基于二值化条件随机场卷积网络的极化SAR海陆分割

Sea-land Segmentation of Polarimetric SAR Image Based on Binarized Conditional Random Field Segmentation Network

针对现有SAR海陆分割预测精度较低,采用的分割网络模型普遍较大、难以星上部署等难点,提出了一种基于二值化条件随机场卷积网络的极化SAR海陆分割方法(BiCSNet)。该模型的轻量化主要通过所设计的适用于海陆分割二元任务的二值化卷积模块实现,为了提高轻量化网络的分割精度,BiCSNet还融入了卷积条件随机场实现端到端的网络预测功能。基于我国沿海区域的全极化SAR图像构建的数据集,验证了所提出网络在精度和轻量化两方面的良好性能。

Due to the low prediction accuracy of existing SAR sea-land segmentation,the segmentation network model adopted is generally large and difficult to deploy on board.A binarized conditional random field segmentation network(BiCSNet)for polarimetric SAR is proposed.The lightweight model is mainly realized by the binarization convolution module designed for the binary task of sea-land segmentation.In order to improve the segmentation accuracy of the lightweight network,the convolution conditional random field is incorporated into BiCSNet to realize the end-to-end network prediction.Based on the data set constructed from quad-polarization SAR images in coastal areas of China,the proposed network has good performance in both accuracy and lightweight.