AF-Net: All-scale Feature Fusion Network for Road Extraction from Remote Sensing Images

Road extraction from high-resolution remote sensing images (RSIs) is a challenging task due to occlusion, irregular structures, complex background, etc. A typical solution for road extraction is semantic segmentation that tries to segment the road region directly from the background region at the pixel level. Because of the narrow and slender structures of roads, high-quality multi-resolution and diverse semantic feature representations are necessary for this task. To this end, this paper introduces an all-scale feature fusion network named as AF-Net to extract roads from RSIs. AF-Net adopts an encoder-decoder architecture, whose encoder and decoder are connected by the introduced all-scale feature fusion module (AF-module). AF-module contains multiple feature fusion stages, corresponding to features of different scales. At each stage of feature fusion, all-scale all-level feature representations are employed to recursively integrate the features from two paths. One path propagates the high-resolution spatial features to the current scale feature and another path merges the current scale feature with high-level semantic features. In this way, we effectively employ all-scale features with varied spatial information and semantic information in each fusion stage, facilitating producing more accurate spatial information and richer semantic information for road extraction. Moreover, a convolutional block attention module is embedded into AF-module to suppress unconducive features from the surrounding background and improve the quality of extracted roads. Due to the features with richer semantic information and more precise spatial information, the proposed AF-Net outperforms other state-of-the-art methods on two benchmark datasets.