AGF-Net: Adaptive Road Extraction Network with Global Feature Fusion

Remote sensing image road extraction is of great significance for vehicle navigation and emergency insurance. However, due to the limited presence of roads in remote sensing images, they are often obstructed by shadows from trees or buildings along the road, leading to fragmented extraction results. Moreover, the variation in road scales within remote sensing images makes it challenging to accurately extract small-sized roads. Therefore, automatic road extraction from remote sensing images poses a significant challenge. To address these issues, this paper proposes an adaptive road extraction method based on global feature fusion. The proposed method employs an encoder-decoder structure for end-to-end road extraction. In the middle module, a dilated convolution strip attention module is designed to enhance the continuity of roads. The dilated convolution module expands the receptive field, while the strip attention module adaptively emphasizes the relevant features along the road direction. Multiple global feature fusion modules in the skip connection are designed to fuse the features of different stages. Meanwhile, a multi-scale strip convolution module is employed to compensate for the discrepancies in feature maps across different stages. Extensive experiments reveal that AGF-Net outperforms existing methods on different road datasets (Massachusetts Road dataset and CHN6-CUG Road dataset).