyramid network to enhance cross-scale feature fusion from a frequency domain perspective, effectively sharpening target boundaries. Additionally, the polarity spatial enhancement block was introduced to enhance the model’s ability to perceive local details. Finally, we design the MAL-WIoU loss function to synergistically optimize the localization and classification of difficult small targets. Compared to the baseline RT-DETR-r18 with an input resolution of 640 \(\times \) × 640, CFPM-DETR reduces its parameters and computational cost by 25.0% and 22.8% respectively, while increasing FPS by 29.9%. On the HIT-UAV dataset, mAP50 and mAP50:95 improved by 4.3% and 2.5% respectively, and on the NCHU-A2G-SIRST dataset, mAP50 increased by 1.5%. The experimental results demonstrate that the proposed model achieves an excellent balance between detection accuracy and computational efficiency.