Wide-scale Building Vision Transformer used for Building Detection in Remote Sensing Images

To detect buildings with diverse structures in high-resolution optical remote sensing images, the Wide-scale Building Vision Transformer (W-BViT) model is proposed by optimizing feature representation hyperparameters via analyzing spatial distribution and morphological characteristics of buildings. To extract the details of buildings, the Spatial-Detailed Context branch is constructed to learn the deep semantical information of local objects. To excavate the building correlation features, the global context branch is generated by applying vision encoding. Compared to conventional ViT, W-BViT improves the BuildFormer dual-path ViT model by constructing building-oriented convolution kernel parameters. Experiment results show that the IoU index of the W-BViT model on the WHU (WHU Building Dataset) and Massachusetts datasets is 0.4% and 0.3% higher than that of the BuildFormer model, respectively, and obtains the best detection results among all the comparison methods.