融合Transformer和CNN的U型神经网络遥感影像道路提取算法
Remote Sensing Image Road Extraction Algorithm Based on U-Type Neural Network and Transformer Combined with CNN

U型网络作为一种经典的编码–解码结构网络，不只在医学影像领域内发挥出色，在图像分割领域也有广泛的影响。以U型网络为基础其它衍生网络层出不穷。U型网络最经典的思想是编码和解码，再加上编解码之间的跳跃连接。由于道路遥感图像和医学影像有众多相似的地方，如今U型网络又被用于从遥感图像中提取道路。U型网络使用跳跃连接的方式将下采样低维特征拼接到上采样的高维特征中，以保留更多的空间位置信息和语义信息。因此U型网络更能处理一些特征信息明显的图像数据。但浅层的UNet无法准确提取道路丰富多维的细节信息，在高分辨率卫星遥感图像上无法奏效。所以本文提出一种融合蛇形动态卷积和Swin-Transformer的U型网络用于提高道路提取任务的分割精确度。
As a classical coding-decoding structure network, U-shaped network not only plays an excellent role in the field of medical imaging, but also has a wide impact in the field of image segmentation. Based on U-shaped network, other derivative networks emerge in an endless stream. The most classic idea of U-shaped networks is encoding and decoding, plus jumping connections between coding and de-coding. Due to the similarities between road remote sensing images and medical images, U-shaped networks are now used to extract roads from remote sensing images. The U-shaped network spliced the low- dimensional features from the down-sampled to the high-dimensional features from the up-sampled by means of jump connection, so as to retain more spatial location information and se-mantic information. Therefore, U-shaped network is more capable of processing some image data with obvious feature information. However, shallow UNet can not accurately extract the rich mul-ti-dimensional detailed information of the road, and can not be effective in high-resolution satellite remote sensing images. Therefore, this paper proposes a U-shaped network combining serpentine dynamic convolution and Swin-Transformer to improve the segmentation accuracy of road extraction tasks.