YOLO-IRLight:基于YOLOv8的轻量级无人机红外小目标检测算法  

YOLO-IRLight:Lightweight UAV Infrared Small Target Detection Algorithm Based on YOLOv8

作　　者：倪梦琪 陈凯源 NI Mengqi;CHEN Kaiyuan(North China University of Water Resources and Electric Power,Zhengzhou 450046,China)

机构地区：[1]华北水利水电大学,河南郑州450046

出　　处：《现代信息科技》2025年第8期46-53,60,共9页Modern Information Technology

摘　　要：针对无人机航拍角度下红外小目标检测精度低、计算量大等问题,基于YOLOv8s提出了一种轻量化红外小目标检测模型YOLO-IRLight。此模型引入EMA(Efficient Multiscale Attention)注意力机制来提升特征提取能力;在网络颈部添加PConv-C2f模块减少计算量,将尺度序列特征融合,并添加P2检测层对网络结构进行优化,增强了小目标检测能力;设计了一种新型轻量化检测头Group-Detect,将NWD(Normalized Gaussian Wasserstein Distance)损失函数以线性组合方式融入模型的损失函数中。公开数据集上的实验结果表明与原始YOLOv8s相比,该模型的检测精度mAP@0.5提高了1.7%,参数量下降了45.9%,计算复杂度GFLOPs下降了33.5%,F1精度提高了0.9%。改进后的算法明显优于传统算法,相较于其他当前主流算法,检测精度也有明显提升。To address the issues of low detection accuracy and high computational load in infrared small target detection from UAV aerial perspectives,a lightweight infrared small target detection model,YOLO-IRLight,is proposed based on YOLOv8s.This model introduces the EMA(Efficient Multiscale Attention)Attention Mechanism to enhance feature extraction capabilities.A PConv-C2f module is added to the neck of the network to reduce computational load and fuse scale sequence features,and a P2 detection layer is incorporated to optimize the network structure,thereby improving small target detection performance.A novel lightweight detection head,Group-Detect,is designed,and the NWD(Normalized Gaussian Wasserstein Distance)loss function is incorporated into the loss function of the model in a linear combination manner.Experimental results on the open dataset show that compared to the original YOLOv8s,the proposed model improves detection accuracy(mAP@0.5)by 1.7%,reduces the number of parameters by 45.9%,decreases computational complexity(GFLOPs)by 33.5%,and increases F1 score by 0.9%.The improved algorithm significantly outperforms traditional algorithms,with notable improvements in detection accuracy compared to current mainstream algorithms.

关 键 词：小目标检测 红外目标 轻量化 YOLOv8 网络优化 

分 类 号：TP3191.4[自动化与计算机技术—计算机软件与理论] TP301.6[自动化与计算机技术—计算机科学与技术]