基于残差混合监督网络的无人机目标阴影检测  

Shadow detection of UAV target based on residual mixed supervision network

作　　者：王潇 刘贞报[1] 史忠科[2] WANG Xiao;LIU Zhenbao;SHI Zhongke(School of Civil Aviation,Northwestern Polytechnical University,Xi’an 710072,China;School of Automation,Northwestern Polytechnical University,Xi’an 710129,China)

机构地区：[1]西北工业大学民航学院,西安710072 [2]西北工业大学自动化学院,西安710129

出　　处：《航空学报》2024年第17期279-292,共14页Acta Aeronautica et Astronautica Sinica

基　　金：国家自然科学基金(52072309);陕西省重点研发计划(2019ZDLGY14-02-01);深圳市基础研究资助项目(JCYJ20190806152203506);航空科学基金(ASFC-2018ZC53026)。

摘　　要：无人机在检测和跟踪目标过程中受到目标伪装、目标遮挡、移动躲避以及假目标等因素的干扰,而无人机目标附近的阴影区域加剧了这些因素对目标检测和跟踪性能的影响,因此检测无人机目标阴影区域是无人机领域的重要研究任务之一。现有无人机目标阴影检测方法面临训练数据数量有限、数据收集标注困难以及无人机目标中存在大量尺寸较小的细碎阴影区域等问题,针对这些问题,提出一种基于残差混合监督网络的无人机目标阴影检测算法。首先针对无人机目标阴影检测任务的特点设计分辨率注意力网络,在结合底层纹理特征和高层语义特征的过程中,更准确地保留底层纹理特征。然后设计混合监督网络扩充训练数据集,结合普通阴影检测数据集和无人机目标阴影检测数据集训练教师网络,使用无人机阴影检测数据集和教师网络的参数训练学生网络。同时设计残差图像,利用教师网络检测结果和标准结果之间的残差图像扩充训练数据集,使阴影检测网络更加关注细碎阴影区域。最后,在2个公开实验数据集上和已有方法进行对比实验,在各个评价参数上取得了最多41.6%的提升效果,证明所提无人机目标阴影检测算法较好的解决了现有方法存在的问题,具有较高的准确性。The targets camouflage,targets occlusion,moving dodge and fake targets deteriorate the performance of UAV object detection and tracking,and the shadow regions around UAV target aggravate the negative effect of these factors.Thus,shadow detection is an important task for UAV.The shadow detection of UAV target suffering from lim⁃ited training images,difficulty in labeling ground truth data and mass of tiny shadow regions.To deal with these prob⁃lems,we propose a UAV target shadow detection method based on residual mixed supervision network.Firstly,we design a resolution-aware attention shadow detection network based on the character of shadow regions in UAV tar⁃get.The newly designed network can maintain the lower texture feature more accurately.Then we design mixed su⁃pervision network to enlarge the number of training images.The teacher network is trained by both ordinary dataset and UAV dataset,while the student network is trained based on UAV dataset and the parameter of teacher network.Meanwhile we design residual images to further enlarge the number of training images and makes the network pay more attention to tiny shadow regions.The residual image is calculated by measuring the difference between detection results of teacher network and ground truth data.At last,the proposed method is compared with existing methods on two public UAV target shadow detection datasets.The evaluation metrics are improved by 41.6%at most.The experi⁃ment proves the effectiveness and accuracy of proposed shadow detection method on UAV target.

关 键 词：无人机目标 阴影检测 混合监督网络 残差图像 分辨率注意力网络 

分 类 号：V249[航空宇航科学与技术—飞行器设计] TP391[自动化与计算机技术—计算机应用技术]