CEBRNet: Connectivity Enhancement and Boundary Refinement Network for Road Extraction From Satellite Imagery

The encoder-decoder architecture is widely adopted in convolutional neural network (CNN)-based road extraction methods. However, due to the absence of long-range dependencies in the upsampling process of a decoder model and the global spatial correlation of occluded road regions, existing methods often produce fragmented roads with poor connectivity. Besides, shallow encoder layer features may capture massive needless low-level textures that can be regarded as noises due to the limited receptive field, blurring the boundaries of roads. To enhance the connectivity and refine the boundary of roads, the connectivity enhancement module (CEM) and the boundary refinement module (BRM) are proposed, respectively. Our CEM is able to capture global relations and extract the linear features of roads effectively, which are critical for topological correctness. By leveraging the output of the CEM as a source of semantic guidance, BRM is introduced to filter out the local noises from encoder features. Then, we construct an interactive global-local decoder (IGD) based on the two proposed modules to enable dynamic interactions between them, resulting in further mutually beneficial improvements. Based on these innovations, our CEBRNet method is proposed for road detection from satellite imagery. With extensive experiments on two public benchmark DeepGlobe and SpaceNet datasets, we demonstrate the superiority of our method against other state-of-the-art methods.