基线网络轻量化的无人机小目标检测算法
CSTR:
作者:
作者单位:

(1.新疆大学 电气工程学院,乌鲁木齐 830049;2.新疆大学 智能科学与技术学院,乌鲁木齐 830049)

作者简介:

杨超(1999—),男,硕士研究生;张宏立(1972—),男,教授,博士生导师

通讯作者:

李新凯,lxk@xju.edu.cn

中图分类号:

TP391

基金项目:

国家自然科学基金(62263030);新疆维吾尔自治区自然科学基金(2023D01C187)


Small target detection algorithm for unmanned aerial vehicles based on baseline network lightweighting
Author:
Affiliation:

(1.School of Electrical Engineering, Xinjiang University, Urumqi 830049, China; 2.School of Intelligent Science and Technology, Xinjiang University, Urumqi 830049, China)

Fund Project:

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    为解决无人机端侧小目标检测中像素低、背景复杂、目标密集且算力受限的矛盾,提出基于改进YOLOv8s的轻量化基线(you only look once v8 small-lightweight,YOLOv8s-LW)算法。首先,利用幽灵卷积与深度卷积优化的高性能图形处理器网络取代YOLOv8s主干网络,通过参数共享与结构裁剪降低冗余,增强端到端可训练性;其次,在颈部以幽灵卷积增强型特征融合模块和快速块高效多尺度注意力融合模块替换跨阶段双卷积瓶颈模块,引入空间混合与多尺度注意力加权,强化多尺度特征聚合与泛化能力;再次,以焦点形状约束交并比损失函数取代原损失函数,从形状约束与匹配精度多维度优化小目标回归鲁棒性;最后,融合动态检测头,提升复杂背景下的小目标响应与定位一致性。实验结果表明:在VisDrone-2019数据集上,YOLOv8s-LW算法的mAP@0.5达到38.5%,参数量为6.4×106,计算量为19.2×109,推理速度为64.0帧/s;相较于基线YOLOv8s,参数量下降约43%,计算量下降约33%,且精度提升0.2%。所提算法在端侧无人机小目标检测中实现了低算力、高实时、稳精度的可用均衡,形成可直接部署的轻量化基线,并为复杂背景场景下的稳健检测与后续多任务扩展提供工程化支持与可迁移的结构设计范式。

    Abstract:

    To alleviate the conflicts of low pixel resolution, complex backgrounds, dense targets, and limited compute in unmanned aerial vehicle (UAV) edge detection for small targets, this paper proposed a lightweight baseline algorithm based on the improved YOLOv8s, namely you only look once v8 small lightweight (YOLOv8s-LW). First, the algorithm replaced the backbone network of YOLOv8s with a high-performance graphics processing unit network optimized with ghost and depthwise convolutions, reducing redundancy through parameter sharing and structural pruning while enhancing end-to-end trainability. Second, it replaced the module of cross-stage partial bottleneck with two convolutions with the ghost-convolution-augmented feature fusion module and the fast block-based efficient multi-scale attention fusion module, which introduced spatial mixing and multi-scale attention weighting to strengthen multi-scale feature aggregation and generalization. Subsequently, the algorithm employed a focaler-shape-wise intersection over union loss function instead of the original loss function to improve the robustness of small target regression from the perspectives of shape constraint and matching accuracy. Finally, it integrated a dynamic detection head to enhance small target responses and localization consistency under complex backgrounds. Experiments on the VisDrone-2019 dataset show that YOLOv8s-LW achieves 38.5% mAP@0.5 with 6.4×106 parameters, 19.2×109 computations, and an inference speed of 64.0 frames per second. Compared with the baseline YOLOv8s, parameters and computation are reduced by about 43% and 33%, respectively, while accuracy is improved by 0.2%. The proposed algorithm achieves a favorable trade-off among computational cost, real-time performance, and accuracy for UAV edge detection for small targets and provides engineering support and a deployable lightweight baseline and transferable architectural design for robust detection and multi-task extensions in complex background scenarios.

    参考文献
    相似文献
    引证文献
引用本文

杨超,李新凯,张宏立,孟月,王聪.基线网络轻量化的无人机小目标检测算法[J].哈尔滨工业大学学报,2026,58(9):70. DOI:10.11918/202412071

复制
分享
相关视频

文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2024-12-24
  • 最后修改日期:
  • 录用日期:
  • 在线发布日期: 2026-10-11
  • 出版日期:
文章二维码