摘要
针对遥感图像地物信息多、目标小、边界模糊、特征不明显,常规检测算法可能使目标被背景噪声淹没造成误检、漏检的问题,以YOLOv8算法为基础,对其主干网络进行改进,设计了新的空间金字塔池化模块ASPPF,强化主干网络的特征提取能力,弱化遥感图像复杂背景对算法的干扰。在改进网络的检测层结构增加一个小目标检测层,同时删除大目标检测层,引入双向特征金字塔网络的跨层连接和加权融合方法,构建一种新的多尺度特征融合网络MFPN,强化模型的特征融合能力,提升算法对小目标检测的能力。引入轻量级卷积GhostNet替换C2f模块的Bottleneck,在保证检测精度的基础上减少算法的参数量和计算量,从而加快算法的检测速度。在自建数据集上的对比实验结果表明,所设计的算法对遥感图像小目标检测具有较好的精确度和实时性。
In response to the problems of multiple land cover information,small targets,blurred boundaries,and unclear features in remote sensing images,conventional detection algorithms may cause false positives and false negatives due to the submergence of targets by background noise.Based on the YOLOv8 algorithm,the backbone network is improved and a new spatial pyramid pooling module ASPPF is designed to enhance the feature extraction ability of the backbone network and weaken the interference of complex backgrounds in remote sensing images on the algorithm.By improving the detection layer structure of the network and adding a small object detection layer while removing the large object detection layer,a new multi-scale feature fusion network MFPN is constructed by introducing the cross layer connection and weighted fusion method of bidirectional feature pyramid network.This strengthens the feature fusion ability of the model and enhances the algorithm's ability to detect small objects.Introducing lightweight convolution GhostNet to replace Bottleneck in C2f module,reducing the number of algorithm parameters and computation while ensuring detection accuracy,thereby accelerating the detection speed of the algorithm.The comparative experimental results on the self built dataset show that the designed algorithm has good accuracy and real-time performance for small object detection in remote sensing images.
作者
张高伦
郑永强
陶陶
ZHANG Gaoun;ZHENG Yongqiang;TAO Tao(IDC Business Department,Baosight Software(Anhui)Co.,Ltd.,Ma'anshan,Anhui 243002,China;School of Computer,Anhui University of Technology,Ma'anshan,Anhui 243002,China)
出处
《自动化应用》
2025年第16期63-65,68,共4页
Automation Application
基金
安徽省自然科学基金面上项目(1908085MF212)。
关键词
深度学习
小目标检测
特征融合
轻量化网络
deep learning
small object detection
feature fusion
lightweight network
