• ISSN 1004-0323     CN 62-1099/TP
• 联合主办：中国科学院遥感联合中心
• 中国科学院兰州文献情报中心
• 中国科学院国家空间科学中心
 遥感技术与应用  2020, Vol. 35 Issue (6): 1377-1385    DOI: 10.11873/j.issn.1004-0323.2020.6.1377
 数据与图像处理

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
Object Accurate Localization of Remote Sensing Image based on Bayesian Optimization
Dong Chai1(),Suhui Xu1,2(),Chang Luo3,Yanchen Lu4
1.Beijing Aviation Engineering Technology Research Center，Beijing 100076，China
2.The Chinese People's Liberation Army（78102），Chengdu，Sichuan 610031，China
3.The Chinese People's Liberation Army（78092），Chengdu，Sichuan 610031，China
4.The Chinese People's Liberation Army（96946），Beijing 102202，China
 全文: PDF(5080 KB)   HTML

Abstract:

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.

Key words: Remote sensing image    Object detection    Object accurate locolization    Region proposal    Gaussian process    Bayesian optimization

 ZTFLH: TP751