摘要
针对大部分模型参数量较多,导致飞机遥感图像目标检测实时性差的问题,提出了一种基于GhostNet-YOLOv4的轻量化遥感图像中飞机目标检测模型。该模型取代YOLOv4的主干网络CSPDarknet53,并使用Ghost卷积取代原网络中的部分标准卷积,进一步减少参数量;使用改进后的空洞空间金字塔池化(Atrous Spatial Pyramid Pooling,ASPP)模块替换SPP模块,增大感受野范围;将ECA模块和SAM模块结合成新的注意力机制,并加入到YOLO Head之前,提高模型的检测能力。实验结果表明,与原始YOLOv4相比,改进后的轻量化YOLOv4模型大小缩减为65.2 MB,检测速率提高至134.23帧/秒,检测精度达到94.63%,满足遥感图像中飞机目标快速和高精度的检测要求。
To address the problem that most models have a large number of parameters,which leads to the poor real-time performance of aircraft target detection in remote sensing images,a lightweight remote sensing image target detection model based on GhostNet-YOLOv4 is proposed.The model replaces the backbone network CSPDarknet53 of YOLOv4,and uses Ghost convolution to replace part of the standard convolution in the original network to further reduce the number of parameters.The improved Atrous Spatial Pyramid Pooling(ASPP)module is used to replace the SPP module to increase the perceptual field range.Finally,the ECA module and SAM module are combined into a new attention mechanism and added in front of YOLO Head to improve the detection capability of the model.The experimental results show that compared with the original YOLOv4,the size of the improved lightweight YOLOv4 model is reduced to 65.2 MB,the detection rate is increased to 134.23 frame/s,and the detection accuracy reaches 94.63%,which meet the requirements of fast and high accuracy detection of aircraft targets in remote sensing images.
作者
张新蕊
邓超
ZHANG Xinrui;DENG Chao(School of Physics and Electronic Information Engineering,Henan Polytechnic University,Jiaozuo 454003,China)
出处
《无线电工程》
北大核心
2023年第7期1554-1562,共9页
Radio Engineering
基金
河南省高校基本科研业务费专项资金资助(NSFRF210427)
河南理工大学光电传感与智能测控河南省工程实验室开放课题(HELPSIMC-2020-002)
河南理工大学基本科研业务费基础研究项目(B类)(NSFRF230601)
河南省科技攻关项目(232102210100)。