MFEBNet: multidirectional feature enhancement and background-aware network for road extraction from high-resolution remote sensing images

Road extraction from high-resolution remote sensing images plays a significant role in various remote sensing applications. However, accurately extracting roads faces challenges due to the road discontinuity caused by dense occlusion and complex background interference. Inspired by prior knowledge of road shapes, we propose a multidirectional feature enhancement and background-aware network (MFEBNet) to address these challenges. Specifically, MFEBNet employs the visual state space blocks as an encoder to extract global information and introduces the multidirectional feature enhancement (MDFE) module into the skip connections. By integrating multidirectional dilated strip convolution with spatial attention, MDFE effectively tackles the challenge of road discontinuity caused by dense occlusion in complex scenarios. In addition, the background-aware distinction module is proposed to enhance road features and mitigate background interference by adaptively learning channel weights. Extensive experiments conducted on two public datasets demonstrate that MFEBNet outperforms existing methods in terms of road extraction connectivity and completeness.