Article(id=1149773874725871618, tenantId=1146029695717560320, journalId=1146123166801305609, issueId=1149773869357167407, articleNumber=null, orderNo=null, doi=10.12404/j.issn.1671-1815.2404266, pmid=null, cstr=null, oa=null, hot=null, price=null, onlineType=0, articleFormat=0, articleType=null, articleTypeStr=null, receivedDate=1717689600000, receivedDateStr=2024-06-07, revisedDate=1736956800000, revisedDateStr=2025-01-16, acceptedDate=null, acceptedDateStr=null, onlineDate=1752057053500, onlineDateStr=2025-07-09, pubDate=1746633600000, pubDateStr=2025-05-08, doiRegisterDate=null, doiRegisterDateStr=null, onlineIssueDate=1752057053500, onlineIssueDateStr=2025-07-09, onlineJustAcceptDate=null, onlineJustAcceptDateStr=null, onlineFirstDate=null, onlineFirstDateStr=null, sourceXml=null, magXml=null, createTime=1752057053500, creator=13701087609, updateTime=1752057053500, updator=13701087609, issue=Issue{id=1149773869357167407, tenantId=1146029695717560320, journalId=1146123166801305609, year='2025', volume='25', issue='13', pageStart='5273', pageEnd='5704', issueExtLink='null', onlineDate='null', pubDate='null', beforeIssueId=null, nextIssueId=null, price=null, status=1, issueComplete=1, articleOrder=1, issueType=-1, specialIssue=0, createTime=1752057052207, creator=13701087609, updateTime=1768456769392, updator=13701087609, preIssue=null, nextIssue=null, ext={EN=IssueExt(id=1218559268744253990, tenantId=1146029695717560320, journalId=1146123166801305609, issueId=1149773869357167407, language=EN, specialIssueTitle=, coverIllustrator=, specialIssueEditor=, specialIssueAbout=), CN=IssueExt(id=1218559268744253991, tenantId=1146029695717560320, journalId=1146123166801305609, issueId=1149773869357167407, language=CN, specialIssueTitle=, coverIllustrator=, specialIssueEditor=, specialIssueAbout=)}, issueFiles=null}, startPage=5527, endPage=5534, ext={EN=ArticleExt(id=1149773875606675465, articleId=1149773874725871618, tenantId=1146029695717560320, journalId=1146123166801305609, language=EN, title=Small Target Detection Algorithm in Aerial Images Based on Improved RT-DETR, columnId=1156262729162810294, journalTitle=Science Technology and Engineering, columnName=Papers·Automation and Computational Technology, runingTitle=null, highlight=null, articleAbstract=

An algorithm has been proposed to detect small targets in unmanned aerial vehicle(UAV) aerial images. The algorithm is based on an improved real-time detection Transformer (RT-DETR) and aims to address the challenges posed by complex backgrounds and a large number of small target samples. To enhance the feature fusion network, a dedicated feature fusion structure for small targets has been incorporated, utilizing rich location information from the shallow feature map to improve the network's ability to detect small targets. Furthermore, the last residual block in the BackBone has been removed to prevent an increase in additional parameters. Additionally, the MCP Block, a reconstructed BasicBlock structure in the backbone network, has been designed, which includes a multi-channel feature partial convolution module (MCPConv) to reduce redundancy in channel features and enhance the acquisition of multi-scale detail features. Moreover, a location encoding mechanism with learning ability has been introduced to obtain more accurate and expressive location information. The normalized weighted deviation(NWD) and mean precision-driven IoU(MPDIoU) positioning loss functions have been incorporated to accelerate the convergence speed of the model and reduce sensitivity to position deviation. Experimental results on the VisDrone2019-DET dataset demonstrate that the improved model reduces parameters by 62% compared to the original model, increases mAP50 by 3.9%, and improves FPS by 17%. The improved model exhibits superior detection performance compared to other mainstream detection models.

, correspAuthors=Wei-dong LI, authorNote=null, correspAuthorsNote=null, copyrightStatement=null, copyrightOwner=null, extLink=null, articleAbsUrl=null, sourceXml=null, magXml=null, pdfUrl=null, pdf=null, pdfFileSize=null, pdfExtLink=null, richHtmlUrl=null, mobilePdfUrl=null, reviewReport=null, pdfFirstPage=null, abstractGraph=null, abstractGraphContent=null, abstractVideo=null, citation=null, cebUrl=null, magXmlContent=null, mapNumber=null, authorCompany=null, fund=null, authors=null, authorsList=Zi-qi ZHAO, Wei-dong LI, Xiao-juan LI), CN=ArticleExt(id=1149773893034008895, articleId=1149773874725871618, tenantId=1146029695717560320, journalId=1146123166801305609, language=CN, title=基于改进实时Transformer的航拍图像小目标检测算法, columnId=1156262729783567290, journalTitle=科学技术与工程, columnName=论文·自动化技术、计算机技术, runingTitle=null, highlight=null, articleAbstract=

针对无人机航拍图像中背景复杂、小目标样本多,难以提取有效特征等问题,提出一种改进实时Transformer(real-time detection Transformer,RT-DETR)的无人机航拍小目标检测算法。首先,在特征融合网络中增加针对微小目标的特征融合结构,利用浅层特征图中丰富的位置信息来增强网络对小目标的检测能力,同时为了防止额外参数的增加,去除主干网络中最后一个残差结构;其次,设计一种多通道特征部分卷积模块(multichannel partial convolution,MCPConv),基于此重新构造了主干网络中的BasicBlock结构,命名为MCP Block,减少通道特征冗余,提升多尺度细节特征的获取能力;引入具有学习能力的位置编码,获取更精确、更具表达能力的位置信息;最后引入归一化加权偏差(normalized weighted deviation,NWD)和平均精度驱动交并比(mean precision-driven IoU,MPDIoU)定位损失函数,降低对位置偏差的敏感性,加快模型收敛速度。实验结果表明,在VisDrone2019-DET数据集上,改进后的模型较原始模型参数量降低了62%,检测精度mAP50提升了3.9%,且FPS较改进前提升了17%,对比其他主流检测模型具有更好的检测效果。

, correspAuthors=李卫东, authorNote=null, correspAuthorsNote=
* 李卫东(1973—),男,汉族,河北石家庄人,博士,教授。研究方向:机器学习、深度学习。E-mail:
, copyrightStatement=null, copyrightOwner=null, extLink=null, articleAbsUrl=null, sourceXml=SYBFEtmVmfVNN9RYK0lJ4w==, magXml=1Od40VZhuAovQUBaC315MA==, pdfUrl=null, pdf=edBDfUZkiUL+Z3QtnUhbgQ==, pdfFileSize=null, pdfExtLink=null, richHtmlUrl=null, mobilePdfUrl=null, reviewReport=null, pdfFirstPage=null, abstractGraph=null, abstractGraphContent=null, abstractVideo=null, citation=null, cebUrl=null, magXmlContent=fo9deKWXZkRyNpAm7SG9tg==, mapNumber=null, authorCompany=null, fund=null, authors=

赵子琪(2000—),男,汉族,河北石家庄人,硕士研究生。研究方向:图像处理、目标检测。E-mail:

, authorsList=赵子琪, 李卫东, 李晓娟)}, authors=[Author(id=1175114407483949601, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, orderNo=0, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=583982413@qq.com, emailSecond=null, emailThird=null, correspondingAuthor=0, authorType=1, ext={EN=AuthorExt(id=1175114407559447074, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, authorId=1175114407483949601, language=EN, stringName=Zi-qi ZHAO, firstName=Zi-qi, middleName=null, lastName=ZHAO, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=1, address=1 School of Management Science and Information Engineering, Hebei University of Economics and Business, Shijiazhuang 050062, China, bio=null, bioImg=null, bioContent=null, aboutCorrespAuthor=null), CN=AuthorExt(id=1175114407685276197, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, authorId=1175114407483949601, language=CN, stringName=赵子琪, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=1, address=1 河北经贸大学管理科学与信息工程学院, 石家庄 050062, bio={"content":"

赵子琪(2000—),男,汉族,河北石家庄人,硕士研究生。研究方向:图像处理、目标检测。E-mail:

"}, bioImg=null, bioContent=

赵子琪(2000—),男,汉族,河北石家庄人,硕士研究生。研究方向:图像处理、目标检测。E-mail:

