Scholarlink

<p>UAV object detection tasks commonly face challenges such as dense small objects, complex multi-scale object distributions, and stringent real-time requirements for edge deployment. A lightweight and efficient improved algorithm LCS-YOLO based on YOLOv12n is proposed to solve these problems. First, Lightweight Re-parameterizable Convolution (LRConv) is proposed to enhance feature expressiveness without increasing inference overhead by decoupling the re-parameterization strategy between training and inference phases, thereby providing more stable underlying feature embeddings for small object detection. Secondly, the Center-guided Cross-scale Fusion Network (CCFN) is designed. By centering on mesoscale features, efficient interaction and semantic alignment among multi-scale features are achieved through centralized cross-scale fusion and center-guided reconstruction. Finally, the Shared Dynamic Aligned Detection Head (SDA-Head) is proposed. A unified and efficient prediction head architecture is constructed through cross-scale parameters sharing, task-decoupled modeling, and dynamic regression alignment mechanisms, multi-scale prediction consistency and regression stability are enhanced while reducing parameters redundancy. Multiple comparative experiments are conducted on the VisDrone2019, HIT-UAV, and CODrone datasets. Compared with YOLOv12n, LCS-YOLO reduces the parameters by 23.3% and improves FPS by 24, the mAP50 is improved by 3.6%, 5.5%, and 1.4% on the three datasets, respectively. These results validate the effectiveness, real-time performance, and deployment advantages of the proposed algorithm in UAV object detection scenarios.</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

LCS-YOLO: A lightweight and efficient small object detection algorithm for UAV

  • 期刊论文

  • 发表日期：

    2026年4月22日

查看全文
  • Xiuman Liang,
  • Yongsheng Zhang,
  • Nan Wu,
  • Haifeng Yu

摘要

UAV object detection tasks commonly face challenges such as dense small objects, complex multi-scale object distributions, and stringent real-time requirements for edge deployment. A lightweight and efficient improved algorithm LCS-YOLO based on YOLOv12n is proposed to solve these problems. First, Lightweight Re-parameterizable Convolution (LRConv) is proposed to enhance feature expressiveness without increasing inference overhead by decoupling the re-parameterization strategy between training and inference phases, thereby providing more stable underlying feature embeddings for small object detection. Secondly, the Center-guided Cross-scale Fusion Network (CCFN) is designed. By centering on mesoscale features, efficient interaction and semantic alignment among multi-scale features are achieved through centralized cross-scale fusion and center-guided reconstruction. Finally, the Shared Dynamic Aligned Detection Head (SDA-Head) is proposed. A unified and efficient prediction head architecture is constructed through cross-scale parameters sharing, task-decoupled modeling, and dynamic regression alignment mechanisms, multi-scale prediction consistency and regression stability are enhanced while reducing parameters redundancy. Multiple comparative experiments are conducted on the VisDrone2019, HIT-UAV, and CODrone datasets. Compared with YOLOv12n, LCS-YOLO reduces the parameters by 23.3% and improves FPS by 24, the mAP50 is improved by 3.6%, 5.5%, and 1.4% on the three datasets, respectively. These results validate the effectiveness, real-time performance, and deployment advantages of the proposed algorithm in UAV object detection scenarios.