基于全卷积神经网络的SAR海面溢油图像分割方法

Image segmentation for SAR oil spill on sea surface based on fully convolutional neural network

针对合成孔径雷达(SAR)图像中普遍存在斑点噪声和强度不均匀等问题,提出一种基于全卷积神经网络语义分割框架(FCN)的SAR图像溢油分割方法。首先该方法采用迁移学习来提高泛化能力,从而有效地抑制了溢油区域普遍存在的斑噪声和强度不均匀现象;然后采用跳跃式架构来提高溢油区域的分割精度;最后基于一个包含4200个样本的溢油数据集,将该方法与一些传统机器学习算法(如支持向量机(SVM)、随机森林(RF)和分类回归树(CART)等)和BP神经网络进行对比实验。实验结果表明,该方法相对其他传统方法在像素精度方面提升了7%,针对SAR图像中存在的斑点噪声、强度不均匀及弱边界现象的暗斑分割效果具有显著的改善。

Aiming at speckle noise and intensity unevenness in synthetic aperture radar (SAR) images, an oil spill segmentation method based on Full Convolutional Neural network (FCN) was proposed. Firstly, the migration learning was used to improve the generalization ability, thereby effectively suppressing the universal phenomenon of speckle noise and intensity unevenness in SAR images of oil spill. Then a jump architecture was used to improve the segmentation accuracy. Finally based on an oil spill dataset containing 4 200 samples, the method was compared with some traditional machine learning methods (such as SVM (Support Vector Machine), RF (Random Forest), CART (Classification And Regression Tree)) and BP neural network. The experimental results show that the proposed method improves the pixel accuracy by 7% compared with other traditional methods, and it has significant improvement on dark spot segmentation effect with speckle noise, uneven intensity and weak boundary phenomenon of SAR images.