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1175114407324566041, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, xref=1, ext=[AuthorCompanyExt(id=1175114407328760345, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, companyId=1175114407324566041, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1 School of Management Science and Information Engineering, Hebei University of Economics and Business, Shijiazhuang 050062, China), AuthorCompanyExt(id=1175114407337148954, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, companyId=1175114407324566041, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1 河北经贸大学管理科学与信息工程学院, 石家庄 050062)])]), Author(id=1175114407848854054, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, orderNo=1, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=14263859@qq.com, emailSecond=null, emailThird=null, correspondingAuthor=1, authorType=1, ext={EN=AuthorExt(id=1175114407974683178, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, authorId=1175114407848854054, language=EN, stringName=Wei-dong LI, firstName=Wei-dong, middleName=null, lastName=LI, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=1, 2, *, address=1 School of Management Science and Information Engineering, Hebei University of Economics and Business, Shijiazhuang 050062, China
2 Hebei Cross border E-commerce Technology Innovation Center, Shijiazhuang 050062, China, bio=null, bioImg=null, bioContent=null, aboutCorrespAuthor=null), CN=AuthorExt(id=1175114408041792044, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, authorId=1175114407848854054, language=CN, stringName=李卫东, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=1, 2, *, address=1 河北经贸大学管理科学与信息工程学院, 石家庄 050062
2 河北省跨境电商技术创新中心, 石家庄 050062, bio=null, bioImg=null, bioContent=null, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1175114407324566041, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, xref=1, ext=[AuthorCompanyExt(id=1175114407328760345, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, companyId=1175114407324566041, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1 School of Management Science and Information Engineering, Hebei University of Economics and Business, Shijiazhuang 050062, China), AuthorCompanyExt(id=1175114407337148954, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, companyId=1175114407324566041, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1 河北经贸大学管理科学与信息工程学院, 石家庄 050062)]), AuthorCompany(id=1175114407383286299, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, xref=2, ext=[AuthorCompanyExt(id=1175114407391674909, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, companyId=1175114407383286299, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=2 Hebei Cross border E-commerce Technology Innovation Center, Shijiazhuang 050062, China), AuthorCompanyExt(id=1175114407400063518, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, companyId=1175114407383286299, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=2 河北省跨境电商技术创新中心, 石家庄 050062)])]), Author(id=1175114408125678127, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, orderNo=2, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=null, emailSecond=null, emailThird=null, correspondingAuthor=0, authorType=1, ext={EN=AuthorExt(id=1175114408331199026, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, authorId=1175114408125678127, language=EN, stringName=Xiao-juan LI, firstName=Xiao-juan, middleName=null, lastName=LI, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=1, 2, address=1 School of Management Science and Information Engineering, Hebei University of Economics and Business, Shijiazhuang 050062, China
2 Hebei Cross border E-commerce Technology Innovation Center, Shijiazhuang 050062, China, bio=null, bioImg=null, bioContent=null, aboutCorrespAuthor=null), CN=AuthorExt(id=1175114408410890805, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, authorId=1175114408125678127, language=CN, stringName=李晓娟, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=1, 2, address=1 河北经贸大学管理科学与信息工程学院, 石家庄 050062
2 河北省跨境电商技术创新中心, 石家庄 050062, bio=null, bioImg=null, bioContent=null, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1175114407324566041, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, xref=1, ext=[AuthorCompanyExt(id=1175114407328760345, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, companyId=1175114407324566041, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1 School of Management Science and Information Engineering, Hebei University of Economics and Business, Shijiazhuang 050062, China), AuthorCompanyExt(id=1175114407337148954, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, companyId=1175114407324566041, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1 河北经贸大学管理科学与信息工程学院, 石家庄 050062)]), AuthorCompany(id=1175114407383286299, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, xref=2, ext=[AuthorCompanyExt(id=1175114407391674909, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, companyId=1175114407383286299, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=2 Hebei Cross border E-commerce Technology Innovation Center, Shijiazhuang 050062, China), AuthorCompanyExt(id=1175114407400063518, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, companyId=1175114407383286299, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=2 河北省跨境电商技术创新中心, 石家庄 050062)])])], keywords=[Keyword(id=1175114408553497145, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, language=EN, orderNo=1, keyword=small object detection), Keyword(id=1175114408662549051, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, language=EN, orderNo=2, keyword=RT-DETR), Keyword(id=1175114408733852220, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, language=EN, orderNo=3, keyword=multi channel partial convolution), Keyword(id=1175114408788378174, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, language=EN, orderNo=4, keyword=learned position embedding), Keyword(id=1175114408863875648, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, language=CN, orderNo=1, keyword=小目标检测), Keyword(id=1175114409002287680, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, language=CN, orderNo=2, keyword=RT-DETR), Keyword(id=1175114409140699715, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, language=CN, orderNo=3, keyword=多通道部分卷积), Keyword(id=1175114409199419971, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, language=CN, orderNo=4, keyword=可学习位置编码)], refs=[Reference(id=1175114411695030910, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, doi=null, pmid=null, pmcid=null, year=2014, volume=null, issue=null, pageStart=580, pageEnd=587, url=null, language=null, rfNumber=[1], rfOrder=0, authorNames=Girshick R, Donahue J, Darrell T, journalName=Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, refType=null, unstructuredReference=Girshick R, Donahue J, Darrell T, et al. Rich feature hierarchies for accurate object detection and semantic segmentation[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Columbus: IEEE, 2014: 580-587., articleTitle=Rich feature hierarchies for accurate object detection and semantic segmentation, refAbstract=null), Reference(id=1175114411774722689, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, doi=null, pmid=null, pmcid=null, year=2016, volume=null, issue=null, pageStart=779, pageEnd=788, url=null, language=null, rfNumber=[2], rfOrder=1, authorNames=Redmon J, Divvala S, Girshick R, journalName=Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, refType=null, unstructuredReference=Redmon J, Divvala S, Girshick R, et al. You only look once: unified, real-time object detection[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas: IEEE, 2016: 779-788., articleTitle=You only look once: unified, real-time object detection, refAbstract=null), Reference(id=1175114411862803074, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, doi=null, pmid=null, pmcid=null, year=2024, volume=24, issue=18, pageStart=7726, pageEnd=7732, url=null, language=null, rfNumber=[3], rfOrder=2, authorNames=桑雨, 李立权, 李铁, journalName=科学技术与工程, refType=null, unstructuredReference=桑雨, 李立权, 李铁. 轻量化YOLOv7-tiny的遥感图像小目标检测[J]. 科学技术与工程, 2024, 24(18): 7726-7732., articleTitle=轻量化YOLOv7-tiny的遥感图像小目标检测, refAbstract=null), Reference(id=1175114411950883461, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, doi=null, pmid=null, pmcid=null, year=2024, volume=24, issue=18, pageStart=7726, pageEnd=7732, url=null, language=null, rfNumber=[3], rfOrder=3, authorNames=Sang Yu, Li Liquan, Li Tie, journalName=Science Technology and Engineering, refType=null, unstructuredReference=Sang Yu, Li Liquan, Li Tie, et al. Lightweight YOLOv7-tiny for remote sensing image small target detection[J]. Science Technology and Engineering, 2024, 24(18): 7726-7732., articleTitle=Lightweight YOLOv7-tiny for remote sensing image small target detection, refAbstract=null), Reference(id=1175114412059935368, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, doi=null, pmid=null, pmcid=null, year=2023, volume=23, issue=1, pageStart=274, pageEnd=282, url=null, language=null, rfNumber=[4], rfOrder=4, authorNames=刘晋川, 黎向锋, 刘安旭, journalName=科学技术与工程, refType=null, unstructuredReference=刘晋川, 黎向锋, 刘安旭, 等. 改进RetinaNet的无人机小目标检测[J]. 科学技术与工程, 2023, 23(1): 274-282., articleTitle=改进RetinaNet的无人机小目标检测, refAbstract=null), Reference(id=1175114412177375883, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, doi=null, pmid=null, pmcid=null, year=2023, volume=23, issue=1, pageStart=274, pageEnd=282, url=null, language=null, rfNumber=[4], rfOrder=5, authorNames=Liu Jinchuan, Li Xiangfeng, Liu Anxu, journalName=Science Technology and Engineering, refType=null, unstructuredReference=Liu Jinchuan, Li Xiangfeng, Liu Anxu, et al. Improved RetinaNet for unmanned aerial vehicle small target detection[J]. Science Technology and Engineering, 2023, 23(1): 274-282., articleTitle=Improved RetinaNet for unmanned aerial vehicle small target detection, refAbstract=null), Reference(id=1175114412319982224, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, doi=null, pmid=null, pmcid=null, year=2020, volume=20, issue=27, pageStart=11185, pageEnd=11191, url=null, language=null, rfNumber=[5], rfOrder=6, authorNames=吴稳稳, 吴晓红, 刘强, journalName=科学技术与工程, refType=null, unstructuredReference=吴稳稳, 吴晓红, 刘强, 等. 基于全局注意力的多级特征融合目标检测算法[J]. 科学技术与工程, 2020, 20(27): 11185-11191., articleTitle=基于全局注意力的多级特征融合目标检测算法, refAbstract=null), Reference(id=1175114412403868306, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, doi=null, pmid=null, pmcid=null, year=2020, volume=20, issue=27, pageStart=11185, pageEnd=11191, url=null, language=null, rfNumber=[5], rfOrder=7, authorNames=Wu Wenwen, Wu Xiaohong, Liu Qiang, journalName=Science Technology and Engineering, refType=null, unstructuredReference=Wu Wenwen, Wu Xiaohong, Liu Qiang, et al. Multi-level feature fusion object detection algorithm based on global attention[J]. Science Technology and Engineering, 2020, 20(27): 11185-11191., articleTitle=Multi-level feature fusion object detection algorithm based on global attention, refAbstract=null), Reference(id=1175114412542280340, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, doi=null, pmid=null, pmcid=null, year=2024, volume=24, issue=9, pageStart=3533, pageEnd=3542, url=null, language=null, rfNumber=[6], rfOrder=8, authorNames=康传利, 张思瑶, 李玄皓, journalName=科学技术与工程, refType=null, unstructuredReference=康传利, 张思瑶, 李玄皓, 等. 高斯Wasserstein距离改进轻量YOLOv7模型的遥感影像道路交叉口检测[J]. 科学技术与工程, 2024, 24(9): 3533-3542., articleTitle=高斯Wasserstein距离改进轻量YOLOv7模型的遥感影像道路交叉口检测, refAbstract=null), Reference(id=1175114412605194902, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, doi=null, pmid=null, pmcid=null, year=2024, volume=24, issue=9, pageStart=3533, pageEnd=3542, url=null, language=null, rfNumber=[6], rfOrder=9, authorNames=Kang Chuanli, Zhang Siyao, Li Xuanhao, journalName=Science Technology and Engineering, refType=null, unstructuredReference=Kang Chuanli, Zhang Siyao, Li Xuanhao, et al. Gaussian Wasserstein distance improvement of lightweight YOLOv7 model for remote sensing image road intersection detection[J]. Science Technology and Engineering, 2024, 24(9): 3533-3542., articleTitle=Gaussian Wasserstein distance improvement of lightweight YOLOv7 model for remote sensing image road intersection detection, refAbstract=null), Reference(id=1175114412722635418, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, doi=null, pmid=null, pmcid=null, year=2019, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=[7], rfOrder=10, authorNames=Tan M, Le Q, journalName=ArXiv, refType=null, unstructuredReference=Tan M, Le Q. Efficientnet: rethinking model scaling for convolutional neural networks[J]. ArXiv, 2019: 1905.11946., articleTitle=Efficientnet: rethinking model scaling for convolutional neural networks, refAbstract=null), Reference(id=1175114412873630361, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, doi=null, pmid=null, pmcid=null, year=2020, volume=null, issue=null, pageStart=213, pageEnd=229, url=null, language=null, rfNumber=[8], rfOrder=11, authorNames=Carion N, Massa F, Synnaeve G, journalName=European Conference on Computer Vision, refType=null, unstructuredReference=Carion N, Massa F, Synnaeve G, et al. End-to-end object detection with transformers[C]// European Conference on Computer Vision. Cham: Springer International Publishing, 2020: 213-229., articleTitle=End-to-end object detection with transformers, refAbstract=null), Reference(id=1175114413079151259, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, doi=null, pmid=null, pmcid=null, year=2020, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=[9], rfOrder=12, authorNames=Zhu X, Su W, Lu L, journalName=ArXiv Preprint ArXiv, refType=null, unstructuredReference=Zhu X, Su W, Lu L, et al. Deformable detr: deformable transformers for end-to-end object detection[J]. ArXiv Preprint ArXiv, 2020: 2010.04159., articleTitle=Deformable detr: deformable transformers for end-to-end object detection, refAbstract=null), Reference(id=1175114413272089243, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, doi=null, pmid=null, pmcid=null, year=2017, volume=null, issue=null, pageStart=764, pageEnd=773, url=null, language=null, rfNumber=[10], rfOrder=13, authorNames=Dai J, Qi H, Xiong Y, journalName=Proceedings of the IEEE International Conference on Computer Vision, refType=null, unstructuredReference=Dai J, Qi H, Xiong Y, et al. Deformable convolutional networks[C]// Proceedings of the IEEE International Conference on Computer Vision. Venice: IEEE, 2017: 764-773., articleTitle=Deformable convolutional networks, refAbstract=null), Reference(id=1175114413410501277, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, doi=null, pmid=null, pmcid=null, year=2021, volume=null, issue=null, pageStart=3651, pageEnd=3660, url=null, language=null, rfNumber=[11], rfOrder=14, authorNames=Meng