Road Extraction Network for Remote Sensing Images Based on Multiscale Feature Fusion and Edge-Guided Attention

With the rapid advancement of geospatial technology, automatic road extraction has become an increasingly important and impactful task. However, the results of existing methods often suffer from incomplete road connectivity and poor topology due to complex background interference, diverse road morphology and occlusion. In this study, we propose a novel convolutional neural network that integrates classical edge detection techniques with attention mechanisms. These techniques effectively preserve high-frequency information, particularly edges and boundaries. Additionally, we introduce a new weighted bi-directional feature pyramid network (BiFPN) designed to capture multi-scale semantic information across different layers, thereby bridging the semantic gaps between low-level features and high-level feature maps. We conduct experiments on two distinct road datasets: the DeepGlobe dataset and the Massachusetts dataset. The results demonstrate that our model enhances overall performance compared to several state-of-the-art (SOTA) algorithms, with Intersection over Union (IoU) metrics improving by 2.32% and 1.37% over Unet34 on the two datasets, respectively.