跨尺度特征融合的遙感微小目標(biāo)檢測(cè)算法

打開文本圖片集

中圖分類號(hào)：TP751.2 文獻(xiàn)標(biāo)志碼：A DOI：10.12305/j.issn.1001-506X.2025.05.05

Remote sensing small target detection algorithm based on cross-scale feature fusion

SHAO Kai1，2，3，*，LI Haogang1，LIANG Yan1，NING Jing1，CHEN Wul （1. SchoolofCommunicationand Information Engineering，Chongqing Universityof Posts and Telecommunications， Chongqing40oo65，China；2.ChongqingKeyLaboratoryof MobileCommunications Technology，Chongqing University ofPostsandTelecommunications，Chongqing465，China；3.EngineeringResearchCenterofMobileCommunications oftheMinistryofEducation，Chongqing Universityof Posts and Telecommunications，Chongqing 40oo65，China）

Abstract：For three problems of shalow thinning features，deep semantic representation and multi-scale information extraction for the detection of small targets in remote sensing images，a crossscale YOLOv7 （CSYOLOv7） network by comprehensively applying multiple technologies is proposed. Firstly，a cro-stage feature extraction module（CFEM） and a receptive field feature enhancement module（RFFEM） are designed. CFEMis to improve the model'sability of refining feature extraction and suppress the loss of features during shallow down-sampling.RFFEM is to increase the network's ability of extracting deep semantic features and improve the model’sabilityof acquiring target context information.Secondly，a cross-gradient space pyramid pool module（CSPPM）is designed to effectively fuse global multi-scaleand local features of smalltargets. Finall，shape intersection over union（Shape IoU） is used to replace the complete intersection over union （CIoU）to improve the accuracy of the model in the bounding box positioning task.Experimental results show that the CSYOLOv7 network achieves detection accuracy of 74% and 89.6% on the Dataset for Image Object Recognition（DIOR）data setand Northwestern Polytechnical University Very High Resolution-10（NWPUVHR-10） data setrespectively，which efectively improves the detection efect of smal targets in remote sensing images.

Keywords：remote sensing image；small target； feature extraction；context information

0 引言

光學(xué)遙感圖像目標(biāo)檢測(cè)是航空和衛(wèi)星圖像分析領(lǐng)域的基礎(chǔ)任務(wù)，旨在對(duì)圖像中的感興趣目標(biāo)進(jìn)行分類和定位]，可以有效地提取和識(shí)別圖像中的重要信息，從而在自然災(zāi)害檢測(cè)、軍事目標(biāo)偵察、城市交通疏導(dǎo)等領(lǐng)域有著廣泛的用途。（剩余17770字）