计算机工程与应用 ›› 2024, Vol. 60 ›› Issue (8): 182-191.DOI: 10.3778/j.issn.1002-8331.2310-0063

• 图形图像处理 • 上一篇    下一篇

改进YOLOv8的轻量化无人机目标检测算法

胡峻峰,李柏聪,朱昊,黄晓文   

  1. 东北林业大学 计算机与控制工程学院,哈尔滨 150036
  • 出版日期:2024-04-15 发布日期:2024-04-15

Improved YOLOv8 Lightweight UAV Target Detection Algorithm

HU Junfeng, LI Baicong, ZHU Hao, HUANG Xiaowen   

  1. College of Computer and Control Engineering, Northeast Forestry University, Harbin 150036, China
  • Online:2024-04-15 Published:2024-04-15

摘要: 针对无人机目标检测算法计算复杂难以部署,且长尾分布的无人机数据导致检测精度较低的问题,提出了基于改进YOLOv8的轻量化无人机目标检测算法(PC-YOLOv8-n),可均衡网络检测精度与计算量,并对长尾分布数据有一定泛化能力。使用部分卷积层(PConv)替换YOLOv8中的3×3卷积层,对网络进行轻量化处理,解决网络冗余和计算量复杂的问题;融合双通道特征金字塔,增加自上而下的路径,将深层信息与浅层信息进行融合,同层引入轻量化注意力机制,提升网络的特征提取能力;采用均衡焦点损失(EFL)作为类别损失函数,通过均衡尾部类别在网络训练时的梯度权重,增加网络的类别检测能力。实验结果表明,PC-YOLOv8-n在VisDrone2019数据集中具有良好的表现,在mAP50精度上比原始YOLOv8-n算法提高了1.6个百分点,同时模型的参数和计算量分别降低为2.6×106和7.6 GFLOPs,检测速度达到77.2 FPS。

关键词: 无人机, YOLOv8, 长尾分布, 目标检测, 部分卷积

Abstract: Aiming at the problem that UAV target detection algorithms are computationally complex and difficult to deploy, and the long-tailed distribution of UAV data leads to low detection accuracy, a lightweight UAV target detection algorithm based on improved YOLOv8 (PC-YOLOv8-n) is proposed, which can balance the network detection accuracy and computation, and has some generalisation ability to long-tailed distribution of data. Using partial convolutional layers (PConv) to replace the 3×3 convolutional layers in YOLOv8, the network is lightweighted to solve the problems of network redundancy and computational complexity; it fuses dual-channel feature pyramids, increases top-down paths, fusion of deep and shallow information, and introduces a lightweight attention mechanism in the same layer to improve the feature extraction ability of the network; it uses the equilibrium focus loss (EFL) as the category loss function to increase the category detection ability of the network by equalising the gradient weights of the tail categories during network training. The experimental results show that PC-YOLOv8-n has good performance in the VisDrone2019 dataset, improving 1.6 percentage points in mAP50 accuracy over the original YOLOv8-n algorithm, while the parameters and computation of the model are reduced to 2.6×106 and 7.6 GFLOPs, respectively, and the detection speed reaches 77.2 FPS.

Key words: unmanned aerial vehicle (UAV), YOLOv8, long-tail distribution, object detection, partial convolution