DC-KD: double-constraint knowledge distillation for optical satellite imagery object detection based on YOLOX model | |
Yang, Hongbo1,2; Qiu, Shi1; Feng, Xiangpeng1 | |
2024 | |
会议名称 | 4th International Conference on Machine Learning and Computer Application, ICMLCA 2023 |
会议录名称 | Fourth International Conference on Machine Learning and Computer Application, ICMLCA 2023 |
卷号 | 13176 |
会议日期 | 2023-11-03 |
会议地点 | Hangzhou, China |
出版者 | SPIE |
产权排序 | 1 |
摘要 | Object detection is an important application of optical satellite remote sensing imagery interpretation. Since the objects of interest, such as aircraft, ships, and vehicles, are small in size with obscure contour and texture, it is difficult for object detection in satellite images. The spatial resolution of aerial images is higher than satellite images, and the object detection model can achieve higher precision. Knowledge distillation has been validated as an effective technique by learning the common features of aerial and satellite images to improve the precision of object detection in satellite images. It means that a teacher model pre-trained on aerial image datasets guides the training of a compact student model on satellite image datasets. However, there are data distribution differences between aerial images and satellite images. The distribution differences may cause the teacher model to give guidance signals that deviate from the ground truth, thus leading to sub-optimization of the student model. In this paper, we proposed a new distillation scheme, termed DC-KD, which updates the teacher model using the predictions of the teacher model that are inconsistent with the ground truth, and the rest are used to guide the training of the student model. We achieved a 3.88% mAP50 improvement on the xView dataset based on the YOLOX-S model. © 2024 SPIE. |
关键词 | Remote sensing image applications Satellite image processing Tiny object detection Knowledge distillation |
作者部门 | 光谱成像技术研究室 |
DOI | 10.1117/12.3029285 |
收录类别 | EI |
ISBN号 | 9781510680258 |
语种 | 英语 |
ISSN号 | 0277786X;1996756X |
EI入藏号 | 20242316199229 |
引用统计 | |
文献类型 | 会议论文 |
条目标识符 | http://ir.opt.ac.cn/handle/181661/97528 |
专题 | 光谱成像技术研究室 |
通讯作者 | Yang, Hongbo |
作者单位 | 1.Key Laboratory of Spectral Imaging Technology CAS, Xi'an Institute of Optics and Precision Mechanics, Chinese Academy of Science, Xi'an, China; 2.University of Chinese Academy of Sciences, Beijing, China |
推荐引用方式 GB/T 7714 | Yang, Hongbo,Qiu, Shi,Feng, Xiangpeng. DC-KD: double-constraint knowledge distillation for optical satellite imagery object detection based on YOLOX model[C]:SPIE,2024. |
条目包含的文件 | ||||||
文件名称/大小 | 文献类型 | 版本类型 | 开放类型 | 使用许可 | ||
DC-KD double-constra(457KB) | 会议论文 | 限制开放 | CC BY-NC-SA | 请求全文 |
除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。
修改评论