D, Chen X, Fan Z, journalName=Proceedings of the IEEE/CVF International Conference on Computer Vision, refType=null, unstructuredReference=Meng D, Chen X, Fan Z, et al. Conditional detr for fast training convergence[C]// Proceedings of the IEEE/CVF International Conference on Computer Vision. Montreal: IEEE, 2021: 3651-3660., articleTitle=Conditional detr for fast training convergence, refAbstract=null), Reference(id=1175114413532136094, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, doi=null, pmid=null, pmcid=null, year=2022, volume=36, issue=3, pageStart=2567, pageEnd=2575, url=null, language=null, rfNumber=[12], rfOrder=15, authorNames=Wang Y, Zhang X, Yang T, journalName=Proceedings of the AAAI Conference on Artificial Intelligence, refType=null, unstructuredReference=Wang Y, Zhang X, Yang T, et al. Anchor DETR: query design for Transformer-based detector[C]// Proceedings of the AAAI Conference on Artificial Intelligence. New York: AAAI, 2022, 36(3): 2567-2575., articleTitle=Anchor DETR: query design for Transformer-based detector, refAbstract=null), Reference(id=1175114413641187999, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, doi=null, pmid=null, pmcid=null, year=2022, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=[13], rfOrder=16, authorNames=Liu S, Li F, Zhang H, journalName=ArXiv Preprint ArXiv, refType=null, unstructuredReference=Liu S, Li F, Zhang H, et al. DAB-DETR: dynamic anchor boxes are better queries for DETR[J]. ArXiv Preprint ArXiv, 2022: 2201.12329., articleTitle=DAB-DETR: dynamic anchor boxes are better queries for DETR, refAbstract=null), Reference(id=1175114413725074081, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, doi=null, pmid=null, pmcid=null, year=2022, volume=46, issue=4, pageStart=13619, pageEnd=13627, url=null, language=null, rfNumber=[14], rfOrder=17, authorNames=Li F, Zhang H, Liu S, journalName=IEEE Transactions on Pattern Analysis and Machine Intelligence, refType=null, unstructuredReference=Li F, Zhang H, Liu S, et al. DN-DETR: accelerate DETR training by introducing query denoising[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022, 46(4): 13619-13627., articleTitle=DN-DETR: accelerate DETR training by introducing query denoising, refAbstract=null), Reference(id=1175114413876069027, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, doi=null, pmid=null, pmcid=null, year=2022, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=[15], rfOrder=18, authorNames=Zhang H, Li F, Liu S, journalName=ArXiv Preprint ArXiv, refType=null, unstructuredReference=Zhang H, Li F, Liu S, et al. Dino: DETR with improved denoising anchor boxes for end-to-end object detection[J]. ArXiv Preprint ArXiv, 2022: 2203.03605., articleTitle=Dino: DETR with improved denoising anchor boxes for end-to-end object detection, refAbstract=null), Reference(id=1175114414027063973, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, doi=null, pmid=null, pmcid=null, year=2023, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=[16], rfOrder=19, authorNames=Lü W, Xu S, Zhao Y, journalName=ArXiv Preprint ArXiv, refType=null, unstructuredReference= W, Xu S, Zhao Y, et al. DETRs beat yolos on real-time object detection[J]. ArXiv Preprint ArXiv, 2023: 2304.08069., articleTitle=DETRs beat yolos on real-time object detection, refAbstract=null), Reference(id=1175114414148698791, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, doi=null, pmid=null, pmcid=null, year=2014, volume=null, issue=null, pageStart=740, pageEnd=755, url=null, language=null, rfNumber=[17], rfOrder=20, authorNames=Lin T Y, Maire M, Belongie S, journalName=Computer Vision-ECCV 2014: 13th European Conference, refType=null, unstructuredReference=Lin T Y, Maire M, Belongie S, et al. Microsoft COCO: common objects in context[C]// Computer Vision-ECCV 2014: 13th European Conference. Zurich: Springer International Publishing, 2014: 740-755., articleTitle=Microsoft COCO: common objects in context, refAbstract=null), Reference(id=1175114414404551337, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, doi=null, pmid=null, pmcid=null, year=2021, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=[18], rfOrder=21, authorNames=Wang J, Xu C, Yang W, journalName=ArXiv Preprint ArXiv, refType=null, unstructuredReference=Wang J, Xu C, Yang W, et al. A normalized Gaussian Wasserstein distance for tiny object detection[J]. ArXiv Preprint ArXiv, 2021: 2110.13389., articleTitle=A normalized Gaussian Wasserstein distance for tiny object detection, refAbstract=null), Reference(id=1175114414501020330, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, doi=null, pmid=null, pmcid=null, year=2023, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=[19], rfOrder=22, authorNames=Siliang M, Yong X, journalName=ArXiv Preprint ArXiv, refType=null, unstructuredReference=Siliang M, Yong X. Mpdiou: a loss for efficient and accurate bounding box regression[J]. ArXiv Preprint ArXiv, 2023: 2307.07662., articleTitle=Mpdiou: a loss for efficient and accurate bounding box regression, refAbstract=null), Reference(id=1175114414614266540, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, doi=null, pmid=null, pmcid=null, year=2016, volume=null, issue=null, pageStart=770, pageEnd=778, url=null, language=null, rfNumber=[20], rfOrder=23, authorNames=He K, Zhang X, Ren S, journalName=Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, refType=null, unstructuredReference=He K, Zhang X, Ren S, et al. Deep residual learning for image recognition[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Chengdu: IEEE, 2016: 770-778., articleTitle=Deep residual learning for image recognition, refAbstract=null), Reference(id=1175114414677181102, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, doi=null, pmid=null, pmcid=null, year=2012, volume=null, issue=null, pageStart=25, pageEnd=null, url=null, language=null, rfNumber=[21], rfOrder=24, authorNames=Krizhevsky A, Sutskever I, Hinton G E, journalName=Advances in Neural Information Processing Systems, refType=null, unstructuredReference=Krizhevsky A, Sutskever I, Hinton G E. ImageNet classification with deep convolutional neural networks[J]. Advances in Neural Information Processing Systems, 2012, 25., articleTitle=ImageNet classification with deep convolutional neural networks, refAbstract=null), Reference(id=1175114414941422256, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, doi=null, pmid=null, pmcid=null, year=2014, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=[22], rfOrder=25, authorNames=Sifre L, Mallat S, journalName=ArXiv Preprint ArXiv, refType=null, unstructuredReference=Sifre L, Mallat S. Rigid-motion scattering for texture classification[J]. ArXiv Preprint ArXiv, 2014: 1403.1687., articleTitle=Rigid-motion scattering for texture classification, refAbstract=null), Reference(id=1175114415092417202, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, doi=null, pmid=null, pmcid=null, year=2023, volume=null, issue=null, pageStart=12021, pageEnd=12031, url=null, language=null, rfNumber=[23], rfOrder=26, authorNames=Chen J, Kao S, He H, journalName=Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, refType=null, unstructuredReference=Chen J, Kao S, He H, et al. Run, don't walk: chasing higher FLOPS for faster neural networks[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Vancouver: IEEE, 2023: 12021-12031., articleTitle=Run, don't walk: chasing higher FLOPS for faster neural networks, refAbstract=null), Reference(id=1175114415193080500, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, doi=null, pmid=null, pmcid=null, year=2019, volume=null, issue=null, pageStart=1314, pageEnd=1324, url=null, language=null, rfNumber=[24], rfOrder=27, authorNames=Howard A, Sandler M, Chu G, journalName=Proceedings of the IEEE/CVF International Conference on Computer Vision, refType=null, unstructuredReference=Howard A, Sandler M, Chu G, et al. Searching for MobileNetV3[C]// Proceedings of the IEEE/CVF International Conference on Computer Vision. Seoul: IEEE, 2019: 1314-1324., articleTitle=Searching for MobileNetV3, refAbstract=null), Reference(id=1175114415360852660, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, doi=null, pmid=null, pmcid=null, year=2018, volume=null, issue=null, pageStart=6848, pageEnd=6856, url=null, language=null, rfNumber=[25], rfOrder=28, authorNames=Zhang X, Zhou X, Lin M, journalName=Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, refType=null, unstructuredReference=Zhang X, Zhou X, Lin M, et al. ShuffleNet: an extremely efficient convolutional neural network for mobile devices[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Salt Lake City: IEEE, 2018: 6848-6856., articleTitle=ShuffleNet: an extremely efficient convolutional neural network for mobile devices, refAbstract=null), Reference(id=1175114415478293174, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, doi=null, pmid=null, pmcid=null, year=2020, volume=null, issue=null, pageStart=1580, pageEnd=1589, url=null, language=null, rfNumber=[26], rfOrder=29, authorNames=Han K, Wang Y, Tian Q, journalName=Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, refType=null, unstructuredReference=Han K, Wang Y, Tian Q, et al. GhostNet: more features from cheap operations[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Seattle: IEEE, 2020: 1580-1589., articleTitle=GhostNet: more features from cheap operations, refAbstract=null), Reference(id=1175114415553790648, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, doi=null, pmid=null, pmcid=null, year=2019, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=[27], rfOrder=30, authorNames=Tan M, Le Q, journalName=ArXiv, refType=null, unstructuredReference=Tan M, Le Q. EfficientNet: rethinking model scaling for convolutional neural networks[J]. ArXiv, 2019: 1905.11946., articleTitle=EfficientNet: rethinking model scaling for convolutional neural networks, refAbstract=null), Reference(id=1175114415700591289, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, doi=null, pmid=null, pmcid=null, year=2020, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=[28], rfOrder=31, authorNames=Zhang Q, Jiang Z, Lu Q, journalName=ArXiv Preprint ArXiv, refType=null, unstructuredReference=Zhang Q, Jiang Z, Lu Q, et al. Split to be slim: an overlooked redundancy in Vanilla convolution[J]. ArXiv Preprint ArXiv, 2020: 2006.12085., articleTitle=Split to be slim: an overlooked redundancy in Vanilla convolution, refAbstract=null)], funds=[Fund(id=1175114411300766327, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, awardId=SZX2020034, language=CN, fundingSource=河北省省级科技计划(SZX2020034), fundOrder=null, country=null), Fund(id=1175114411443372665, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, awardId=BJK2022041, language=CN, fundingSource=河北省高等学校科学研究计划(BJK2022041), fundOrder=null, country=null)], companyList=[AuthorCompany(id=1175114407324566041, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, xref=1, ext=[AuthorCompanyExt(id=1175114407328760345, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, companyId=1175114407324566041, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1 School of Management Science and Information Engineering, Hebei University of Economics and Business, Shijiazhuang 050062, China), AuthorCompanyExt(id=1175114407337148954, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, companyId=1175114407324566041, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1 河北经贸大学管理科学与信息工程学院, 石家庄 050062)]), AuthorCompany(id=1175114407383286299, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, xref=2, ext=[AuthorCompanyExt(id=1175114407391674909, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, companyId=1175114407383286299, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=2 Hebei Cross border E-commerce Technology Innovation Center, Shijiazhuang 050062, China), AuthorCompanyExt(id=1175114407400063518, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, companyId=1175114407383286299, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=2 河北省跨境电商技术创新中心, 石家庄 050062)])], figs=[ArticleFig(id=1175114409388163659, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, language=EN, label=Fig.1, caption=Improved RT-DETR model structure, figureFileSmall=K3XwWYNeOWgoqB6villyZA==, figureFileBig=Ch3bN5pVs+3Wrvdw3wE8lw==, tableContent=null), ArticleFig(id=1175114409446883916, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, language=CN, label=图1, caption=改进RT-DETR算法模型结构, figureFileSmall=K3XwWYNeOWgoqB6villyZA==, figureFileBig=Ch3bN5pVs+3Wrvdw3wE8lw==, tableContent=null), ArticleFig(id=1175114409513992783, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, language=EN, label=Fig.2, caption=Model structure comparison, figureFileSmall=cPNoqzbaf2tmb1jKuEM04g==, figureFileBig=DMG940CKfqwVtaxLWvtLXA==, tableContent=null), ArticleFig(id=1175114409618850386, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, language=CN, label=图2, caption=模型结构对比图, figureFileSmall=cPNoqzbaf2tmb1jKuEM04g==, figureFileBig=DMG940CKfqwVtaxLWvtLXA==, tableContent=null), ArticleFig(id=1175114409753068118, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, language=EN, label=Fig.3, caption=Multichannel partial convolution, figureFileSmall=17KV/nIoqo6paP+JiUr++Q==, figureFileBig=qNf06OOwuF35cyHfs5HyNg==, tableContent=null), ArticleFig(id=1175114409832759895, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, language=CN, label=图3, caption=MCPConv, figureFileSmall=17KV/nIoqo6paP+JiUr++Q==, figureFileBig=qNf06OOwuF35cyHfs5HyNg==, tableContent=null), ArticleFig(id=1175114409916645978, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, language=EN, label=Fig.4, caption=Comparison of Basic Block and MCP Block design, figureFileSmall=qiUBZslDv0Vq8VAQvRwLdA==, figureFileBig=n0PSeqUSPhlT4P0R6g6xgw==, tableContent=null), ArticleFig(id=1175114410004726364, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, language=CN, label=图4, caption=Basic Block 与MCP Block设计对比, figureFileSmall=qiUBZslDv0Vq8VAQvRwLdA==, figureFileBig=n0PSeqUSPhlT4P0R6g6xgw==, tableContent=null), ArticleFig(id=1175114410134749791, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, language=EN, label=Fig.5, caption=Comparison between the improved model and the original model, figureFileSmall=mippP/sKgC6Fu4ecPF5R+w==, figureFileBig=Dnwg8CQZMciNWTEgJQqzhQ==, tableContent=null), ArticleFig(id=1175114410252190306, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, language=CN, label=图5, caption=改进后的模型与原模型检测效果对比, figureFileSmall=mippP/sKgC6Fu4ecPF5R+w==, figureFileBig=Dnwg8CQZMciNWTEgJQqzhQ==, tableContent=null), ArticleFig(id=1175114410336076389, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, language=EN, label=Table 1, caption=

