Research on building extraction from high-resolution UAV images based on CSDF-UNet

To address the issues of detail information loss, false detection of non-buildings with similar textures to buildings, missed detection of small individual buildings, and inaccurate building edges in existing convolutional neural networks for building extraction, a CSDF-UNet network based on VGG16-UNet is proposed. By introducing conditional parameterized convolution, the weights of the convolution kernel can be adaptively adjusted according to the conditions of the input data, thereby enhancing the receptive field of the network and reducing false building detections. The SMU-1 activation function is introduced to improve detection details and accuracy. Two loss functions, Dice Loss and Focal Loss, are introduced to enhance the network's ability to detect small individual buildings, ultimately constructing the CSDF-UNet model. Meanwhile, high-resolution images of the study area are obtained through unmanned aerial vehicle (UAV) aerial survey technology, and a self-built urban building dataset is used for deep learning training to explore the applicability of the CSDF-UNet model in extracting buildings from different high-resolution UAV images. Experimental results show that compared with SegFormer, PSPNet, DeepLabv3+, and U-Net, the CSDF-UNet constructed in this paper achieves the best performance in terms of IoU, Recall, Accuracy, and F1, improving the accuracy and completeness of building extraction and verifying the feasibility of the method.