Scholarlink

<p>Infrared imaging plays a pivotal role in overcoming the limitations of traditional visible-light imaging, especially under complex environmental conditions. Object detection within infrared images is crucial across various domains, but challenges such as low target pixel resolution often hinder detection performance. This paper introduces the ADaptive DEtection TRansformer (AD-DETR), designed to address these challenges. Firstly, we propose the adaptive weight division network (AWD-Net) to enhance multi-scale target feature extraction through dynamic weight partitioning. Secondly, the adaptive hierarchical feature path aggregation network (AHF-PAN) integrates local–global attention mechanisms with multi-branch feature fusion, improving the model’s multi-scale feature representation. Finally, the polarized dynamic attention fusion (PDAF) mechanism strengthens nonlinear modeling capabilities. Experimental results demonstrate that AD-DETR achieves 77.6% detection precision, with 16.3% and 13.5% reductions in parameter count and computational complexity, respectively, compared to Real-Time DEtection TRansformer (RT-DETR), while improving mAP@0.5 and mAP@0.5:0.95 by 4.9% and 4.5%. These results validate the superior capability of our proposed model in infrared image target recognition tasks. Our source code is available at: <a href="https://github.com/zcfanhua/AD-DETR">https://github.com/zcfanhua/AD-DETR</a>.</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

Adaptive transformer-based detection: enhancing infrared image target recognition

  • 期刊论文

  • 发表日期：

    2026年1月11日

查看全文
  • Chong Zhang,
  • Jie Yue,
  • Jianglong Fu

摘要

Infrared imaging plays a pivotal role in overcoming the limitations of traditional visible-light imaging, especially under complex environmental conditions. Object detection within infrared images is crucial across various domains, but challenges such as low target pixel resolution often hinder detection performance. This paper introduces the ADaptive DEtection TRansformer (AD-DETR), designed to address these challenges. Firstly, we propose the adaptive weight division network (AWD-Net) to enhance multi-scale target feature extraction through dynamic weight partitioning. Secondly, the adaptive hierarchical feature path aggregation network (AHF-PAN) integrates local–global attention mechanisms with multi-branch feature fusion, improving the model’s multi-scale feature representation. Finally, the polarized dynamic attention fusion (PDAF) mechanism strengthens nonlinear modeling capabilities. Experimental results demonstrate that AD-DETR achieves 77.6% detection precision, with 16.3% and 13.5% reductions in parameter count and computational complexity, respectively, compared to Real-Time DEtection TRansformer (RT-DETR), while improving mAP@0.5 and mAP@0.5:0.95 by 4.9% and 4.5%. These results validate the superior capability of our proposed model in infrared image target recognition tasks. Our source code is available at: https://github.com/zcfanhua/AD-DETR.