Model performance comparison

, figureFileSmall=null, figureFileBig=null, tableContent=
方法 参数量/106 计算量 准确率/% 召回率/% mAP50/% mAP50:95/% FPS/(帧·s-1)
RT-DETR 19.9 57.3 55.4 37.6 36.6 21.0 60
方法一 18.6 78.8 55.2 40.3 39.3 23.2 46
方法二 7.9 69.4 56.2 39.9 39.1 23.1 61
), ArticleFig(id=1175114410411573864, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, language=CN, label=表1, caption=

模型效果对比

, figureFileSmall=null, figureFileBig=null, tableContent=
方法 参数量/106 计算量 准确率/% 召回率/% mAP50/% mAP50:95/% FPS/(帧·s-1)
RT-DETR 19.9 57.3 55.4 37.6 36.6 21.0 60
方法一 18.6 78.8 55.2 40.3 39.3 23.2 46
方法二 7.9 69.4 56.2 39.9 39.1 23.1 61
), ArticleFig(id=1175114410537402986, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, language=EN, label=Table 2, caption=

Cryogenic experiment

, figureFileSmall=null, figureFileBig=null, tableContent=
序号 方法 参数量/106 计算量 准确率/% 召回率/% mAP50/% mAP50:95/% FPS/(帧·s-1)
RT-DETR 19.9 57.3 55.4 37.6 36.6 21.0 60
A RT-DETR+MCP Block 16.8 49.8 54.5 38.6 37.5 21.8 58
B RT-DETR+AIFI-LPE 20.0 57.3 55.5 38.8 37.8 22.0 63
C RT-DETR+NWD-MPDIoU 19.9 57.3 55.6 38.2 37.5 21.7 59
D RT-DETR+多尺度特征 7.9 69.4 56.2 39.9 39.1 23.1 61
E A+B 16.9 49.8 56.2 39.1 38.0 22.1 60
F A+B+C 16.9 49.8 56.0 39.4 38.5 22.4 61
本文 A+B+C+D 7.5 63.2 57.7 41.8 40.5 23.7 70
), ArticleFig(id=1175114410738729581, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, language=CN, label=表2, caption=

消融实验

, figureFileSmall=null, figureFileBig=null, tableContent=
序号 方法 参数量/106 计算量 准确率/% 召回率/% mAP50/% mAP50:95/% FPS/(帧·s-1)
RT-DETR 19.9 57.3 55.4 37.6 36.6 21.0 60
A RT-DETR+MCP Block 16.8 49.8 54.5 38.6 37.5 21.8 58
B RT-DETR+AIFI-LPE 20.0 57.3 55.5 38.8 37.8 22.0 63
C RT-DETR+NWD-MPDIoU 19.9 57.3 55.6 38.2 37.5 21.7 59
D RT-DETR+多尺度特征 7.9 69.4 56.2 39.9 39.1 23.1 61
E A+B 16.9 49.8 56.2 39.1 38.0 22.1 60
F A+B+C 16.9 49.8 56.0 39.4 38.5 22.4 61
本文 A+B+C+D 7.5 63.2 57.7 41.8 40.5 23.7 70
), ArticleFig(id=1175114410944250479, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, language=EN, label=Table 3, caption=

Comparative Experiment

, figureFileSmall=null, figureFileBig=null, tableContent=
方法 参数量/106 计算量 准确率/% 召回率/% mAP50/% mAP50:95/% FPS/(帧·s-1)
YOLOv5-m 25.0 64.2 48.5 35.3 34.1 19.8 112
YOLOv5-l 53.1 135.0 48.4 37.1 35.5 20.8 99
YOLOv8-m 25.9 79.3 48.1 36.0 34.4 20.0 108
YOLOv8-L 43.6 165.7 49.2 36.8 35.3 20.7 101
本文 7.5 63.2 57.7 41.8 40.5 23.7 70
), ArticleFig(id=1175114411023942257, tenantId=1146029695717560320, journalId=1146123166801305609, articleId=1149773874725871618, language=CN, label=表3, caption=

