Scholarlink

<p>At present, applying object detection models to complex and unknown aerial drone scenes is a challenge. The model not only needs to consider detection accuracy but also needs to balance computation speed and size parameters on low-power devices. To address these challenges, we propose a lightweight object detection model, LAO-YOLO, based on YOLOv10s. Firstly, we propose a new C2f structure, C2f-FRFC, which introduces the newly designed bottleneck structure SCRB (Spatial and Channel Reuse Bottleneck), achieving light-weighting of the backbone. To improve the performance of small object detection in aerial photography scenes, we adopt three strategies to improve the neck. In addition, we introduce a EMASlideLoss(Exponential Moving Average SlideLoss) to solve the problem of imbalanced sample classification. Finally, we adopt the LAMP(Layer-Adaptive Magnitude-based Pruning) strategy to maximize the size compression of the model while considering its accuracy. Experiments on VisdroneDET-2021 showed that compared to YOLOv10s, LAO-YOLO parameter count decreased by 55.9%, floating-point operations decreased by 61.2%, model size decreased by 56.3%, and mAP50 increased by 0.3%. The results show that the proposed method achieves precise detection while reducing model size, meeting the portability requirements for deployment on edge mobile devices.</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

Lao-Yolo: improved YOLOv10 model for lightweight aerial object detection

  • 期刊论文

  • 发表日期：

    2025年4月2日

查看全文
  • ZhiLin Gao,
  • QiXiang Meng,
  • JinTao Wang,
  • FanLiang Bu

摘要

At present, applying object detection models to complex and unknown aerial drone scenes is a challenge. The model not only needs to consider detection accuracy but also needs to balance computation speed and size parameters on low-power devices. To address these challenges, we propose a lightweight object detection model, LAO-YOLO, based on YOLOv10s. Firstly, we propose a new C2f structure, C2f-FRFC, which introduces the newly designed bottleneck structure SCRB (Spatial and Channel Reuse Bottleneck), achieving light-weighting of the backbone. To improve the performance of small object detection in aerial photography scenes, we adopt three strategies to improve the neck. In addition, we introduce a EMASlideLoss(Exponential Moving Average SlideLoss) to solve the problem of imbalanced sample classification. Finally, we adopt the LAMP(Layer-Adaptive Magnitude-based Pruning) strategy to maximize the size compression of the model while considering its accuracy. Experiments on VisdroneDET-2021 showed that compared to YOLOv10s, LAO-YOLO parameter count decreased by 55.9%, floating-point operations decreased by 61.2%, model size decreased by 56.3%, and mAP50 increased by 0.3%. The results show that the proposed method achieves precise detection while reducing model size, meeting the portability requirements for deployment on edge mobile devices.