YOLO‑v8 with Multidimensional Attention and Upsampling Fusion for Small Air Target Detection in Radar Images

作     者：JIANG Zhenyu LI Xiaodong DU Chen CHEN An HAN Yanqiang LI Jinjin 江振宇;李晓东;杜晨;陈安;韩彦强;李金金

作者机构：College of Aerospace ScienceNational University of Defense TechnologyChangsha 410073P.R.China National Key Laboratory of Advanced Micro and Nano Manufacture TechnologyShanghai Jiao Tong UniversityShanghai 200240P.R.China Department of Micro/Nano ElectronicsSchool of Electronic Information and Electrical EngineeringShanghai Jiao Tong UniversityShanghai 200240P.R.China 

出 版 物：《Transactions of Nanjing University of Aeronautics and Astronautics》 (南京航空航天大学学报(英文版))

年 卷 期：2024年第41卷第6期

页      面：710-724页

核心收录：

学科分类：0810[工学-信息与通信工程] 08[工学] 081002[工学-信号与信息处理] 

基　　金：supported by the Na‑tional Natural Science Foundation of China Joint Fund(No.U21B2028) the National Key R&D Program of China(No.2021YFC 2100100) the Shanghai Science and Technology Project(Nos.21JC1403400,23JC1402300) 

主　　题：YOLO radar images object detection machine learning 

摘      要：This study presents an innovative approach to improving the performance of YOLO-v8 model for small object detection in radar ***,a local histogram equalization technique was applied to the original images,resulting in a notable enhancement in both contrast and detail ***,the YOLO-v8 backbone network was augmented by incorporating convolutional kernels based on a multidimensional attention mechanism and a parallel processing strategy,which facilitated more effective feature information *** the model’s head,an upsampling layer was added,along with the fusion of outputs from the shallow network,and a detection head specifically tailored for small object detection,thereby further improving ***,the loss function was modified to incorporate focal-intersection over union(IoU)in conjunction with scaled-IoU,which enhanced the model’s performance.A weighting strategy was also introduced,effectively improving detection accuracy for small *** results demonstrate that the customized model outperforms traditional approaches across various evaluation metrics,including recall,precision,F1-score,and the receiver operating characteristic(ROC)curve,validating its efficacy and innovation in small object detection within radar *** results indicate a substantial improvement in accuracy compared to conventional methods such as image segmentation and standard convolutional neural networks.