旋转矩形区域的遥感图像舰船目标检测模型  

Ship Detection in Remote Sensing Based with Rotated Rectangular Region

作　　者：仲伟峰[1] 郭峰 向世明[2] 潘春洪[2] Zhong Weifeng;Guo Feng;Xiang Shiming;Pan Chunhong(School of Automation,Harbin University of Science and Technology,Harbin 150080;National Laboratory of Pattern Recognition,Institute of Automation,Chinese Academy of Sciences,Beijing 100190)

机构地区：[1]哈尔滨理工大学自动化学院,哈尔滨150080 [2]中国科学院自动化研究所模式识别国家重点实验室,北京100190

出　　处：《计算机辅助设计与图形学学报》2019年第11期1935-1945,共11页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(91646207);科技部重点研发计划(2016YFB0501100)

摘　　要：高分辨率遥感图像舰船目标检测是遥感图像理解任务中的热点研究问题.由于遥感图像中舰船目标存在成像视角单一、目标分布密集和目标尺度变化大等特点,直接将自然场景目标检测方法应用于遥感图像舰船检测任务中,并不能获得满意的效果.此外,自然场景目标检测任务中常用的水平矩形框对细长型舰船目标的定位精确度无法满足实际应用需求.因此,提出了基于旋转矩形区域的遥感舰船目标检测算法.首先,采用旋转矩形框完成舰船目标的定位.其次,提出兴趣区域特征金字塔池化模块,融合兴趣区域的多尺度池化特征以处理目标尺度变化较大的问题.最后,设计定位准确度预测分支,利用定位准确度预测值指导非极大值抑制算法,从而优化后处理的结果.在遥感舰船公开数据集HRSC2016上,通过3个级别任务(分别为单类、4类和19类舰船检测识别)上的实验结果验证了算法的有效性.The detection and recognition of ships in high-resolution remote sensing images play a vital role in the understanding of remote sensing images. Due to the characteristics of ships in high-resolution remote sensing images(such as viewing angle of images, distribution of objects, various scale of targets, etc.), simply applying the detection algorithm for natural images to remote sensing images can hardly obtain satisfactory performance. To this end, this paper proposes a ship object detection algorithm in remote sensing based with Rotated Rectangular Region. Firstly, a rotation region representation method is introduced to locate and classify the ship objects precisely. Secondly, pyramid pooling module of region of interest(RoI) is proposed, which integrates the multi-scale pooling features of RoI to adapt to the large scale range of the ship target. Finally, localization confidence prediction branch is designed to use intersection over union(IoU) guided non-maximum suppression, which optimizes the post-processing results. Experiments on HRSC2016 dataset show that our method outperforms exiting methods.

关 键 词：遥感图像 卷积神经网络 舰船检测 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]