基于改进Yolov5s的迷彩伪装目标检测  

Research on Camouflage Target Detection Based on Improved Yolov5S

作　　者：杨凡 张洁[1] YANG Fan;ZHANG Jie(Nanjing University of Posts and Telecommunications,Nanjing 210046,China)

机构地区：[1]南京邮电大学,南京210046

出　　处：《火力与指挥控制》2025年第2期148-155,共8页Fire Control & Command Control

基　　金：国家重点研发计划资助项目(2018YFB1500902)。

摘　　要：伪装目标检测在军事作战中扮演了关键角色,针对现有迷彩伪装目标检测方法精度不高、漏检误检等问题,提出一种改进的Yolov5s算法。在特征提取网络中融合SE注意力,选择性增强目标的关键信息,抑制背景环境的干扰因素。引入SPPFCSPC替换原池化方式,生成多尺度特征,在加快速度的同时增加检测精度。采用双立方插值取代最近邻插值,以减少上采样过程中丢失的图像细节。实验结果显示,改进的算法在一种公开的迷彩伪装数据集上的mAP、Recall分别达到96.9%和93.8%,较当前研究有显著的提升。Camouflage target detection plays a key role in military operations,and an improved Yolov5s algorithm is proposed to address the problems of low accuracy and missed and false detection of existing camouflage target detection methods.Firstly,SE attention is fused in the feature extraction network to selectively enhance the key information of the target and suppress the interference factors of the background environment.Secondly,SPPFCSPC is introduced to replace the original pooling method to generate multi-scale features,which increases the detection accuracy while speeding up.Finally,bi-cubic interpolation is used instead of nearest-neighbor interpolation to reduce the image details lost during upsampling process.The experimental results show that mAP、Recall on the open camouflage dataset of the improved algorithm reach 96.9% and 93.8% respectively,there is a remarkable improvement compared to the current study.

关 键 词：迷彩伪装士兵 目标检测 Yolov5s算法 双立方插值 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]