Scholarlink

<p>Rapid advancement in computer vision and deep learning techniques has revolutionised tiny object detection in several real-world applications such as maritime surveillance, traffic monitoring, autonomous driving, and disaster management. Tiny targets often carry essential information, particularly in applications where every pixel matters. In aerial and medical images, tiny-object detection plays a significant role in ensuring safety by processing all essential visual cues in an image. Detecting meticulous details from an image becomes a considerable challenge in a standard object detection method, and it struggles with the loss of small details due to scale variations. Many researchers have addressed these challenges and developed deep learning-based techniques such as You Only Look Once (YOLO), Faster-RCNN, and RetinaNet, and used with multi-size feature fusion to improve the ability to detect small objects in drone-captured images. Despite these developments, many of these techniques face difficulty in detecting fine-grained details and tiny targets in low-illuminance environments. This paper introduces YOLOTR, a novel optimised transformer-based object detection technique designed to enhance global contextual information and improve detection performance under low-illumination conditions. This proposed model uses YOLOv11 as a baseline architecture and introduce C3TR module with substantial architectural modifications for enhanced multi-scale feature fusion and global contextual modelling. By utilising a transformer-based attention mechanism, YOLOTR improves long-range dependency learning and strengthens semantic consistency across scales. This model enables efficient detection of tiny and low-illuminance objects under UAV imagery. Various experiments were conducted using the VisDrone data to examine the efficiency of the YOLOTR model. The experimental results indicate that the YOLOTR model achieves better precision, recall, and mAP@50 than the existing models. It predicts objects in low-light conditions more effectively than the baseline model.</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

YOLOTR: improved multi-size feature fusion using transformer-based YOLO model for tiny-object detection in low-illuminance drone images

  • 期刊论文

  • 发表日期：

    2026年4月28日

查看全文
  • R. Priyadharshini,
  • P. Shanthi Bala,
  • S. Ajeeth,
  • C. Kaviyazhiny

摘要

Rapid advancement in computer vision and deep learning techniques has revolutionised tiny object detection in several real-world applications such as maritime surveillance, traffic monitoring, autonomous driving, and disaster management. Tiny targets often carry essential information, particularly in applications where every pixel matters. In aerial and medical images, tiny-object detection plays a significant role in ensuring safety by processing all essential visual cues in an image. Detecting meticulous details from an image becomes a considerable challenge in a standard object detection method, and it struggles with the loss of small details due to scale variations. Many researchers have addressed these challenges and developed deep learning-based techniques such as You Only Look Once (YOLO), Faster-RCNN, and RetinaNet, and used with multi-size feature fusion to improve the ability to detect small objects in drone-captured images. Despite these developments, many of these techniques face difficulty in detecting fine-grained details and tiny targets in low-illuminance environments. This paper introduces YOLOTR, a novel optimised transformer-based object detection technique designed to enhance global contextual information and improve detection performance under low-illumination conditions. This proposed model uses YOLOv11 as a baseline architecture and introduce C3TR module with substantial architectural modifications for enhanced multi-scale feature fusion and global contextual modelling. By utilising a transformer-based attention mechanism, YOLOTR improves long-range dependency learning and strengthens semantic consistency across scales. This model enables efficient detection of tiny and low-illuminance objects under UAV imagery. Various experiments were conducted using the VisDrone data to examine the efficiency of the YOLOTR model. The experimental results indicate that the YOLOTR model achieves better precision, recall, and mAP@50 than the existing models. It predicts objects in low-light conditions more effectively than the baseline model.