轻量级红外目标检测算法研究  

Research on Lightweight Infrared Target Detection Algorithm

作　　者：张上[1,2] 陈永麟 王恒涛 黄俊锋 ZHANG Shang;CHEN Yonglin;WANG Hengtao;HUANG Junfeng(Hubei Province Engineering Technology Research Center for Construction Quality Testing Equipment,China Three Gorges University,Yichang 443002,China;College of Computer and Information Technology,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：《无线电工程》2024年第11期2558-2565,共8页Radio Engineering

基　　金：国家级大学生创新创业训练计划(202011075013,202111075019)。

摘　　要：针对红外图像分辨率差、对比度低、信噪比低、视觉效果模糊等检测难点,提出一种基于YOLOv5的轻量化红外目标检测(Lightweight Infrared Target Detector-YOLO,LITD-YOLO)算法。LITD-YOLO算法重新设计网络结构,针对红外目标成像特征,将特征提取网络与特征融合网络结构重构。提高小感受野权重,重建浅层特征和深层特征之间的多尺度融合关系,提高浅层网络语义信息表征能力权重,增强对红外小目标的检测能力。引入Varifocal loss以实现交并比感知分类评分(Intersection over Union-Aware Classification Score, IACS)回归,使模型对密集目标的检测能力进一步加强。使用SIoU作为边框损失函数,用于提升预测框的准确度,同时加速模型收敛。实验结果表明,在FLIR和OSU数据集下模型检测精度分别提高至88.5%、99.7%,模型体积仅3.9 MB,参数量和算法复杂度大幅降低;与主流算法相比,LITD-YOLO在各项指标上均取得了不错的进步,在检测精度、模型体积和推理速度等方面具有先进性,能满足对红外目标的高质量检测。For poor resolution,low contrast,low signal to noise ratio,visual effect blurring and other detection difficulties of infrared images,a Lightweight Infrared Target Detection-YOLO(LITD-YOLO)algorithm based on YOLOv5 is proposed.Firstly,the LITD-YOLO algorithm redesigns the network structure and reconstructs the feature extraction network and feature fusion network structure for the infrared target imaging features.The small sensory field weight is increased to reconstruct the multi-scale fusion relationship between shallow and deep features,and the semantic information representation capability weight of the shallow network is increased to enhance the detection capability of small infrared targets;then,Varifocal loss is introduced to achieve Intersection over Union-Aware Classification Score(IACS)regression to further enhance the detection capability of the model for dense targets;finally,SIoU is used as the border loss function to improve the accuracy of the prediction frame while accelerating the convergence of the model.The experimental results show that the model detection accuracy improves to 88.5%and 99.7%for FLIR and OSU datasets respectively,and the model volume is only 3.9 MB,with a significant reduction in the number of parameters and algorithm complexity;compared with mainstream algorithms,LITD-YOLO has made good progress in all indicators,and is advanced in terms of detection accuracy,model volume and inference speed,and can meet the requirements for high quality detection of infrared targets.

关 键 词：目标检测 模型轻量化 YOLOv5 Varifocal loss SIoU 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]