文献类型：期刊文献

英文题名：Improved YOLOX-X based UAV aerial photography object detection algorithm

作者：Wang, Xin[1];He, Ning[1];Hong, Chen[2];Wang, Qi[2];Chen, Ming[2]

通讯作者：He, N[1]

机构：[1]Beijing Union Univ, Coll Smart City, Beijing 100101, Peoples R China;[2]Beijing Union Univ, Coll Robot, Beijing 100101, Peoples R China

第一机构：北京联合大学

通讯机构：[1]corresponding author), Beijing Union Univ, Coll Smart City, Beijing 100101, Peoples R China.|[11417]北京联合大学;

年份：2023

卷号：135

外文期刊名：IMAGE AND VISION COMPUTING

收录：;EI(收录号：20232214173867);Scopus(收录号：2-s2.0-85160515858);WOS:【SCI-EXPANDED(收录号:WOS:001011531900001)】；

基金：This work is supported by the National Natural Science Foundation of China (62272049, 61972375, 62172045) , the Key Project of Beijing Municipal Commission of Education (KZ201911417048) , the National Key R & D Program of China (No. 2018AAA0100804) , Premium Funding Project for Academic Human Resources Development in Beijing Union University (BPHR2020AZ01, BPH2020EZ01) , the Science and Technology Project of Beijing Municipal Commission of Education (KM202111417009, KM201811417005) .

语种：英文

外文关键词：UAV aerial photography; Object detection; YOLOX; Small objects

摘要：Unmanned Aerial Vehicle (UAV) aerial photography object detection has high research significance in the fields of disaster rescue, ecological environmental protection, and military reconnaissance. The larger width of UAV photography introduces background interference into the detection task, whereas the relatively high imaging height of the UAV results in mostly small objects in the aerial images. YOLOX-X operated fast and achieved ad-vanced results on MS COCO of natural scene images, so YOLOX-X was used as the baseline network in this paper. A UAV aerial photography object detection algorithm YOLOX_w with improved YOLOX-X is proposed to handle the characteristics of complex backgrounds and the large number of small objects in UAV aerial photog-raphy images. The model's performance in detecting small objects is first improved by preprocessing the training set with the slicing aided hyper inference (SAHI) algorithm and by data augmentation. Then, a shallow feature map with rich spatial information is introduced into the path aggregation network (PAN), and a detection head is added to detect small objects. Next, the ultra-lightweight subspace attention module (ULSAM) is added to the PAN stage to highlight the target features and weaken the background features, which improves the detection accuracy of the network. Finally, the loss function of the bounding box regression is optimized to further improve network prediction accuracy. Experimental results on the VisDrone dataset demonstrate that the detection accuracy of the proposed YOLOX_w algorithm improved by 8% when compared with the baseline YOLOX-X. Moreover, migration experiments on the DIOR dataset verify the effectiveness and robustness of the improved method.& COPY; 2023 Elsevier B.V. All rights reserved.