基于可变形部件模型的遥感影像船只检测方法

Ship Detection in Highresolution Remote Sensing Images Based on Deformable Part Model

摘要:

提出了一种基于混合可变形部件模型的船只检测方法.该方法采用基于梯度方向直方图（HOG）特征的可变形部件模型（DPM）来描述船只，在HOG特征金字塔空间通过滑窗方式检测目标.基于HOG特征的DPM不具有方向不变性，因此根据船只目标特点，分别从模型训练阶段和目标检测阶段进行改进.在模型训练阶段，为了减少模板数量，将所有船只样本旋转到相同方向进行参数学习；在目标检测阶段，将兴趣区旋转至特定方向后进行模板匹配，实现遥感影像上任意方向的船只检测.利用该方法在高分二号光学遥感影像上进行船只检测，实验结果表明此方法可以有效检测船只.

Abstract:

A method of ship detection in highresolution remote sensing images using mixture of deformable part models(DPMs) is proposed in this paper. The method is robust to size difference by constructing multiscale histogram of oriented gradients (HOG) feature pyramids. Deformable part models are introduced to deal with the deformation of the key parts of ships. And a sliding window detection strategy is adopted to separate the clustered ships. Considering that HOG features are orientationsensitive, improved training and detection methods are also proposed. In the training phase, all the samples are rotated to the same direction for parameter learning to reduce the number of templates. In the detection phase, the regions of interest are rotated to a specific direction to implement template matching. The experiments of ship detection in GaoFen2 highresolution remote sensing images are carried out. It is shown that the proposed method is effective.