对比实验

, figureFileSmall=null, figureFileBig=null, tableContent=
方法 参数量/106 计算量 准确率/% 召回率/% mAP50/% mAP50:95/% FPS/(帧·s-1)
YOLOv5-m 25.0 64.2 48.5 35.3 34.1 19.8 112
YOLOv5-l 53.1 135.0 48.4 37.1 35.5 20.8 99
YOLOv8-m 25.9 79.3 48.1 36.0 34.4 20.0 108
YOLOv8-L 43.6 165.7 49.2 36.8 35.3 20.7 101
本文 7.5 63.2 57.7 41.8 40.5 23.7 70
)], attaches=null, journal=Journal(id=1146119176004939786, delFlag=0, nameCn=科学技术与工程, nameEn=Science Technology and Engineering, nameHistory1=null, nameHistory2=null, issn=1671-1815, eissn=, cn=11-4688/T, coden=null, periodic=4, language=CN, oaType=是, ccby=null, superviseOffice=null, ownerOffice=null, pubOffice=null, editorOffice=null, officeType=null, aims=null, clcCode=null, officeProv=null, officeCity=null, officeAddr=null, officeZip=null, officeEmail=null, officePhone=null, editDirector=null, officeDirector=null, officeDirectorPhone=null, officeStaffNum=null, officeEmpNum=null, coverPicUrl=UKU/O7GSka5polgCTkbIIw==, journalPrice=null, startedYear=null, abbrevIsoEn=Sci Technol Eng, journalRemark=null, publicationField=null, createdTime=null, updatedTime=1754445529766, createdBy=null, updatedBy=13701087609, firstLetterCn=S, firstLetterEn=S, subjectCode=Natural Sciences, subjectName=自然科学, subjectCodeEn=Natural Sciences, subjectNameEn=null, picCn=UKU/O7GSka5polgCTkbIIw==, picEn=5hwlULoNwcbj3xUmVi9MAQ==, jcr=null, cjcr=null, exts=[JournalExt(id=1159791870395564357, language=CN, name=科学技术与工程, nameHistory1=null, nameHistory2=null, managedBy=, sponsoredBy=, publishedBy=, editorOffice=, officeProv=null, officeCity=null, officeAddr=, officeZip=, editDirector=null, officeDirector=null, officePhone=null, coverPicUrl=null, journalRemark=, submitArticleUrl=null, websiteUrl=http://www.stae.com.cn/jsygc/home, createdTime=1754445529793, updatedTime=1754445529793, createdBy=13701087609, updatedBy=13701087609, submissionGuidelinesUrl=http://www.stae.com.cn/jsygc/site/menus/20090429150146001, submissionAuthorUrl=http://www.stae.com.cn/jsygc/author/login, submissionEditorUrl=http://www.stae.com.cn/jsygc/editor/login, submissionReviewUrl=http://www.stae.com.cn/jsygc/reviewer/login, submissionCeEditorUrl=, submissionAeEditorUrl=, option={"copyright":""}), JournalExt(id=1159791870441701702, language=EN, name=Science Technology and Engineering, nameHistory1=null, nameHistory2=null, managedBy=, sponsoredBy=, publishedBy=, editorOffice=, officeProv=null, officeCity=null, officeAddr=, officeZip=, editDirector=null, officeDirector=null, officePhone=null, coverPicUrl=null, journalRemark=, submitArticleUrl=null, websiteUrl=http://www.stae.com.cn/jsygc/home, createdTime=1754445529804, updatedTime=1754445529804, createdBy=13701087609, updatedBy=13701087609, submissionGuidelinesUrl=, submissionAuthorUrl=http://www.stae.com.cn/jsygc/author/login, submissionEditorUrl=http://www.stae.com.cn/jsygc/editor/login, submissionReviewUrl=http://www.stae.com.cn/jsygc/reviewer/login, submissionCeEditorUrl=, submissionAeEditorUrl=, option={"copyright":""})], databaseList=null, tenantJournalId=1146123166801305609, websiteList=[Website(id=1148243202391400884, webName=null, webTitle=null, webDomain=null, webCopyrigh=null, webIpcNo=null, seoTitle=null, seoKeywords=null, seoDescription=null, tenantJournalId=null, journalId=1146123166801305609, journalNameCn=null, journalNameEn=null, grayFlag=null, tenantId=1146029695717560320, platformId=null, journalGroupId=null, journalGroupNameCn=null, journalGroupNameEn=null, type=1, domain=https://castjournals.cast.org.cn/joweb/kxjsygc/CN, language=CN, createTime=1751692112777, createBy=18614031015, updateTime=1753520965431, updateBy=18614031015, name=科学技术与工程-中文站点, tplId=1146099689490845704, title=科学技术与工程, delFlag=0, indexPage=/home, props=[WebsiteProps(id=1148622798802673703, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1148243202391400884, code=articleTextType, value=kx, createTime=1751782615614, updateTime=1751782615614, creator=18614031015, updator=18614031015), WebsiteProps(id=1148622798781702180, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1148243202391400884, code=banner, value=null, createTime=1751782615609, updateTime=1751782615609, creator=18614031015, updator=18614031015), WebsiteProps(id=1148622798769119267, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1148243202391400884, code=logo, value=https://castjournals.cast.org.cn/joweb/kjdb/CN/file/pic?fileId=j86gbwi+p0Idkyl5SzIlmQ==, createTime=1751782615606, updateTime=1751782615606, creator=18614031015, updator=18614031015), WebsiteProps(id=1148622798794285094, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1148243202391400884, code=picServerUrl, value=https://castjournals.cast.org.cn/joweb/kjdb/CN/file/pic, createTime=1751782615612, updateTime=1751782615612, creator=18614031015, updator=18614031015), WebsiteProps(id=1148622798790090789, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1148243202391400884, code=staticResourcePath, value=https://castjournals.cast.org.cn/joweb/cast_kjdb_cn_619/, createTime=1751782615611, updateTime=1751782615611, creator=18614031015, updator=18614031015)]), Website(id=1155914124811976731, webName=null, webTitle=null, webDomain=null, webCopyrigh=null, webIpcNo=null, seoTitle=null, seoKeywords=null, seoDescription=null, tenantJournalId=null, journalId=1146123166801305609, journalNameCn=null, journalNameEn=null, grayFlag=null, tenantId=1146029695717560320, platformId=null, journalGroupId=null, journalGroupNameCn=null, journalGroupNameEn=null, type=1, domain=https://castjournals.cast.org.cn/joweb/kxjsygc/EN, language=EN, createTime=1753521003206, createBy=18614031015, updateTime=1753521003206, updateBy=18614031015, name=科学技术与工程-英文站点, tplId=1146101810881728533, title=Science Technology and Engineering, delFlag=0, indexPage=/home, props=[WebsiteProps(id=1155914371227308235, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1155914124811976731, code=articleTextType, value=kx, createTime=1753521061952, updateTime=1753521061952, creator=18614031015, updator=18614031015), WebsiteProps(id=1155914371210531016, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1155914124811976731, code=banner, value=null, createTime=1753521061947, updateTime=1753521061947, creator=18614031015, updator=18614031015), WebsiteProps(id=1155914371202142407, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1155914124811976731, code=logo, value=https://castjournals.cast.org.cn/joweb/kjdb/CN/file/pic?fileId=j86gbwi+p0Idkyl5SzIlmQ==, createTime=1753521061945, updateTime=1753521061945, creator=18614031015, updator=18614031015), WebsiteProps(id=1155914371223113930, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1155914124811976731, code=picServerUrl, value=https://castjournals.cast.org.cn/joweb/kjdb/CN/file/pic, createTime=1753521061950, updateTime=1753521061950, creator=18614031015, updator=18614031015), WebsiteProps(id=1155914371218919625, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1155914124811976731, code=staticResourcePath, value=https://castjournals.cast.org.cn/joweb/cast_kjdb_cn_619/, createTime=1753521061949, updateTime=1753521061949, creator=18614031015, updator=18614031015)])], journalTitle=科学技术与工程, weixinUrl=null, journalUrl=null, iacademicId=null, status=0, seqNo=null, journalTitleEn=Science Technology and Engineering, journalPhotoCn=UKU/O7GSka5polgCTkbIIw==, journalPhotoEn=5hwlULoNwcbj3xUmVi9MAQ==, journalFirstLetter=S, journalRecommend=null, journalNew=null, journalCollection=null, jcrJf=null, cjcrJf=null, jcrJfStr=null, cjcrJfStr=null, submissionFirstDecision=null, sciSubjectClassification=null, casSubjectClassification=null, citeScore=null, totalCitationFrequency=null, icpCode=null, psCode=null, advertisingLicenseCode=null, copyrightInformation=null, country=null, option=null, provinceCode=null, provinceName=null, collectFlag=false), detailUrlCn=https://castjournals.cast.org.cn/joweb/kxjsygc/CN/10.12404/j.issn.1671-1815.2404266, detailUrlEn=https://castjournals.cast.org.cn/joweb/kxjsygc/EN/10.12404/j.issn.1671-1815.2404266, pdfUrlCn=https://castjournals.cast.org.cn/joweb/kxjsygc/CN/PDF/10.12404/j.issn.1671-1815.2404266, pdfUrlEn=https://castjournals.cast.org.cn/joweb/kxjsygc/EN/PDF/10.12404/j.issn.1671-1815.2404266, aliStartDate=null, aliEndDate=null, collectionFlag=false, citedCount=null, citedUrl=null, reference=null)
收藏切换
基于改进实时Transformer的航拍图像小目标检测算法
收藏切换
PDF下载
赵子琪 1 , 李卫东 1, 2, * , 李晓娟 1, 2
科学技术与工程 | 论文·自动化技术、计算机技术 2025,25(13): 5527-5534
收起
收藏切换
科学技术与工程 | 论文·自动化技术、计算机技术 2025, 25(13): 5527-5534
基于改进实时Transformer的航拍图像小目标检测算法
全屏
赵子琪1 , 李卫东1, 2, * , 李晓娟1, 2
作者信息
  • 1 河北经贸大学管理科学与信息工程学院, 石家庄 050062
  • 2 河北省跨境电商技术创新中心, 石家庄 050062
  • 赵子琪(2000—),男,汉族,河北石家庄人,硕士研究生。研究方向:图像处理、目标检测。E-mail:

通讯作者:

* 李卫东(1973—),男,汉族,河北石家庄人,博士,教授。研究方向:机器学习、深度学习。E-mail:
Small Target Detection Algorithm in Aerial Images Based on Improved RT-DETR
Zi-qi ZHAO1 , Wei-dong LI1, 2, * , Xiao-juan LI1, 2
Affiliations
  • 1 School of Management Science and Information Engineering, Hebei University of Economics and Business, Shijiazhuang 050062, China
  • 2 Hebei Cross border E-commerce Technology Innovation Center, Shijiazhuang 050062, China
出版时间: 2025-05-08 doi: 10.12404/j.issn.1671-1815.2404266
文章导航
收藏切换

