面向遥感影像的轻量级卷积神经网络目标检测  

Lightweight Convolutional Neural Network Object Detection for Remote Sensing Images

作　　者：谭海英 杨军 TAN Haiying;YANG Jun(Faculty of Geomatics,Lanzhou Jiaotong University,Lanzhou 730070,China;School of Electronic and Information Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China)

机构地区：[1]兰州交通大学测绘与地理信息学院,甘肃兰州730070 [2]兰州交通大学电子与信息工程学院,甘肃兰州730070

出　　处：《遥感技术与应用》2025年第1期167-176,共10页Remote Sensing Technology and Application

基　　金：国家自然科学基金项目(42261067)。

摘　　要：针对背景复杂、目标尺度差异性大且分布不均的高分辨率遥感影像,现有算法无法同时兼顾检测精度与检测速度等问题,以YOLO v7为基础提出了一种局部—全局检测器(Local-Global Detector,LGDet)的轻量化目标检测网络。首先,采用部分卷积(Partial Convolution,PConv)对其主干和颈部网络进行压缩,减少模型的参数量与运算量;其次,设计了快速傅里叶联合部分卷积模块,构建了局部和全局感受野结合的轻量化特征提取网络;最后,提出了一种轻量化三元注意力模块,增强了有用特征。在RSOD和NWPU VHR-10数据集上进行了实验,本算法的mAP分别为93.4%和90.5%,FLOPs为87.9 G,与YOLO v7相比mAP分别提高了2.8%和2.3%,FLOPs减少17.3 G。结果表明本算法在计算复杂度更低的情况下具有更优异的遥感影像目标检测精度。Aiming at the problems of complex backgrounds,large-scale variations,and uneven distributions in high-resolution remote sensing images,as well as the trade-off between detection accuracy and speed in exist⁃ing algorithms,a lightweight object detection network called Local-Global Detector(LGDet)is proposed based on YOLO v7.Firstly,the main backbone and neck network were compressed using Partial Convolution(PConv)to reduce the model's parameters and computational complexity.Secondly,a lightweight feature ex⁃traction network was constructed by designing a fast Fourier and partial convolution block that combines local and global receptive fields.Lastly,a lightweight triplet attention module was proposed,which enhances useful features.Experimental results on the RSOD and NWPU VHR-10 datasets demonstrate that the proposed algo⁃rithm achieves mAP of 93.4%and 90.5%respectively,with FLOPs of 87.9 G.Compared with YOLO v7,the mAP is improved by 2.8%and 2.3%respectively,while reducing FLOPs by 17.3 G.These findings demon⁃strate that the proposed algorithm achieves superior remote sensing object detection accuracy under lower com⁃putational complexity.

关 键 词：遥感影像 轻量化 目标检测 注意力机制 YOLO v7 

分 类 号：TP75[自动化与计算机技术—检测技术与自动化装置]