BMP-Net: a semantic-segmentation model for road extraction in remote-sensing images

With the ongoing advancement of urban infrastructure, the precise extraction of road networks from remote sensing imagery has grown increasingly significant. Semantic segmentation offers an effective approach to accomplish this task; however, conventional convolutional neural network (CNN) methods still exhibit limitations in global perception and multiscale information processing, which constrains the enhancement of segmentation accuracy. We propose a network architecture BMP-Net, including BiFormer, a multilayer feature cross-fusion module (MFCMF), and a pyramid information-fusion module (PIFM). BiFormer achieves more flexible arithmetic allocation through a bilevel routing attention, and it has higher computational efficiency when extracting features. In MFCMF, feature maps of different layers are cross-fertilized to capture richer semantic information. In PIFM, shifted multilayer perceptron and partial large-kernel convolution are used for feature extraction to obtain a global sensory view while reducing the computational effort as much as possible. Experiments on two datasets demonstrate that BMP-Net achieves superior performance compared with existing methods, with intersection over union reaching 63.76% and 74.76%, respectively. The code is available at https://github.com/FDZ-XL/BMP_Net