针对无人机航拍图像中背景复杂、小目标样本多,难以提取有效特征等问题,提出一种改进实时Transformer(real-time detection Transformer,RT-DETR)的无人机航拍小目标检测算法。首先,在特征融合网络中增加针对微小目标的特征融合结构,利用浅层特征图中丰富的位置信息来增强网络对小目标的检测能力,同时为了防止额外参数的增加,去除主干网络中最后一个残差结构;其次,设计一种多通道特征部分卷积模块(multichannel partial convolution,MCPConv),基于此重新构造了主干网络中的BasicBlock结构,命名为MCP Block,减少通道特征冗余,提升多尺度细节特征的获取能力;引入具有学习能力的位置编码,获取更精确、更具表达能力的位置信息;最后引入归一化加权偏差(normalized weighted deviation,NWD)和平均精度驱动交并比(mean precision-driven IoU,MPDIoU)定位损失函数,降低对位置偏差的敏感性,加快模型收敛速度。实验结果表明,在VisDrone2019-DET数据集上,改进后的模型较原始模型参数量降低了62%,检测精度mAP50提升了3.9%,且FPS较改进前提升了17%,对比其他主流检测模型具有更好的检测效果。

小目标检测  /  RT-DETR  /  多通道部分卷积  /  可学习位置编码

An algorithm has been proposed to detect small targets in unmanned aerial vehicle(UAV) aerial images. The algorithm is based on an improved real-time detection Transformer (RT-DETR) and aims to address the challenges posed by complex backgrounds and a large number of small target samples. To enhance the feature fusion network, a dedicated feature fusion structure for small targets has been incorporated, utilizing rich location information from the shallow feature map to improve the network's ability to detect small targets. Furthermore, the last residual block in the BackBone has been removed to prevent an increase in additional parameters. Additionally, the MCP Block, a reconstructed BasicBlock structure in the backbone network, has been designed, which includes a multi-channel feature partial convolution module (MCPConv) to reduce redundancy in channel features and enhance the acquisition of multi-scale detail features. Moreover, a location encoding mechanism with learning ability has been introduced to obtain more accurate and expressive location information. The normalized weighted deviation(NWD) and mean precision-driven IoU(MPDIoU) positioning loss functions have been incorporated to accelerate the convergence speed of the model and reduce sensitivity to position deviation. Experimental results on the VisDrone2019-DET dataset demonstrate that the improved model reduces parameters by 62% compared to the original model, increases mAP50 by 3.9%, and improves FPS by 17%. The improved model exhibits superior detection performance compared to other mainstream detection models.

