面向战场低空目标的轻量化识别方法  

A Lightweight Recognition Method for Low Altitude Targets in the Battlefield

作　　者：徐图南 高昂[1] 陈昱成 闫守成 邓斌 XU Tunan;GAO Ang;CHEN Yucheng;YAN Shoucheng;DENG Bin(School of Electronics And Information,Northwestern Polytechnical University,Xi'an 710072,Shaanxi,China;Faculty of Electronic and Information Engineering,Xi'an Jiaotong University,Xi'an 710049,Shaanxi,China;Key Laboratory of Near Ground Detection Technology,Wuxi 214035,Jiangsu China)

机构地区：[1]西北工业大学电子信息学院,陕西西安710072 [2]西安交通大学电子与信息学部,陕西西安710049 [3]近地面探测技术重点实验室,江苏无锡214035

出　　处：《兵工学报》2025年第2期124-134,共11页Acta Armamentarii

基　　金：近地面探测技术重点实验室基金(6142414220406);陕西省重点研发计划项目(2021GXLH-01-15);太仓市重点研发计划项目(TC2019SF03)。

摘　　要：针对在战场目标识别过程中存在的图像较少、对比度低、资源受限等问题,提出一种名为YOLOF的战场低空目标识别算法。该算法基于YOLOv5s网络模型,使用循环对抗网络应用于图像增强;通过融合RepVGG模块和SiLU激活函数,依靠结构重参数化和更高效的激活函数,提升了模型的特征提取能力和推理速度;再以基于滤波器重要性的剪枝算法,精确评估和删除权重影响较低的滤波器,优化了模型的结构,提升了计算和存储效率;最后通过基于特征层次的知识蒸馏方法,使教师模型向学生模型的特征层进行知识传递,保持了剪枝后模型的高性能。实验结果表明,所提的YOLOF算法相较初始YOLOv5s算法,可以在保证高精度目标识别的同时,实现网络结构的轻量化,即参数量仅有3.6×10^(6),平均精度均值在自制数据集中达到86.3%,已满足战场低空目标的识别要求。In the process of intelligent radar target detection,the real-time recognition of targets by edge devices is particularly important.Considering the resource constraints inherent in embedded devices,the need for lightweight target recognition networks has become increasingly prominent.To address the challenges of limited battlefield target images and low contrast,a battlefield low-altitude target recognition algorithm named YOLOF is proposed.This algorithm is based on the YOLOv5s network model and incorporates a cycle-generative adversarial network(CycleGAN)for image enhancement.By integrating the RepVGG module and SiLU activation function,the algorithm enhances the feature extraction capability and inference speed of model through structural reparameterization and more efficient activation functions.Additionally,a pruning algorithm based on filter importance is employed to accurately evaluate and remove the filters with low weight impact,thereby optimizing the model's structure and improving the computational and storage efficiencies.Furthermore,the knowledge distillation based on feature layers allows the transfer of knowledge from the teacher model to the student model's feature layers,thus maintaining the high performance of the pruned model.Experimental results demonstrate that the proposed YOLOF algorithm,compared to the original YOLOv5s algorithm,achieves network lightweighting while ensuring high-precision target recognition.Specifically,the parameter count is reduced to just 3.6×10^(6),and the mean average precision(mAP)reaches 86.3%on a custom dataset,meeting the requirements for battlefield low-altitude target recognition.

关 键 词：目标识别 YOLOv5s网络 剪枝 网络轻量化 知识蒸馏 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]