文献类型：期刊文献

英文题名：Lightweight Small Target Detection Algorithm with Multi-Feature Fusion

作者：Yang, Rujin[1];Zhang, Jingwei[2];Shang, Xinna[1,3];Li, Wenfa[1,4]

第一作者：Yang, Rujin

通讯作者：Shang, XN[1];Li, WF[1];Shang, XN[2];Li, WF[3]

机构：[1]Beijing Union Univ, Beijing Key Lab Informat Serv Engn, Beijing 100101, Peoples R China;[2]Southern Fed Univ, Vorovich Inst Math Mech & Comp Sci, Rostov Na Donu 344090, Russia;[3]Beijing Union Univ, Coll Robot, Beijing 100101, Peoples R China;[4]Univ Sci & Technol Beijing, Inst Artificial Intelligence, Beijing 100083, 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 100101, Peoples R China;[2]corresponding author), Beijing Union Univ, Coll Robot, Beijing 100101, Peoples R China;[3]corresponding author), Univ Sci & Technol Beijing, Inst Artificial Intelligence, Beijing 100083, Peoples R China.|[1141739]北京联合大学机器人学院;[11417]北京联合大学;[11417103]北京联合大学北京市信息服务工程重点实验室;

年份：2023

卷号：12

期号：12

外文期刊名：ELECTRONICS

收录：;EI(收录号：20230148299);Scopus(收录号：2-s2.0-85163825062);WOS:【SCI-EXPANDED(收录号:WOS:001014134000001)】；

基金：This research was funded by the National Natural Science Foundation of China (Grant Nos. 61972040).

语种：英文

外文关键词：UAV target detection; lightweighting; GhostNet module; EIoU loss; K-means++

摘要：Unmanned aerial vehicles (UAVs) are a highly sought-after technology with numerous applications in both military and non-military uses. The identification of targets is a crucial aspect of UAV applications, but there are challenges associated with complex detection models and difficulty in detecting small targets. To address these issues, this study proposes the lightweight L-YOLO algorithm for target detection tasks from a UAV perspective. The L-YOLO algorithm improves on YOLOv5 by improving the model's detection performance for small targets while reducing the number of parameters and computational effort. The GhostNet module replaces the relevant convolution in the YOLOv5 model to create a lightweight model. The EIoU loss is used as the loss function of the algorithm to accelerate convergence and improve regression accuracy. Furthermore, feature-level extensions based on YOLOv5 are implemented, and a new detection head is proposed to improve the model's detection accuracy for small targets. The size of the anchor boxes is redesigned to suit the small targets using the K-means++ clustering algorithm. The experiments were conducted on the VisDrone-2022 dataset, and the L-YOLO algorithm demonstrated a reduction in computational effort by 42.42% and number of parameters by 48.6% compared to the original algorithm. Furthermore, recall and mAP@0.5 improved by 2.1% and 1.4%, respectively. These results demonstrate that the L-YOLO algorithm not only has better detection performance for small targets but is also a lighter model, indicating promising prospects for target detection from a UAV perspective.