small object detection  /  RT-DETR  /  multi channel partial convolution  /  learned position embedding
赵子琪, 李卫东, 李晓娟. 基于改进实时Transformer的航拍图像小目标检测算法. 科学技术与工程, 2025 , 25 (13) : 5527 -5534 . DOI: 10.12404/j.issn.1671-1815.2404266
Zi-qi ZHAO, Wei-dong LI, Xiao-juan LI. Small Target Detection Algorithm in Aerial Images Based on Improved RT-DETR[J]. Science Technology and Engineering, 2025 , 25 (13) : 5527 -5534 . DOI: 10.12404/j.issn.1671-1815.2404266
随着计算机视觉和无人机技术的不断发展,结合无人机对航拍图像进行目标检测的方法越来越成熟。然而,无人机航拍图像存在背景复杂度高、目标尺寸小、外观模糊等问题。基于传统方法的目标检测技术对小目标检测精度低,对复杂背景下目标的适应性弱,容易发生漏检和误检。相比之下,基于深度学习的目标检测算法在检测速度和精度方面具有巨大的优势,能够显著提升目标检测的性能。
现阶段目标检测器有两种典型架构:基于卷积神经网络(convolutional neural networks,CNN)和基于Transformer的目标检测器。基于CNN架构的目标检测器近年来发展迅速,由最初的以区域卷积神经网络(region convolutional neural networks,R-CNN)等[1]双阶段算法发展到以YOLO(you only look once)系列[2]为代表的单阶段算法。YOLO系列在兼顾速度的同时仍能保持较高的检测精度,目前已广泛应用于各种工业场景中。
在无人机航拍领域,由于背景复杂,目标尺寸小,难以提取有效特征,会导致航拍场景下目标检测效果不佳,从而引发误检、漏检、目标定位不准确等问题。同时航拍目标检测任务对时间、空间复杂度都有很高的要求。针对这些问题,研究者们提出了很多改进策略。桑雨等[3]通过对预测头添加语义引导,并设计轻量级池化结构,强化小目标位置信息,减少了计算成本。刘晋川等[4]结合注意力机制构建特征挖掘模块,融入深层语义信息丰富小目标特征。吴稳稳等[5]针对小目标检测精度较低的问题,设计多级特征融合并引入注意力机制,对重要特征增强,提高特征图表达能力。康传利等[6]在网络中加入三维注意力机制并采用高斯距离判定误差,优化特征并优化损失函数。通过对模型的不断改进,单阶段目标检测器目前在航拍领域取得了较好的效果,但依然存在着无法解决的问题。
单阶段目标检测是“密集预测”思想的典型代表,单阶段目标检测器会在整个图像上密集地生成锚框[7],并对每个锚框进行目标分类和位置预测,而在众多锚框当中真正含有目标的锚框占比极低,正负样本数量的极大不平衡是单阶段目标检测网络始终面临的问题。
传统的基于卷积神经网络的目标检测离不开锚框设计、阈值筛选、非极大值抑制(non-maximum suppression,NMS)等前后处理步骤,并且其中超参数的设置极大影响模型的性能,这导致当前的目标检测架构难以做到真正意义上的端到端检测。直到Transformer应用到目标检测领域后,这一问题有了新的解决方案。Carion等[8]提出基于Transfomer架构的新的端到端目标检测器(detection transformer,DETR),DETR使用可学习的查询(query)来预测目标的存在,通过二分匹配来分配标签,解决了模型对锚框及非极大值抑制等后处理操作的依赖问题,将原本的“密集检测”变成了“稀疏检测”,实现了真正意义上的端到端检测。尽管有明显的优势,但DETR目前依然存在两个主要问题:训练收敛缓慢;查询过程难以优化。针对这两个问题目前也存在一些改进方案,Deformable DETR[9]通过提高注意力机制的效率[10]加速收敛过程,同时引入多尺度特征解决了小目标检测性能不足的问题。Conditional DETR[11]和Anchor DETR[12]通过提高目标查询的定位能力加速模型收敛过程。DAB-DETR[13]引入4D参考点逐层迭代优化预测框。DN-DETR[14]通过引入查询去噪来加速训练收敛。DINO[15]沿用之前工作的思想并加以改进,取得了更好的效果,然而现有DETR算法还是无法满足实时性的要求。
在此基础上,Lü等[16]提出了一个新的基于Transformer的实时目标检测算法RT-DETR,相较于YOLOv8,RT-DETR以较少的训练时长在同等测试条件下展现出了更强的性能和更好的平衡,且检测速度不输YOLOv8。在摆脱单阶段目标检测器对手工组件的依赖问题之后,DETR系列模型经过不断的更迭,首次达到了实时检测的效果。这项研究为无人机航拍目标检测提供了新的思路。
RT-DETR算法在COCO数据集[17]上取得了良好的结果。但在无人机航拍领域,尤其是小目标检测方面,仍然面临诸多挑战。具体包括图像中背景复杂、小目标样本多、尺度差异大以及难以提取有效特征等问题。此外,现有的损失函数对于不同尺度的空间敏感性差异较大,也会导致误检和漏检的现象。因此,现提出一种基于改进RT-DETR的航拍小目标检测方法,旨在增强其在航拍场景中的检测效果。
针对航拍场景下小目标样本多、难以提取有效特征的问题,构建针对小目标的多尺度特征提取与融合模块,显著提高模型对小目标的感知与捕获能力。同时优化特征融合结构,提升特征融合过程中小目标特征信息的比重,以增强小目标特征信息的表达能力。
针对航拍场景下目标尺度差异大的问题,提出一种多通道特征部分卷积模块,并基于此设计轻量化的MCP Block。该模块通过捕捉来自不同尺度的特征信息,显著提升网络对多尺度特征的提取能力,从而增强模型对不同尺度目标的捕捉效果。与传统的Basic Block相比,MCP Block在提升特征提取能力的同时,有效减少复杂度和计算冗余。
针对航拍场景下背景复杂导致目标定位不准确的问题,提出使用可学习位置编码替代原始编码,使模型在复杂背景下能够对小目标进行更精准的定位。
针对航拍场景下误检、漏检目标过多的问题,结合归一化加权偏差(normalized weighted deviation,NWD)[18]和平均精度驱动交并比(mean precision-driven IoU,MPDIoU)[19]对损失函数进行重新设计,从而显著降低模型对小目标位置偏移的敏感度,提高模型对目标尺寸变化的鲁棒性,加快模型收敛速度。
RT-DETR是第一个实时端到端目标检测器,不仅在速度和精度上都优于目前最先进的实时检测器,而且不需要后处理,因此其推理速度无延迟,保持稳定。RT-DETR共有6个版本:RT-DETR-R18、RT-DETR-R34、RT-DETR-R50、RT-DETR-R101、RT-DETR-L、RT-DETR-X,6个版本的区别在于主干网络的选择,其中RT-DETR-R18模型网络小,运行速度快,对小目标应用场景的考虑,选择基于该模型进行改进。改进后的网络结构如图1所示。
RT-DETR模型由主干网络、颈部网络以及解码器组成。使用ResNet18[20]作为主干网络,利用主干的最后3个阶段的特征S3、S4和S5作为编码器的输入。S5经过尺度内特征交互(attention-based intra-scale feature interaction,AIFI)后和S3、S4一起输入跨尺度特征融合模块(cross-scale feature fusion module,CCFM)将多尺度特征转化为图像特征序列。经过IoU-aware感知查询选择固定数量的图像特征作为解码器的初始对象查询。最后带有辅助预测头的解码器迭代优化对象查询以生成框和置信度分数。
RT-DETR模型中,输入图像经过主干网络后输出的3个特征图大小分别为20×20,40×40和80×80,分别用来检测大、中以及小目标。针对航拍图像中小目标较多且分布集中的特点,提出在特征融合部分增加一层针对小目标的检测结构。由于浅层特征相比深层特征而言具有更丰富的位置信息及细节特征信息,更有利于小目标的检测,因此在主干网络中提取出160×160大小的特征图并将其与更深层次的图像特征进行特征融合,以实现对小目标的精准定位与捕捉(方法一)。然而这种方法虽然在性能上有一定的提升,但网络的计算开销也随之增大,基于此,提出删除20×20大目标检测层及主干网络中对应的卷积下采样操作,同时在特征融合模块去除该层相关的特征融合操作,在极大降低参数量和计算量的同时仍能够使模型保持较好的检测效果(方法二)。
模型结构对比如图2所示,方法一为仅添加小目标检测层的模型,方法二为提出的模型,在去除P5相关层后将P4送入Transformer Encoder得到F4,然后通过跨尺度特征融合模块增强特征的表达能力,提高小目标的检测性能和模型的收敛速度。模型效果对比如表1所示,可以看出提出的结构在不影响模型速度的前提下参数量相比原模型结构获得了极大的降低,同时检测精度大幅提高。同时和方法一相比在相对检测精度仅降低0.2%的情况下极大减少了模型的参数和计算量,提高了模型检测速度,同时准确率也有了部分提升,实现了更高效的特征获取。
无人机航拍图像中目标尺寸小、分辨率低,由于主干网络下采样过程中卷积的不断堆叠,在深层特征图中虽然能获得更多的全局语义信息,但无法充分捕捉到小目标的细节特征。同时考虑到模型应用于无人机图像目标检测领域,在实际部署时速度与准确性同样重要。目前来看实时性较强的模型一般采用GConv(group convlution)[21]或者DWConv(depthwise convolution)[22]的方式实现更高的效率[23],如MobileNets[24],ShuffleNets[25],GhostNet[26],EffcientNets[27]。DWConv这种拆分通道并对每个通道分别进行卷积的方式虽然有效减少了FLOPs(每秒浮点运算次数),却因无法捕捉通道间的关系而导致全局语义的丢失,因此通常以DWConv+PWConv(pointwise convolution)的组合形式出现,PWConv通过对生成特征图进行通道间的加权组合,生成新的特征图以更好的捕捉图像中的语义信息和特征关联性。然而在实践中通道数的增加又会导致更高的内存访问,降低整体计算速度。
在主干网络下采样过程中丰富多尺度的特征信息更有利于对不同尺寸目标进行特征提取,但同时会带来参数量和计算量的增加。观察到不同通道的特征图具有高度的相似性,存在大量冗余[23,26,28]。由此提出了一种新的MCPConv,在减少特征冗余的同时提高模型对不同尺度特征的感知能力。首先对输入特征的1/4通道进行线性变换以进行空间特征提取,将该通道分别送入一个3×3的卷积核和一个5×5的卷积核(为了降低参数量及计算量这里使用两个3×3的卷积核替代并通过跳跃连接的方式实现),获取不同尺度的特征信息,融合多尺度的目标特征,经过ReLU激活函数,最后将经过特征提取后的通道与未经处理的3/4通道进行拼接得到最终输出。MCPConv结构如图3所示。
假设hw为输入数据的高度和宽度,c为输入通道数,cp为部分输入通道数,k为卷积核大小,标准卷积和多通道特征部分卷积的FLOPs及Params计算过程如下。
FLOPs1=hwk2c2
FLOPs2=2hwk2 c p 2
Params1=k2c2
Params2=2k2 c p 2
cpc的1/4,则MCPConv的FLOPs为标准卷积的1/8,其Params为普通卷积的1/8,有效地降低了模型所需的计算资源。
在RT-DETR模型中,主干网络的下采样过程是通过BasicBlock残差结构完成的,该结构由于卷积的堆叠导致冗余计算余较多,因此基于本文提出的MCPConv,重新设计了Basic Block,命名为MCP Block,其结构如图4所示。
与原始的Basic Block相比,使用MCPConv对常规卷积进行替换,在其后添加两个1×1大小的卷积,并且只在中间层使用了BN以及激活函数,在保持特征多样性的同时保证较低的延迟。设计的MCP Block有效降低了计算量并丰富了下采样过程中的特征信息,实现了更高效的特征获取。
目前主流的Transformer方法主要依靠不可学习的相对/绝对位置编码描述像素或patch序列的先验位置关系。然而在小目标检测任务中目标往往在图像中占据的像素数量较少,因此其定位的准确度对检测结果至关重要。使用可学习位置编码能够使模型更准确地感知图像中目标的绝对位置信息,从而提高对小目标检测的准确性。
使用1D Position Embedding,随机初始化生成可学习的位置编码,与输入特征图相加后输入网络中,通过反向传播算法,更新位置编码参数。
由于IoU在小目标检测时对像素的位置偏差过于敏感,在目标尺度足够小时,轻微的边界框偏差就会导致IoU的急剧降低,从而导致对小目标的泛用性极差。因此引入NWD和MPDIoU损失相结合的方式替代GIoU。
Wang等[18]提出的NWD损失对不同尺度的物体不敏感,对边界框的重叠要求更低,甚至没有重叠也可以度量分布之间的相似性,因此更适用于小目标的检测。该方法先将边界框建模为2D高斯分布,然后使用归一化Wasserstein距离衡量高斯分布的相似性。首先,对于水平边界框R=(Cx,Cy,h,w),其中(Cx,Cy)、wh分别表示中心坐标、宽度和高度。将其建模为高斯分布N(μ,σ),表达式为
μ = C x C y σ = - w 2 4 0 0 h 2 4
将边界框AB之间的相似度转换为两个高斯分布之间的距离,两个二维高斯分布μ1=N(m1,σ1)和μ2=N(m2,σ2),μ1μ2之间的二阶Wasserstein距离为
$\begin{aligned} W_{2}^{2}\left(\boldsymbol{\mu}_{1}, \boldsymbol{\mu}_{2}\right)= & \left\|\boldsymbol{m}_{1}-\boldsymbol{m}_{2}\right\|_{2}^{2}+\operatorname{Tr}\left[\boldsymbol{\sigma}_{1}+\boldsymbol{\sigma}_{2}-\right. \\ & \left.2\left(\boldsymbol{\sigma}_{2}^{1 / 2} \boldsymbol{\sigma}_{1} \boldsymbol{\sigma}_{2}^{1 / 2}\right)^{1 / 2}\right] \end{aligned}$
对于由边界框A= c x a , c y a , w a 2 , h a 2B= c x b , c y b , w b 2 , h b 2建模的高斯分布NaNb,最终简化为
$\begin{aligned} W_{2}^{2}\left(N_{a}, N_{b}\right)= & \Bigg\|\left(\left[c x_{a}, c y_{a}, \frac{w_{a}}{2}, \frac{h_{a}}{2}\right]^{\mathrm{T}},\right. \\ & {\left.\left[c x_{b}, c y_{b}, \frac{w_{b}}{2}, \frac{h_{b}}{2}\right]^{\mathrm{T}}\right) \Bigg\| _{2}^{2} } \end{aligned}$
然而 W 2 2(Na,Nb)是距离度量,不能直接用作相似性度量,对其进行归一化,得到最终的NWD损失表达式为
NWD(Na,Nb)=exp - W 2 2 ( N a , N b ) C
式(8)中:C为常量。
NWD损失提高了模型对小目标的检测精度,但会对模型的收敛速度产生一定影响。同步使用MPDIoU,MPDIoU简化了边界框之间的相似性比较,并且对边界框重叠与不重叠的情况都能很好处理,因此对小目标场景下检测的适用性较好,同时能够加快模型的收敛速度。MPDIoU的计算过程如下。
MPDIoU=IoU- d 1 2 h 2 + w 2- d 2 2 h 2 + w 2
式(9)中: d 1 2 d 2 2分别为预测边界框与真值边界框左上角点和右下角点坐标差值的平方之和,即左上角点和右下角点的距离的平方。
最终将NWD损失和MPDIoU结合得到损失函数表达式为
loss=a(1-MPDIoU)+(1-a)(1-NWD)
经实验对比,其中a取0.7时效果最优。
实验平台为Ubuntu20.04系统,内存为24 GB,GPU为NVIDIA RTX4090,Python版本为3.8,PyTorch版本为1.13.1,CUDA版本为11.7。网络未使用预训练模型权重,在训练和测试阶段采用640×640的图像输入大小,实验设置200个epoch,bathsize为4,使用adamw优化器,学习率为0.000 1。同时为了保证实验的一致性,在测试阶段模型的bathsize均设置为1。
采用天津大学机器学习与数据挖掘实验室AISKYEYE收集的VisDrone2019数据集进行实验,共包含8 629张图像,其中训练集包含6 471张图片,验证集548张,测试集1 610张。该数据集包含10个种类的目标,分别为行人、人、自行车、汽车、面包车、卡车、三轮车、遮阳棚三轮车、巴士、摩托车。
实验采用Params、GFLOPs、mAP50、mAP50:95和FPS作为模型性能的评价指标。mAP、FPS的计算公式如下。
P= T P T P + F P×100%
R= T P T P + F N×100%
AP= 0 1 P(r)dr
mAP= 1 n i = 1 nAPi
