文献类型：期刊文献

英文题名：DBYOLOv8: Dual-Branch YOLOv8 Network for Small Object Detection on Drone Image

作者：Tan, Yawei[1];Xu, Bingxin[1];Sun, Jiangsheng[2];Xu, Cheng[1];Pan, Weiguo[1];Dai, Songyin[1];Liu, Hongzhe[1]

第一作者：Tan, Yawei

通讯作者：Tan, YW[1]

机构：[1]Beijing Union Univ, Beijing Key Lab Informat Serv Engn, Beijing, Peoples R China;[2]Army Res Acad, Sci & Technol Innovat Res Ctr, Beijing, Peoples R China

第一机构：北京联合大学北京市信息服务工程重点实验室

通讯机构：[1]corresponding author), Beijing Union Univ, Beijing Key Lab Informat Serv Engn, Beijing, Peoples R China.|[11417103]北京联合大学北京市信息服务工程重点实验室;[11417]北京联合大学;

年份：2025

卷号：16

期号：1

起止页码：1301-1309

外文期刊名：INTERNATIONAL JOURNAL OF ADVANCED COMPUTER SCIENCE AND APPLICATIONS

收录：EI(收录号：20250617818345);Scopus(收录号：2-s2.0-85216853736);WOS:【ESCI(收录号:WOS:001435368500001)】；

语种：英文

外文关键词：Drone images; dual-branch; small object detection; YOLOv8

摘要：Object detection based on drone platforms is a valuable yet challenging research field. Although general object detection networks based on deep learning have achieved breakthroughs in natural scenes, drone images in urban environments often exhibit characteristics such as a high proportion of small objects, dense distribution, and significant scale variations, posing significant challenges for accurate detection. To address these issues, this paper proposes a dual-branch object detection algorithm based on YOLOv8 improvements. Firstly, an auxiliary branch is constructed by extending the YOLOv8 backbone to aggregate high-level semantic information within the network, enhancing the feature extraction capability. Secondly, a MultiBranch Feature Enhancement (MBFE) module is designed to enrich the feature representation of small objects and enhance the correlation of local features. Third, Spatial-to-Depth Convolution (SPDConv) is utilized to mitigate the loss of small object information during downsampling, preserving more small object feature information. Finally, a dual-branch feature pyramid is designed for feature fusion to accommodate the dual-branch input. Experimental results on the VisDrone benchmark dataset demonstrate that DBYOLOv8 outperforms state-of-the-art object detection methods. Our proposed DBYOLOv8s achieve mAP@0.5 of 49.3% and mAP@0.5:0.95 of 30.4%, which are 2.8% and 1.5% higher than YOLOv9e, respectively.