联合目标分割和关键点检测的飞机型号识别  

Aircraft type recognition method by integrating target segmentation and key points detection

作　　者：刘思婷 王庆栋[1] 张力[1] 韩晓霞 王保前 刘玉贤 LIU Siting;WANG Qingdong;ZHANG Li;HAN Xiaoxia;WANG Baoqian;LIU Yuxian(Chinese Academy of Surveying and Mapping,Beijing 100830,China;Faculty of Geomatics,Lanzhou Jiaotong University,Lanzhou 730070,China;Shenzhen Investigation&Research Institute Co.,Ltd,Shenzhen 518026,China)

机构地区：[1]中国测绘科学研究院,北京100830 [2]兰州交通大学测绘与地理信息学院,兰州730070 [3]深圳市勘察研究院有限公司,深圳518026

出　　处：《遥感学报》2024年第4期1010-1024,共15页NATIONAL REMOTE SENSING BULLETIN

基　　金：国家重点研发计划(编号:2019YFB1405600);深圳市技术攻关(编号:JSGG20191129103003903)。

摘　　要：目前,受限于数据集精细度与网络结构,深度学习技术仍难以应对飞机目标型号识别这类精细化识别任务。本文针对遥感影像中飞机目标型号识别问题,提出一种融合目标分割与关键点检测的飞机型号识别方法。该方法有机地结合多任务深度神经网络与条件随机场和模板匹配算法,利用“预训练+微调+后处理”的方式实现飞机型号的高精度识别。首先,基于多任务深度神经网络迁移学习技术实现飞机目标物位置、掩膜与关键点信息识别。其次,为了便于后期高精度模板匹配,利用本文提出的融合条件随机场的飞机目标掩膜精化算法和基于关键点的姿态调整算法,实现识别目标的边界精细化与机体姿态调整;最后,在本文构建的飞机型号模板库基础上,将经过精化后处理的飞机掩膜信息与模板库进行匹配,实现飞机目标的型号识别。为了验证所提方法的有效性,本文进行了相关实验,并与传统算法及完全端到端深度学习方法进行了对比,结果表明,本文所提方法具有更高准确率,并且在实用性方面更具优势。Aircraft detection via deep learning is a popular field in remote sensing image analysis.However,given the limited perspectives of satellite imagery and high similarities in image appearance,aircraft type recognition remains a challenging task.The existing deep learning methods cannot be satisfactorily applied to fine-grained aircraft type recognition tasks,which require refined labels for datasets.With the aim of effectively recognizing aircraft types in remote sensing images,we propose an integrated target segmentation and key point detection method for aircraft type recognition.The proposed method combines an organic multitask deep neural network with a conditional random field and template matching algorithm to achieve high-precision recognition of aircraft types by pretraining,fine-tuning,and postprocessing.First,we performed target aircraft position and mask and keypoint recognition by deploying multitask learning and transfer learning technology.Second,to facilitate high-precision template matching in the later stage,we utilized an aircraft target mask refinement algorithm and a keypoint-based mask attitude adjustment algorithm to achieve boundary refinement of the recognition target and aircraft target mask attitude adjustment.Finally,on the basis of the aircraft type template library constructed in this study,we matched the refined aircraft mask information with the template library to identify the aircraft type.The proposed algorithm was applied to the MTARSI dataset and remote sensing images for verification.The results showed that the recognition accuracy of the 11 types of images was 89%.Aircraft with simple structures and unique shapes,such as B-2 and B-1,exhibited high recognition accuracy,whereas aircraft with complex structures and high similarity with other shapes,such as E-3 reconnaissance aircraft,exhibited low recognition accuracy.Subsequently,the algorithm was compared with traditional algorithms and end-to-end deep learning methods.Eleven types of aircraft were studied.The results showed th

关 键 词：目标检测 分割 关键点检测 条件随机场 飞机型号识别 

分 类 号：P2[天文地球—测绘科学与技术]