FPS= 1000 t
式中:P为准确率;R为召回率;n为检测类别数,TP、FP、FN、TN分别为实际为正样本检测为正样本、实际为负样本检测为正样本、实际为正样本检测为负样本、实际为负样本检测为负样本;mAP50为IoU阈值为0.5时所有类别的平均检测精度(mAP);FPS为模型1 s内推理的图片数量,用来表示模型的推理速度,帧/s;t为时间,s。
为了验证所提每个改进部分的有效性,对每个改进分别在VisDrone2019-DET数据集上进行消融实验,并在测试集上验证其对小目标的检测性能,实验结果如表2所示。可以看出,在加入MCP Block后,由于残差结构中计算量减少并且过程中融合了多尺度的特征信息,显著降低了模型的参数量和计算量,并且检测精度提升0.9%。AIFI-LPE模块提供了可学习位置编码使模型能够更加准确的感知输入图像中的位置信息,使模型精度提升1.2%。在引入NWD和MPDIoU后模型对小目标的定位效果更佳,检测精度提升0.9%。最后加入提出的多尺度特征融合模块后,虽然模型计算量有少量增加,但参数量降低了60%,并且精度提升2.5%。从实验结果来看,本文模型在检测精度和网络体积上取得了恰当的平衡,更满足无人机航拍图像目标检测任务的要求。
为了验证改进模型的检测性能,将提出的算法与其他实时端到端目标检测器进行比较,实验结果如表3所示。在VisDrone数据集上,本文算法的AP达到40.5%,FPS达到70帧/s,相较于同规模甚至更大规模的YOLO系列算法,本文算法能够在很好地满足实时性(FPS>60帧/s)的同时,极大降低了模型的参数量,且大幅提高了对小目标的检测精度,相较于其他最新的优秀模型具有更好的检测性能。
为了直观呈现本文模型在无人机航拍图像目标检测这一任务中的表现,将部分目标检测效果进行对比展示。如图5所示,上排图像为RT-DETR模型的检测效果,下排为本文模型的检测效果,可以看出本文模型在远距离微小目标检测上展现出了更好的效果,极大提高了对目标的捕获能力,同时显著降低了原模型的漏检以及误检问题。本文模型在微小目标检测任务中表现出较强的能力,在对小目标精准定位的同时能够准确对其类别进行区分,展现了出色的鲁棒性和精确度。
针对无人机航拍图像中小目标检测中常见的漏检、误检等问题,提出了一种针对无人机航拍场景的小目标检测算法。首先,提出多尺度特征融合模块,并增加针对微小目标的检测层,以此增强网络对小目标的感知能力。同时为了防止模型过大,优化主干网络,去除了一个残差块。其次,在主干网络中使用MCP Block替换了传统的Basic Block,使模型更加轻量化的同时,提升对小目标多尺度特征的提取能力,有效减少了特征冗余。此外,通过引入可学习位置编码,模型对复杂背景下小目标的精准定位能力得到了加强。最后,提出NWD和MPDIoU相结合的损失函数,在降低模型对位置偏差敏感度的同时加快了模型的收敛速度。实验结果表明,本文算法在Visdrone2019数据集上取得了良好的效果,领先于目前大多数主流目标检测算法,在实时无人机航拍目标检测领域有更大的优势。
研究在无人机技术领域的发展中发挥了积极作用。通过技术创新,提出的算法显著提升了无人机航拍图像中小目标的检测精度和实时处理能力,同时实现了模型计算成本的降低。这使无人机在资源受限的条件下依然能够执行高效的目标识别任务。这对于农业现代化、地理测绘、环境监测等关键领域的发展具有重要的作用,同时也对加强国防安全具有积极影响,为中国无人机在复杂地理环境和多变气候条件下的应用提供了坚实的技术基础。
  • 河北省省级科技计划(SZX2020034)
  • 河北省高等学校科学研究计划(BJK2022041)
参考文献 引证文献
排序方式:
[1]
Girshick R, Donahue J, Darrell T, et al. Rich feature hierarchies for accurate object detection and semantic segmentation[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Columbus: IEEE, 2014: 580-587.
[2]
Redmon J, Divvala S, Girshick R, et al. You only look once: unified, real-time object detection[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas: IEEE, 2016: 779-788.
[3]
桑雨, 李立权, 李铁. 轻量化YOLOv7-tiny的遥感图像小目标检测[J]. 科学技术与工程, 2024, 24(18): 7726-7732.
Sang Yu, Li Liquan, Li Tie, et al. Lightweight YOLOv7-tiny for remote sensing image small target detection[J]. Science Technology and Engineering, 2024, 24(18): 7726-7732.
[4]
刘晋川, 黎向锋, 刘安旭, 等. 改进RetinaNet的无人机小目标检测[J]. 科学技术与工程, 2023, 23(1): 274-282.
Liu Jinchuan, Li Xiangfeng, Liu Anxu, et al. Improved RetinaNet for unmanned aerial vehicle small target detection[J]. Science Technology and Engineering, 2023, 23(1): 274-282.
[5]
吴稳稳, 吴晓红, 刘强, 等. 基于全局注意力的多级特征融合目标检测算法[J]. 科学技术与工程, 2020, 20(27): 11185-11191.
Wu Wenwen, Wu Xiaohong, Liu Qiang, et al. Multi-level feature fusion object detection algorithm based on global attention[J]. Science Technology and Engineering, 2020, 20(27): 11185-11191.
[6]
康传利, 张思瑶, 李玄皓, 等. 高斯Wasserstein距离改进轻量YOLOv7模型的遥感影像道路交叉口检测[J]. 科学技术与工程, 2024, 24(9): 3533-3542.
Kang Chuanli, Zhang Siyao, Li Xuanhao, et al. Gaussian Wasserstein distance improvement of lightweight YOLOv7 model for remote sensing image road intersection detection[J]. Science Technology and Engineering, 2024, 24(9): 3533-3542.
[7]
Tan M, Le Q. Efficientnet: rethinking model scaling for convolutional neural networks[J]. ArXiv, 2019: 1905.11946.
[8]
Carion N, Massa F, Synnaeve G, et al. End-to-end object detection with transformers[C]// European Conference on Computer Vision. Cham: Springer International Publishing, 2020: 213-229.
[9]
Zhu X, Su W, Lu L, et al. Deformable detr: deformable transformers for end-to-end object detection[J]. ArXiv Preprint ArXiv, 2020: 2010.04159.
[10]
Dai J, Qi H, Xiong Y, et al. Deformable convolutional networks[C]// Proceedings of the IEEE International Conference on Computer Vision. Venice: IEEE, 2017: 764-773.
[11]
Meng D, Chen X, Fan Z, et al. Conditional detr for fast training convergence[C]// Proceedings of the IEEE/CVF International Conference on Computer Vision. Montreal: IEEE, 2021: 3651-3660.
[12]
Wang Y, Zhang X, Yang T, et al. Anchor DETR: query design for Transformer-based detector[C]// Proceedings of the AAAI Conference on Artificial Intelligence. New York: AAAI, 2022, 36(3): 2567-2575.
[13]
Liu S, Li F, Zhang H, et al. DAB-DETR: dynamic anchor boxes are better queries for DETR[J]. ArXiv Preprint ArXiv, 2022: 2201.12329.
[14]
Li F, Zhang H, Liu S, et al. DN-DETR: accelerate DETR training by introducing query denoising[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022, 46(4): 13619-13627.
[15]
Zhang H, Li F, Liu S, et al. Dino: DETR with improved denoising anchor boxes for end-to-end object detection[J]. ArXiv Preprint ArXiv, 2022: 2203.03605.
[16]
W, Xu S, Zhao Y, et al. DETRs beat yolos on real-time object detection[J]. ArXiv Preprint ArXiv, 2023: 2304.08069.
[17]
Lin T Y, Maire M, Belongie S, et al. Microsoft COCO: common objects in context[C]// Computer Vision-ECCV 2014: 13th European Conference. Zurich: Springer International Publishing, 2014: 740-755.
[18]
Wang J, Xu C, Yang W, et al. A normalized Gaussian Wasserstein distance for tiny object detection[J]. ArXiv Preprint ArXiv, 2021: 2110.13389.
[19]
Siliang M, Yong X. Mpdiou: a loss for efficient and accurate bounding box regression[J]. ArXiv Preprint ArXiv, 2023: 2307.07662.
[20]
He K, Zhang X, Ren S, et al. Deep residual learning for image recognition[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Chengdu: IEEE, 2016: 770-778.
[21]
Krizhevsky A, Sutskever I, Hinton G E. ImageNet classification with deep convolutional neural networks[J]. Advances in Neural Information Processing Systems, 2012, 25.
[22]
Sifre L, Mallat S. Rigid-motion scattering for texture classification[J]. ArXiv Preprint ArXiv, 2014: 1403.1687.
[23]
Chen J, Kao S, He H, et al. Run, don't walk: chasing higher FLOPS for faster neural networks[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Vancouver: IEEE, 2023: 12021-12031.
[24]
Howard A, Sandler M, Chu G, et al. Searching for MobileNetV3[C]// Proceedings of the IEEE/CVF International Conference on Computer Vision. Seoul: IEEE, 2019: 1314-1324.
[25]
Zhang X, Zhou X, Lin M, et al. ShuffleNet: an extremely efficient convolutional neural network for mobile devices[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Salt Lake City: IEEE, 2018: 6848-6856.
[26]
Han K, Wang Y, Tian Q, et al. GhostNet: more features from cheap operations[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Seattle: IEEE, 2020: 1580-1589.
[27]
Tan M, Le Q. EfficientNet: rethinking model scaling for convolutional neural networks[J]. ArXiv, 2019: 1905.11946.
[28]
Zhang Q, Jiang Z, Lu Q, et al. Split to be slim: an overlooked redundancy in Vanilla convolution[J]. ArXiv Preprint ArXiv, 2020: 2006.12085.
2025年第25卷第13期
PDF下载
388
152
引用本文
BibTeX
文章信息
doi: 10.12404/j.issn.1671-1815.2404266
  • 接收时间:2024-06-07
  • 首发时间:2025-07-09
  • 出版时间:2025-05-08
补充材料
相关文章
文章信息
作者
出版历史
  • 收稿日期:2024-06-07
  • 修回日期:2025-01-16
基金
河北省省级科技计划(SZX2020034)
河北省高等学校科学研究计划(BJK2022041)
作者信息
    1 河北经贸大学管理科学与信息工程学院, 石家庄 050062
    2 河北省跨境电商技术创新中心, 石家庄 050062

通讯作者:

* 李卫东(1973—),男,汉族,河北石家庄人,博士,教授。研究方向:机器学习、深度学习。E-mail:
参考文献
分享链接
https://castjournals.cast.org.cn/joweb/kxjsygc/CN/10.12404/j.issn.1671-1815.2404266
分享至
全文二维码

扫描看全文

引用本文
BibTeX
本文的引用情况
2种不同金属材料的力学参数

Family
属数
Number of
genus
种数
Number of
species
占总种数比例
Percentage of
total species (%)

Genus
种数
Number of
species
占总种数比例
Percentage of total
species (%)
鹅膏菌科Amanitaceae 2 11 5.26 鹅膏菌属 Amanita 10 4.78
小菇科 Mycenaceae 2 12 5.74 丝盖伞属 Inocybe 5 2.39
多孔菌科 Polyporaceae 8 14 6.70 蜡蘑属 Laccaria 5 2.39
红菇科 Russulaceae 3 23 11.00 小皮伞属 Marasmius 6 2.87
小菇属 Mycena 11 5.26
光柄菇属 Pluteus 5 2.39
红菇属 Russula 17 8.13
栓菌属 Trametes 5 2.39
关闭全屏