High-Resolution Road Information Extraction Based on an Improved Dual-Path Neural Network

With the rapid global urbanization, the extraction of road networks has grown in significance and complexity. Current research on road extraction methods often involves high computational demands, struggles with accurately extracting narrow roads, and encounters challenges in maintaining road connectivity. In this letter, we introduce a dual-path deep learning network. The semantic path integrates insights from the watershed segmentation algorithm as prior knowledge. Meanwhile, the high-resolution path circumvents downsampling during feature extraction, employing a bilateral edge fusion method to enhance information exchange between both paths. In addition, we incorporate a large kernel convolution attention mechanism to enhance the network’s focus on roads. Finally, we devise a strip-shaped convolution module to efficiently capture multiscale information and reduce the parameter count. On the CHN-CUG dataset, our method achieves an intersection over union (Iou) score of 73.36%, surpassing the second-best method by 6.11%. The computational complexity is measured at 60.677 G GFLOPS, with 25.398 M parameters. Our research validates the efficacy of the proposed method for road extraction from remote sensing images.