舰船遥感图像数据集DSTD的构建研究  

Research on DSTD Construction of Ship Remote Sensing Image Dataset

作　　者：何维娟 江涛[1] 王林飞 徐权峰 王欣 HE Wei-juan;JIANG Tao;WANG Lin-fei;XU Quan-feng;WANG Xin(School of Mathematics and Computer Science,Yunnan Minzu University,Kunming 650500,China;School of Information,Yunnan University,Kunming 650091,China)

机构地区：[1]云南民族大学数学与计算机科学学院,云南昆明650500 [2]云南大学信息学院,云南昆明650091

出　　处：《计算机技术与发展》2022年第7期82-86,92,共6页Computer Technology and Development

基　　金：国家自然科学基金(61363022);云南民族大学研究生创新基金项目(SJXY-2021-002)。

摘　　要：目标检测算法在PASCAL VOC、COCO等一系列数据集中都取得了一定的效果,但是都是面向自然场景下的多目标检测任务,而这些数据集中的单类目标场景往往很单一,相应目标数量也很少,针对特定场景和特定目标的数据集并不多见。而对于此类任务的数据集往往也是很有价值的,例如遥感场景下的舰船检测或者飞机检测。针对此问题,构建一种面向遥感场景的大规模水面舰船目标检测数据集,其数据集的主要来源为DIOR、DOTA、NWPU-VHRdataset、TGRS-HRRSD-Dataset-master等几个开源数据集,将其命名为DSTD(dataset for ship target detection),数据集中包含4854张舰船图片,87076个舰船实例。DSTD数据集具有数量多、多尺度和成像差异大以及较高的类内多样性等特点。在构建数据集的基础上,进一步分析了遥感舰船图像的细节特征,评估了一些经典目标检测方法的性能,并进行了实验对比,验证了该数据集的可行性,同时发现了当前较适合舰船的检测方法:即YOLOV5。使用YOLOV5算法进行舰船图像目标检测,在保证高检测精度的同时,能保持极快的检测速度。Target detection algorithms have achieved certain effects in a series of datasets such as PASCAL VOC and COCO,but they are all for multi-target detection tasks in natural scenes.The single-class target scenes in these data sets are often single and the number of corresponding targets is small.Datasets for specific scenes and targets are rare.Datasets for such tasks are often valuable,such as ship detection or aircraft detection in remote sensing scenarios.To solve this problem,a large-scale surface ship target detection dataset for remote sensing scene is proposed.The main sources of the dataset are DIOR,DOTA,NWPU-VHRdataset,TGRS-HRRSD-Dataset-master and so on,called DSTD(dataset for ship target detection).The dataset contains 4854 ship pictures and 87076 ship instances.DSTD are characterized by large number,large multi-scale and imaging differences,and high intra-class diversity.On the basis of the proposed dataset,we further analyze the detailed features of remote sensing ship images,evaluate the performance of some classical target detection methods,and carry out experiments to verify the feasibility of the dataset.At the same time,YOLOV5,a more suitable ship detection method,is found.The YOLOV5 algorithm is used to detect ship image targets,which ensures high detection accuracy while maintaining extremely fast detection speed.

关 键 词：目标检测 遥感 数据集 舰船 YOLOV5 

分 类 号：TP7[自动化与计算机技术—检测技术与自动化装置]