摘要：

语义分割技术被广泛应用于遥感图像水体提取任务中, 然而语义分割的结果极大依赖于数据集的规模, 针对遥感图像中水体数据集较少、获得精准标注数据成本高的问题, 提出一种用于水体提取的半监督对抗语义分割方法。作为生成器的分割网络中的卷积操作具有受限的感受野, 缺乏对长距离上下文关系的建模能力, Transformer能够建模图像的全局信息。该方法在分割网络中采用Swin Transformer建模深层特征的全局上下文信息, 挖掘像素之间的语义关系, 提高网络的特征提取能力。采用双卷积块提取图像的局部特征, 保留高分辨率细节信息。特征增强模块(FEM)用于抑制图像的背景噪声干扰, 进一步提高水体提取的精度。分割网络和判别器网络共同训练, 以提高在使用少量有标签数据条件下模型提取水体的性能。在GID数据集上进行大量实验, 结果表明, 该方法在不同比例有标签数据条件下均提高了水体提取的精度, 在仅1/8有标签数据的条件下, 该方法取得的F1-Score和交并比(IoU)分别为90.02%和81.86%, 优于U-Net、MWEN等语义分割网络。

关键词: 语义分割, 半监督, 卷积神经网络, Swin Transformer模块, 遥感图像

Abstract:

Semantic segmentation techniques are widely used in water body extraction tasks in remote sensing images. However, the results of semantic segmentation considerably depend on the scale of the dataset. A semi-supervised adversarial semantic segmentation method for water body extraction is proposed herein to address the problems of fewer water body datasets and the high cost of obtaining accurate labeled data in remote sensing images. As a generator, the convolution operation of the segmentation network has a limited receptive field and lacks the ability to model long-range contextual relationships, whereas a Transformer can model the global information of the images. The method uses a Swin Transformer to model the global contextual information of deep features in the segmentation network, mining the semantic relationships between pixels, and improving the feature extraction ability of the network. Double convolution blocks are used to extract the local features of the image and retain the high-resolution detail information. A Feature Enhancement Module (FEM) is used to suppress background noise interference in the image, further improving the accuracy of water body extraction. The segmentation and discriminator networks are jointly trained to improve the performance of the model in extracting water bodies using a small amount of labeled data. Many experiments are conducted on the GID dataset, and the results indicate that the method improves the accuracy of water body extraction under different scales of labeled data. For example, under the condition of only 1/8 labeled data, the F1-Score and Intersection over Union (IoU) of the method achieves 90.02% and 81.86%, respectively, which is superior to other semantic segmentation networks such as U-Net and MWEN.

Key words: semantic segmentation, semi-supervised, Convolutional Neural Network (CNN), Swin Transformer module, remote sensing images