基于改进双边网络的SAR图像海陆分割方法

海陆分割是海岸线提取、近岸目标检测的一个基本步骤。传统的海陆分割算法分割准确度差，参数调节繁琐，难以满足实际应用要求。卷积神经网络能够高效地提取图像多个层次特征，广泛应用于图像分类任务，可作为海陆分割新的技术途径。其中双边网络(BiSeNet)能有效平衡分割精度和速度，在自然场景图像语义分割任务上取得了较好的表现。但对于SAR图像海陆分割任务，双边网络难以有效提取SAR图像的上下文语义信息和空间信息，分割效果较差。针对上述问题，该文根据SAR图像特点减少双边网络中空间路径的卷积层数，从而降低空间信息的损失，并选用ResNet18轻量化模型作为上下文路径骨干网络，减少过拟合现象并提供较广阔的特征感受野，同时提出边缘增强损失函数策略，提升模型分割性能。基于高分三号SAR图像数据的实验表明，所提方法可有效提升网络的预测精度和分割速率，其分割准确度和F₁分数分别达到了0.9889和0.9915，对尺寸大小为1024×1024的SAR图像切片处理速率为12.7 frames/s，均优于当前主流的分割网络框架。此外，所提网络的规模较BiSeNet减少50%以上，并小于轻量级的U-Net架构，同时网络有较强的泛化性能，具有较高的实际应用价值。 

Sea-land Segmentation Method for SAR Images Based on Improved BiSeNet

Sea–land segmentation is a basic step in coastline extraction and nearshore target detection. Because of poor segmentation accuracy and complicated parameter adjustment, the traditional sea–land segmentation algorithm is difficult to adapt in practical applications. Convolutional neural networks, which can extract multiple hierarchical features of images, can be used as an alternative technical approach for sea–land segmentation tasks. Among them, BiSeNet exhibits good performance in the semantic segmentation of natural scene images and effectively balances segmentation accuracy and speed. However, for the sea–land segmentation of SAR images, BiSeNet cannot extract the contextual semantic and spatial information of SAR images; thus, the segmentation effect is poor. To address the aforementioned problem, this study reduced the number of convolution layers in the spatial path to reduce the loss of spatial information and selected the ResNet18 lightweight model as the backbone network for the context path to reduce the overfitting phenomenon and provide a broad receptive field. At the same time, strategies for edge enhancement and loss function are proposed to improve the segmentation performance of the network in the land and sea boundary region. Experimental results based on GF3 data showed that the proposed method effectively improves the prediction accuracy and segmentation rate of the network. The segmentation accuracy and F₁ score of the proposed method are 0.9889 and 0.9915, respectively, and the processing rate of SAR image slices with the resolution of 1024 × 1024 is 12.7 frames/s, which are better than those of other state-of-the-art approaches. Moreover, the size of the network is more than half of that of BiSeNet and smaller than that of U-Net. Thus, the network exhibits strong generalization performance.