使用贝叶斯优化对遥感影像目标进行精确定位  

Object Accurate Localization of Remote Sensing Image based on Bayesian Optimization

作　　者：柴栋 许夙晖 罗畅 鲁彦辰 Chai Dong;Xu Suhui;Luo Chang;Lu Yanchen(Beijing Aviation Engineering Technology Research Center,Beijing 100076,China;The Chinese People's Liberation Army(78102),Chengdu,Sichuan 610031,China;The Chinese People's Liberation Army(78092),Chengdu,Sichuan 610031,China;The Chinese People's Liberation Army(96946),Beijing 102202,China)

机构地区：[1]空军研究院,北京100085 [2]中国人民解放军78102部队,四川成都610031 [3]中国人民解放军78092部队,四川成都610031 [4]中国人民解放军96946部队,北京102202

出　　处：《遥感技术与应用》2020年第6期1377-1385,共9页Remote Sensing Technology and Application

基　　金：军内某科研项目。

摘　　要：针对大尺寸遥感影像目标检测中检测边框不精确的问题,提出使用高斯过程贝叶斯优化对遥感影像中的目标进行精确检测与定位。研究分为两个阶段,第一阶段使用基于边缘信息的EdgeBoxes算法对大尺寸遥感影像进行目标候选区域的选取,用分类器得到初始检测结果;为了得到更加准确的边框,在第二阶段,基于高斯过程的贝叶斯优化对每个目标的边框进行微调:①以目标初始边框为基准,在其周围选取与初始边框相交的边框集合,并得到一个高斯过程分布;②使用贝叶斯优化估计出下一个边框,并将其加入边框集;③求分类器对所有边框的得分,得分最高的边框作为下次迭代的基准边框;④重复若干次贝叶斯优化后得到最终的边框。实验结果表明:EdgeBoxes方法以较少的候选框可以得到较大的召回率,使用高斯过程的贝叶斯优化可以明显地提高检测边框的精度。To solve the issues of inaccurate bounding box in large-scale remote sensing image object detection,an accurate object detection and localization appoarch of remote sensing image based on Bayesian Optimization is proposed. The method consists of two stages:In the first stage,the EdgeBoxes which is based on edges information is adopted to generate object proposals. The classifier is applied to get initial object detection result. To obtain more accurate bounding box,a bayesian optimization based on gaussian process is applied to fine-tune the bounding box around each object in the second stage. Firstly,a set of boxes that intersect the initial bounding box around each initial box is selected to form a gaussian process. Secondly,a new bounding box is estimated through bayesian optimization and added to the set of boxes. Thirdly,the score of each box is calculated by the classifier,and the box with the highest score is set as the base box in the next iteration. At last,the bayesian optimization process is repeatedand and final bounding boxes is obtained. Experiments demonstrate the EdgeBoxes method can achive a better recall evaluation with less number of propsals. The bayesian optimization based on gaussian process can significantly improve the localization accuracy of the detection bounding box.

关 键 词：遥感影像 目标检测 目标精确定位 区域候选框 高斯过程 贝叶斯优化 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]