计算机工程与应用 ›› 2024, Vol. 60 ›› Issue (18): 114-125.DOI: 10.3778/j.issn.1002-8331.2404-0058

• YOLOv8 改进及应用专题 • 上一篇    下一篇

改进YOLOv8的轻量级军事飞机检测算法

刘丽,张硕,白宇昂,李宇健,张初夏   

  1. 1.华北电力大学 计算机系,河北 保定 071000
    2.复杂能源系统智能计算教育部工程研究中心,河北 保定 071000
  • 出版日期:2024-09-15 发布日期:2024-09-13

Improved Lightweight Military Aircraft Detection Algorithm of YOLOv8

LIU Li, ZHANG Shuo, BAI Yu’ang, LI Yujian, ZHANG Chuxia   

  1. 1.Department of Computer Science, North China Electric Power University, Baoding, Hebei 071000, China
    2.Engineering Research Center of Intelligent Computing for Complex Energy Systems, Ministry of Education, Baoding, Hebei 071000, China
  • Online:2024-09-15 Published:2024-09-13

摘要: 遥感图像军事飞机检测在侦察预警、情报分析等领域具有重要意义。为使军事飞机检测模型能在算力受限的设备上高效运行,从网络设计与模型压缩两个方面对YOLOv8n进行轻量化改进。在网络设计方面,使用FAS_C2f替换原始主干网络中的C2f模块,减少计算冗余并加快网络特征提取的速度;根据军事飞机目标的尺度特征对网络结构进行优化,缓解因过度下采样导致的小目标信息丢失问题;使用Inner-SIoU作为新的定位回归损失函数,提升对小目标样本的学习能力并加快回归边界框的收敛。在模型压缩方面,使用基于LAMP分数的通道剪枝对重设计后的模型进行压缩,进一步减少参数和模型大小;并利用通道级知识蒸馏(channel-wise knowledge distillation,CWD)将模型精度恢复到接近剪枝前的水平。实验结果表明,在公开军用飞机数据集MAR20上,轻量化后的模型mAP为97.2%,体积仅有0.7 MB,较原始模型缩小了88.3%,FPS提高了14帧/s,满足军事飞机目标检测的实时性要求。

关键词: 目标检测, 军事飞机, YOLOv8, 模型剪枝, 知识蒸馏

Abstract: Military aircraft detection with remote sensing images is of great significance in the fields of reconnaissance and early warning, intelligence analysis and so on. In order to make the military aircraft inspection model run efficiently on the equipment with limited computing power, the lightweight improvement of YOLOv8n is carried out from two aspects: network design and model compression. In the aspect of network design, firstly, FAS_C2f is used to replace the C2f module in the original backbone network, which reduces the computational redundancy and speeds up the network feature extraction. Secondly, the network structure is optimized according to the scale characteristics of military aircraft targets to alleviate the problem of small target information loss caused by excessive downsampling. Thirdly, Inner-SIoU is used as a new localization regression loss function to improve the learning ability of small target samples and accelerate the convergence of regression bounding box. In terms of model compression, channel pruning based on LAMP fraction is used to compress the redesigned model to further reduce parameters and model size. With channel-wise knowledge distillation (CWD), the accuracy of the model is restored to the level close to that before pruning. The experimental results show that on the open military aircraft data set MAR20, the mAP of the lightweight model is 97.2%, the volume is only 0.7 MB, which is 88.3% smaller than the original model, and the FPS is increased by 14 frames per second, which meets the real-time requirements of military aircraft target detection.

Key words: object detection, military aircraft, YOLOv8, model pruning, knowledge distillation