Scholarlink

<p>Small target detection in UAV aerial imagery encounters challenges such as low target resolution and significant background interference. Current detection models in such scenarios exhibit issues with insufficient accuracy and computational redundancy. To overcome these challenges, this paper introduces an advanced UAV object detection algorithm based on YOLOv8-DDH-YOLO. This algorithm employs a dual-detection-head (DDH) architecture to replace the original three-detection-head structure, thereby enhancing small target detection capabilities while reducing the parameter count by 43.8%. Furthermore, a novel C2f-DD module is designed, integrating dilated convolution and residual connections to improve feature representation through Dilated Wise Residual (DWR) and Dilated Reparam Block (DRB). To further optimize feature fusion, a polymorphic feature fusion pyramid network (PFPN) is developed, utilizing a cross-layer feature interaction mechanism to enhance detail representation while reducing the model size. To further improve small target feature extraction, a feature enhancement bottleneck module is incorporated at the front end of the detection head. Experimental results on the VisDrone2019 dataset show that the DDH-YOLO model achieves an mAP@50 of 40.4% and an mAP@50:95 of 24.2%, representing improvements of 7.6% and 5.1% over the baseline model, with a 44.5% reduction in parameter count to 1.67M. The proposed model effectively balances detection accuracy and computational efficiency, providing a valuable reference for real-time detection tasks on UAV platforms.</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

DDH-YOLO: A dual-head YOLOv8 model for small object detection in UAV aerial images

  • 期刊论文

  • 发表日期：

    2025年6月27日

查看全文
  • Yangye Liu,
  • Zongchang Yang

摘要

Small target detection in UAV aerial imagery encounters challenges such as low target resolution and significant background interference. Current detection models in such scenarios exhibit issues with insufficient accuracy and computational redundancy. To overcome these challenges, this paper introduces an advanced UAV object detection algorithm based on YOLOv8-DDH-YOLO. This algorithm employs a dual-detection-head (DDH) architecture to replace the original three-detection-head structure, thereby enhancing small target detection capabilities while reducing the parameter count by 43.8%. Furthermore, a novel C2f-DD module is designed, integrating dilated convolution and residual connections to improve feature representation through Dilated Wise Residual (DWR) and Dilated Reparam Block (DRB). To further optimize feature fusion, a polymorphic feature fusion pyramid network (PFPN) is developed, utilizing a cross-layer feature interaction mechanism to enhance detail representation while reducing the model size. To further improve small target feature extraction, a feature enhancement bottleneck module is incorporated at the front end of the detection head. Experimental results on the VisDrone2019 dataset show that the DDH-YOLO model achieves an mAP@50 of 40.4% and an mAP@50:95 of 24.2%, representing improvements of 7.6% and 5.1% over the baseline model, with a 44.5% reduction in parameter count to 1.67M. The proposed model effectively balances detection accuracy and computational efficiency, providing a valuable reference for real-time detection tasks on UAV platforms.