Abstract:

The uncertainty of ship orientation in remote sensing images, the diversity of ship types and the similarities with other sea and port objects have caused the performance of ship detection to deteriorate seriously. Aiming at this problem, a simple and effective method is used to train the Mask R-CNN ship detection model with rotation invariance and Fisher discriminant. By optimizing the model’s objective function to improve ship detection performance, the original detection is maintained. Based on the invariant model structure, two regularizers are introduced. The first regularizer enhances the feature association before and after training sample rotation. The second regularizer limits the convolutional neural network to have small intra-class divergence and large inter-class divergence. Experimental results on Kaggle remote sensing image ship detection dataset verify the proposed method, which improves performance of detecting ship targets in remote sensing images.

Key words: remote sensing image, Mask R-CNN, ship inspection, rotation unchanged, Fisher discrimination

摘要：

遥感图像中舰船朝向不确定性，舰船种类的多样性以及和其他海上及港口物体之间的相似性，使舰船检测的性能下降严重。针对这一问题，使用一种简单且有效的方法来训练有旋转不变性和Fisher判别的Mask R-CNN舰船检测模型，通过优化模型的目标函数以提高舰船检测性能，在保持原有检测模型结构不变的基础上引入两个正则化器，第一个正则化器加强训练样本旋转之前和之后的特征联系，第二个正则化器限制卷积神经网络有小的类内散度和大的类间散度。实验中，在Kaggle遥感图像船只检测数据集上验证了所提出的方法提高了检测遥感图像中舰船目标的性能。

关键词: 遥感图像, Mask R-CNN, 舰船检测, 旋转不变, Fisher判别