Scholarlink

<p>Real-time and accurate detection of small military aircraft in aerial images is critical for battlefield reconnaissance, yet remains challenging due to significant scale variations, complex backgrounds, and constrained onboard computing resources. In this paper, we propose an improved lightweight real-time detection transformer framework tailored for detecting small objects in drone-based aerial photography. Specifically, we replace the original ResNet18 backbone with ConvNextV2 to enhance fine-grained feature extraction while reducing model parameters. Additionally, we introduce EfficientViT’s Cascaded Group Attention in the encoder to replace the original AIFI attention, thereby improving the model’s ability to focus on key regions. To further refine bounding box regression, we integrate Inner IoU and MPDIoU strategies, which offer more robust spatial evaluations of predicted boxes. Extensive experiments on a military aircraft recognition dataset, containing 3842 images and over 20 aircraft types, demonstrate that our improved model reduces parameters by 38.91% compared to the baseline while increasing the mean average precision by 1.0 at a 50% IoU threshold. The final design also achieves near real-time performance, making it suitable for deployment in resource-constrained environments such as drone-based military reconnaissance. Overall, our contributions offer an efficient, accurate, and practical solution for small object detection tasks, laying a foundation for future lightweight object detection research and applications.</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

IRT-DETR: a lightweight transformer with ConvNextV2 and Cascaded Group Attention for real-time military aircraft detection in aerial imagery

  • 期刊论文

  • 发表日期：

    2026年1月19日

查看全文
  • Xinyong Lu,
  • Zhi Li,
  • Zheyu Zhang,
  • Runpeng Liu,
  • Yantong Liu

摘要

Real-time and accurate detection of small military aircraft in aerial images is critical for battlefield reconnaissance, yet remains challenging due to significant scale variations, complex backgrounds, and constrained onboard computing resources. In this paper, we propose an improved lightweight real-time detection transformer framework tailored for detecting small objects in drone-based aerial photography. Specifically, we replace the original ResNet18 backbone with ConvNextV2 to enhance fine-grained feature extraction while reducing model parameters. Additionally, we introduce EfficientViT’s Cascaded Group Attention in the encoder to replace the original AIFI attention, thereby improving the model’s ability to focus on key regions. To further refine bounding box regression, we integrate Inner IoU and MPDIoU strategies, which offer more robust spatial evaluations of predicted boxes. Extensive experiments on a military aircraft recognition dataset, containing 3842 images and over 20 aircraft types, demonstrate that our improved model reduces parameters by 38.91% compared to the baseline while increasing the mean average precision by 1.0 at a 50% IoU threshold. The final design also achieves near real-time performance, making it suitable for deployment in resource-constrained environments such as drone-based military reconnaissance. Overall, our contributions offer an efficient, accurate, and practical solution for small object detection tasks, laying a foundation for future lightweight object detection research and applications.