Scholarlink

To address the critical challenges of weak feature representation, high computational complexity, and insufficient multi-scale adaptability in tiny object detection of high-resolution remote sensing images, this study proposes an improved lightweight model based on the YOLOv8n architecture. Firstly, an innovative GhostConv-Head module is designed, which establishes a collaborative mechanism of primary feature compression and phantom expansion through decoupled feature generation processes, significantly reducing model parameters and computational complexity while maintaining feature representation capability. Secondly, a multi-scale Slim-Neck feature fusion architecture with dynamic kernel size adaptation is proposed, employing a receptive field dynamic adjustment strategy to enhance detection sensitivity for tiny targets. Furthermore, a lightweight LiteCBAM dual-dimensional attention mechanism is developed, achieving balanced optimization between computational efficiency and feature focusing capability through dual-pooling feature compression strategy and decomposed spatial attention computation. Comparative experiments on the AI-TOD remote sensing dataset demonstrate that the improved model achieves a mean Average Precision (mAP) of 24.9%, representing a 2.8 percentage point improvement over the baseline YOLOv8n. The model size is compressed to 2.73 MB, meeting the real-time detection requirements of edge computing devices while preserving lightweight characteristics.

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

Lightweight Remote Sensing Tiny Object Detection Model Based on YOLOv8n Architecture

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Jinyin Bai,
  • Wei Zhu,
  • Qinglin Xu,
  • Xiangchen Wang,
  • Peng Zheng

摘要

To address the critical challenges of weak feature representation, high computational complexity, and insufficient multi-scale adaptability in tiny object detection of high-resolution remote sensing images, this study proposes an improved lightweight model based on the YOLOv8n architecture. Firstly, an innovative GhostConv-Head module is designed, which establishes a collaborative mechanism of primary feature compression and phantom expansion through decoupled feature generation processes, significantly reducing model parameters and computational complexity while maintaining feature representation capability. Secondly, a multi-scale Slim-Neck feature fusion architecture with dynamic kernel size adaptation is proposed, employing a receptive field dynamic adjustment strategy to enhance detection sensitivity for tiny targets. Furthermore, a lightweight LiteCBAM dual-dimensional attention mechanism is developed, achieving balanced optimization between computational efficiency and feature focusing capability through dual-pooling feature compression strategy and decomposed spatial attention computation. Comparative experiments on the AI-TOD remote sensing dataset demonstrate that the improved model achieves a mean Average Precision (mAP) of 24.9%, representing a 2.8 percentage point improvement over the baseline YOLOv8n. The model size is compressed to 2.73 MB, meeting the real-time detection requirements of edge computing devices while preserving lightweight characteristics.