低空轻量级红外弱小目标检测算法  

Low altitude lightweight infrared weak small targetdetection algorithm

作　　者：张上[1,2] 黄俊锋 王恒涛 陈永麟 王康 ZHANG Shang;HUANG Jun-feng;WANG Heng-tao;CHEN Yong-lin;WANG Kang(Hubei Province Engineering Technology Research Center for Construction Quality Testing Equipment,China Three Gorges University,Yichang 443002,China;College of Computer and Information,China Three Gorges University,Yichang 443002,China)

机构地区：[1]三峡大学湖北省建筑质量检测装备工程技术研究中心,湖北宜昌443002 [2]三峡大学计算机与信息学院,湖北宜昌443002

出　　处：《激光与红外》2024年第1期122-129,共8页Laser & Infrared

基　　金：国家级大学生创新创业训练计划项目(No.202111075019,No.202011075013)资助。

摘　　要：精准的红外弱小目标检测是实时监控、追踪、制导的关键;红外弱小目标存在检测难度高、误检高、漏检严重的问题。为了提高红外弱小目标检测算法的实时性和检测精度,提出了一种超轻量红外弱小目标检测算法SL-YOLO。首先,重设计下采样方案,针对红外图像特征信息调节网络架构,解决红外弱小目标特征梯度降低和特征消失问题;然后设计网络模型剪枝算法,实现剪枝算法与网络结构的融合,去除冗余参数,实现检测速度的提高;最后设计Varifocal-SIoU损失函数,在均衡正负样本与重叠损失的同时,对正样本进行加权处理,解决背景干扰问题。实验结果表明,在SIRST和IDSAT数据集下检测精度分别提高至96.4%、98.1%,模型体积和计算量可压缩至190 kB、0.9 GFLOPs,推理速度降至3 ms以下。与主流算法进行对比,改进后算法在检测精度、模型体积、计算量等方面均取得了不错的成绩。能够满足实时性检测需求。Accurate infrared small and weak target detection is the key to real-time monitoring,tracking,and guidance.Infrared weak and small targets have problems of high detection difficulty,high false detection,and serious missed detection.In this paper,an ultra-lightweight infrared dim small target detection algorithm SL-YOLO is proposed to improve the real-time performance and detection accuracy of infrared dim small target detection algorithms.Firstly,the downsampling scheme is redesigned to adjust the network architecture for the infrared image feature information to solve the problem of feature gradient reduction and feature disappearance for infrared weak targets.Then,a network model pruning algorithm is designed to integrate pruning algorithm with network structure,removing redundant parameters,and improving detection speeds.Finally,the SIoU Varifocal loss function is designed to equalise the positive and negative samples with overlapping losses while weighting the positive samples to solve the problem of background interference.The experimental results show that the detection accuracy is improved to 96.4%and 98.1%under the SIRST and IDSAT datasets,respectively.The model volume and computational complexity can be compressed to 190 kB and 0.9 GFLOPs,and the inference speed is reduced to less than 3 ms.Comparing with the mainstream algorithms,the improved algorithm has achieved good results in terms of detection accuracy,model volume;computational complexity.It can meet the real-time detection requirements.

关 键 词：目标检测 模型剪枝 YOLOv5 SIoU Varifocal loss 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TN219[自动化与计算机技术—计算机科学与技术]