文献类型：期刊文献

英文题名：A real-time detector for small-object remote sensing

作者：Wang, Xin[1];Xu, Guangmei[1];Hong, Chen[2];He, Ning[1];Li, Runjie[1];Sun, Fengxi[1];Han, Wenjing[1]

机构：[1]Beijing Union Univ, Coll Smart City, 97 North Fourth Ring East Rd, 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

年份：2025

卷号：125

外文期刊名：COMPUTERS & ELECTRICAL ENGINEERING

收录：;EI(收录号：20252118459828);Scopus(收录号：2-s2.0-105005402461);WOS:【SCI-EXPANDED(收录号:WOS:001497931500003)】；

基金：This work is supported by the National Natural Science Foundation of China (62272049, 62236006, 62172045) , the Key Project of Beijing Municipal Commission of Education, China (KZ201911417048) , the Major Project of Technological Innovation 2030-"New Generation Artificial Intelligence", China (2018AAA0100800) , the Science and Technology Project of Beijing Municipal Commission of Education, China (KM202111417009,KM201811417005) , the Academic Research Projects of Beijing Union University, China (No. ZKZD202301) .

语种：英文

外文关键词：Object detection; Remote sensing images; Small object; Real-time

摘要：In the field of object detection, small-object detection has always been a difficult task. Remote sensing images have complex backgrounds and small objects can be densely distributed. Moreover, remote sensing detection must meet real-time requirements. To address these challenges, this paper proposes a detector called NanoDet-Drone for the real-time detection of small objects in remote sensing scenes. The baseline model lacks a sufficient receptive field to capture both local and long-distance information, and cannot achieve satisfactory detection results when directly applied to remote sensing detection. Our project improves the baseline network. First, the receptive field module is proposed, which uses dilated convolution at different dilation rates to expand the model's receptive field while fully exploiting the contextual information of the small objects, incorporating the coordinate attention mechanism to highlight the features of small objects. Then, the adaptive fusion feature pyramid network (AF-FPN) is proposed to reasonably fuse the features of different branches; this efficiently uses multi-scale features and provides the network with more detailed information about small objects. Finally, the improved training auxiliary module, called the assign guidance module, is used to guide the detection head training and help the network learn richer feature representations to improve the accuracy and robustness of the model. In this study, we conducted extensive experiments on two challenging remote sensing datasets, VisDrone and AI-TOD, to demonstrate the effectiveness and robustness of NanoDet-Drone. Results show that NanoDet-Drone is capable of running at 56.8 frames per second on a CPU, outperforming other advanced detectors (YOLOv9-T and YOLOv10-N) at the same scale. Our model achieves a better trade-off between accuracy and inference speed. The proposed AF-FPN can be easily embedded into a one-stage detector, which effectively improves detection performance while significantly reducing the number of model parameters and computations. Compared with the baseline, NanoDet-Drone increased average precision (AP) and AP0.5 by 5.2% and 8.6%, respectively, on VisDrone, and increased AP and AP0.5 by 4.8% and 10.9%, respectively, on AI-TOD.