计算机工程与应用 ›› 2019, Vol. 55 ›› Issue (22): 146-151.DOI: 10.3778/j.issn.1002-8331.1902-0144

• 模式识别与人工智能 • 上一篇    下一篇

遥感图像舰船检测的旋转卷积集成YOLOv3模型

吴止锾,李磊,高永明   

  1. 1.航天工程大学 研究生院,北京 101416
    2.中国人民解放军63883部队
    3.航天工程大学 电子与光学工程系,北京 101416
    4.航天工程大学 航天信息学院,北京 101416
  • 出版日期:2019-11-15 发布日期:2019-11-13

Rotation Convolution Ensemble YOLOv3 Model for Ship Detection in Remote Sensing Images

WU Zhihuan, LI Lei, GAO Yongming   

  1. 1.Graduate School, Space Engineering University, Beijing 101416, China
    2.Unit 63883 of PLA, China
    3.Department of Electronic and Optical Engineering, Space Engineering University, Beijing 101416, China
    4.School of Space Information, Space Engineering University, Beijing 101416, China
  • Online:2019-11-15 Published:2019-11-13

摘要: 遥感图像俯视角带来的目标朝向多样性影响了大长宽比舰船目标检测的旋转不变性。针对这一问题,提出了一个基于改进YOLOv3的倾斜边界框检测模型。通过引入角度预测实现倾斜边界框回归;提出一种旋转卷积集成模块,通过旋转卷积和旋转激活提高深度卷积网络(Deep Convolutional Neural Networks,DCNN)特征图对于角度变化的敏感性;将目标边界框倾斜角度预测建模为由粗粒度到细粒度的两次角度分类问题;将角度惩罚引入模型的多任务损失函数中,使得模型能够学习目标的角度偏移。通过对舰船目标标注数据集上的实验可以看到,所提的模型和经典YOLOv3模型相比平均精度提高了12.7%,同时能够保持单阶段目标检测的速度优势。

关键词: 遥感图像, 目标检测, 舰船检测, 旋转卷积, 深度学习

Abstract: The orientation diversity of target in remote sensing images caused by the bird-eye view affects the rotation invariance of ship detection with large aspect ratio. Aiming at this problem, this paper proposes an incline bounding box detection model based on improved YOLOv3. The incline bounding box regression is implemented by introducing angle prediction. A rotation convolution ensemble module is proposed to increase the awareness for Deep Convolutional Neural Networks(DCNN) feature map to angle change by rotation convolution and rotation activation. The bounding box angle prediction is modeled as a two-step classification problem from coarse to fine. The error of angle classification punishment is introduced into the multi-task loss function of the model, so that the model can learn the angular offset of the object. Experiments on ship dataset show that the proposed model increases the Average Precision(AP) of ship detection in the test dataset by 12.7% compared with the classic YOLOv3 model, while maintaining the advantage of single-stage model in detection speed.

Key words: remote sensing images, object detection, ship detection, rotation convolution, deep learning