融合全局信息与双域注意力机制的光学遥感飞机目标检测  

Integrating global information and dual-domain attention mechanism for optical remote sensing aircraft target detection

作　　者：林珊玲 张雪[1,2] 陈燕 林坚普 吕珊红 林志贤 郭太良[2,3] LIN Shanling;ZHANG Xue;CHEN Yan;LIN Jianpu;LÜShanhong;LIN Zhixian;GUO Tailiang(School of Advanced Manufacturing,Fuzhou University,Quanzhou 362251,China;Fujian Science and Technology Innovation Laboratory for Photoelectric Information,Fuzhou 350116,China;College of Physics and Information Engineering,Fuzhou University,Fuzhou 350116,China)

机构地区：[1]福州大学先进制造学院,福建泉州362251 [2]中国福建光电信息科学与技术实验室,福建福州350116 [3]福州大学物理与信息工程学院,福建福州350116

出　　处：《光学精密工程》2024年第20期3085-3098,共14页Optics and Precision Engineering

基　　金：国家重点研发计划(No.2022YFB3603705);国家自然科学基金青年基金(No.62101132)。

摘　　要：针对光学遥感图像中飞机目标检测算法因背景复杂、受测飞机目标较小以及飞机外观差异较小导致检测精度不足的问题,本文基于YOLOv8n模型,提出一种融合全局信息与双域注意力机制的光学遥感图像飞机目标检测算法。首先,设计了SPPF_Global模块,通过全局最大池化层提供全局特征概览,帮助模型在复杂环境中更好地区分目标与背景;其次,提出了双域注意力机制,通过空间域和通道域的信息引导,提高模型对机翼形状等重要区域的关注度,增强对不同飞机型号的细微差别的区分能力;最后,采用并行路径改进下采样模块并引入Powerful-IoU损失函数,通过自适应惩罚因子加速模型收敛,提高对小目标飞机的识别能力和预测框的回归效率。实验结果表明:与原始YOLOv8n相比,改进后的模型在公开数据集MAR20上的精确率、召回率、mAP50以及mAP50-95分别提高了3.3%,2.6%,3.2%和2.6%,在NWPU VHR-10数据集上分别提高了5%,5.1%,2.5%和0.3%;同时,参数量和运算量分别降低了6.6%和3.7%。证明了本文模型的高效性和优越性,有效提升了光学遥感图像中飞机目标检测算法的应用价值。To address the problem of insufficient detection accuracy of aircraft targets in optical remote sensing images due to complex backgrounds,small targets,and similar appearances among aircraft,an aircraft target detection algorithm was proposed in this paper based on the YOLOv8n model that integrated the global information and the dual-domain attention mechanism in optical remote sensing images.Firstly,the SPPF_Global module was designed to provide a global feature overview through the global maximum pooling layer,which helped the model better distinguish objects from the background in complex environments.Secondly,a dual-domain attention mechanism was proposed to improve the attention to important areas such as wing shape and other distinctive structures through the information guidance of space domain and channel domain,and enhanced the ability to distinguish the nuances of different aircraft models.Finally,the parallel path downsampling method and the Powerful-IoU loss function was introduced,and the adaptive penalty factor was used to accelerate the convergence of the model,which improved the recognition ability of the model for small target aircraft and the regression efficiency of the prediction frame.The experimental results show that compared with the original YOLOv8n,the accuracy rate,recall rate,mAP50 and MAP50-95 of the proposed model on the open data set MAR20 are increased by 3.3%,2.6%,3.2%and 2.6%respectively.On the NWPU VHR-10 dataset,the parameters are increased by 5%,5.1%,2.5%and 0.3%respectively,while the number of parameters and the calculation amount are decreased by 6.6%and 3.7%respectively,which proves the efficiency and superiority of the proposed model,and effectively improves the application value of the aircraft target detection algorithm in optical remote sensing images.

关 键 词：光学遥感图像 飞机目标检测 YOLOv8 注意力机制 Powerful-IoU 

分 类 号：TP394.1[自动化与计算机技术—计算机应用技术]