SegRoadv2: a hybrid deformable self-attention and convolutional network for road extraction with connectivity structure

Abstract

Road extraction is crucial for navigation, autonomous driving, and smart city development. With advancements in remote sensing and deep learning, the extraction of road information from remote sensing images has emerged as a prevalent research area. Nevertheless, the complexity of roads and image characteristics pose various challenges. To address this issu e, we propose SegRoadv2, a road extraction algorithm based on SegRoad. SegRoadv2 employs a transformer block with a deformable self-attention (DSA) module and a CNN structure with a new groupable deformable convolution (GroupDCN). Additionally, the novel re-parameterized strip convolutions in the decoder and a pixel connectivity structure improve segmentation connectivity. Tested on the DeepGlobe, Massachusetts, and CHN6-CUG datasets, SegRoadv2 exhibits a novel, state-of-the-art performance, achieving an IoU of 69.88% on DeepGlobe and excellent results on the other datasets. These findings highlight the potential of this algorithm for urban development applications.