Scholarlink

<p>In recent years, UAV aerial object detection has been widely applied in fields such as urban planning, environmental monitoring, agricultural management, and disaster assessment. However, challenges such as low accuracy, missed detections, and occlusion still persist when detecting small objects from aerial perspectives. This study proposes the PSO—YOLO (Precise Small Object—You Only Look Once) algorithm, which is developed based on multi-scale feature fusion and extraction techniques for small object detection. Firstly, the MSFE (Multi-Scale Feature Extraction) module was developed, leveraging Transformer-based multi-scale feature fusion and self-attention mechanisms to enhance the acquisition of fine-grained multi-scale details. Secondly, a small object-enhanced STE-Neck (Small Object Enhancement-Neck) network is introduced to fuse multi-scale features during downsampling, effectively capturing detailed and local features to prevent missed detections and feature loss due to occlusion. The fused features are then input into a small object detection head. Finally, the Swin Transformer is incorporated into the YOLOv8n backbone network to improve the C2f module, thereby enhancing its ability to capture long—range contextual information and further improve small object feature extraction. Additionally, the SPPF—LSKA (Spatial Pyramid Pooling Fast—Large Separable Kernel Attention) module is introduced to replace the SPPF (Spatial Pyramid Pooling Fast) module, aiming to enhance the algorithm's ability to extract key features. Experiments on the VisDrone2019 dataset show that the PSO-YOLO algorithm achieves an mAP@50% of 30.9% and an mAP@50–95% of 17.4%, representing improvements of 4.6% and 2.8%, respectively, over YOLOv8n.</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

PSO-YOLO: a contextual feature enhancement method for small object detection in UAV aerial images

  • 期刊论文

  • 发表日期：

    2025年2月15日

查看全文
  • Zhihong Zhao,
  • Xinyue Liu,
  • Peng He

摘要

In recent years, UAV aerial object detection has been widely applied in fields such as urban planning, environmental monitoring, agricultural management, and disaster assessment. However, challenges such as low accuracy, missed detections, and occlusion still persist when detecting small objects from aerial perspectives. This study proposes the PSO—YOLO (Precise Small Object—You Only Look Once) algorithm, which is developed based on multi-scale feature fusion and extraction techniques for small object detection. Firstly, the MSFE (Multi-Scale Feature Extraction) module was developed, leveraging Transformer-based multi-scale feature fusion and self-attention mechanisms to enhance the acquisition of fine-grained multi-scale details. Secondly, a small object-enhanced STE-Neck (Small Object Enhancement-Neck) network is introduced to fuse multi-scale features during downsampling, effectively capturing detailed and local features to prevent missed detections and feature loss due to occlusion. The fused features are then input into a small object detection head. Finally, the Swin Transformer is incorporated into the YOLOv8n backbone network to improve the C2f module, thereby enhancing its ability to capture long—range contextual information and further improve small object feature extraction. Additionally, the SPPF—LSKA (Spatial Pyramid Pooling Fast—Large Separable Kernel Attention) module is introduced to replace the SPPF (Spatial Pyramid Pooling Fast) module, aiming to enhance the algorithm's ability to extract key features. Experiments on the VisDrone2019 dataset show that the PSO-YOLO algorithm achieves an mAP@50% of 30.9% and an mAP@50–95% of 17.4%, representing improvements of 4.6% and 2.8%, respectively, over YOLOv8n.