| 摘要: |
| 舰船目标检测是进行海洋环境监管，保障海上权益的重要手段。基于深度学习的目标检测算法能在复杂环境下保持良好性能，为测试不同深度学习目标检测算法在舰船检测中的效果，构建了一个包含3 893张图像的数据集，涵盖了复杂背景下不同类型的舰船，基于此数据集分别采用Faster RCNN、SSD、RetinaNet、YOLOv3、YOLOv4算法进行实验，结果表明，YOLOv4、YOLOv3、RetinaNet、Faster RCNN平均精度均在83%以上，其中YOLOv4最高达到91.77%，Faster RCNN误检较多，而SSD平均精度最低，只有79.23%，总的舰船检测数目偏少。将5种模型训练结果在高分二号影像上进行测试，得到较好的检测效果，对舰船检测未来理论研究的开展具有一定的指导意义。 |
| 关键词: 遥感 目标检测 舰船 深度学习 |
| DOI：10.11759/hykx20201108009 |
| 分类号:P237 |
| 基金项目:国家重点研发计划项目（2017YFC1405600） |
|
| Comparative analysis of ship detection algorithms based on deep learning in optical remote sensing images |
|
Xia Wen-hui, WAN Jian-hua, ZHENG Hong-xia, Xu Ming-ming, Qu Chuan-ping
|
|
China University of Petroleum, Qingdao 266580, China
|
| Abstract: |
| Ship target detection is an important means to supervize a marine environment and protect maritime rights and interests. A target detection algorithm based on deep learning can maintain good performance in a complex environment. This study describes the construction of a dataset containing 3, 893 images, covering different types of ships in a complex background. Based on this dataset, the performances of Faster RCNN, SSD, RetinaNet, YOLOv3, and YOLOv4 algorithms are examined. The results show that the mean average precision of YOLOv4, YOLOv3, RetinaNet, and Faster RCNN are above 83%. Furthermore, YOLOv4 reached 91.77%, Faster RCNN produced more false detections, and SSD achieved the lowest mean average precision, only 79.23%, with a low total number of ship detections. The training results of the five models are tested on a GF-2 image, with good detection results obtained; these results have certain guiding significance for the future theoretical research of ship detection. |
| Key words: remote sensing target detection ship deep learning |
