The extraction of buildings from high-resolution remote sensing images is of great significance in three-dimensional reconstruction of urban scenes. A remote sensing image building extraction shape correction generative adversarial network (SCGAN) with improved high-resolution network (HRNet) is proposed to address the problem of low segmentation accuracy caused by mutual occlusion and blurred boundaries of buildings in complex background remote sensing images when using traditional convolutional methods. Based on the HRNet structure, shape correction units are introduced to enhance the model's perception of building edges and shapes, and adversarial learning strategies are used to strengthen detailed features such as building boundaries and geometric shapes. The experimental results show that the SCGAN model based on adversarial learning and shape correction units effectively improves segmentation accuracy in building extraction, with IoU of 90.94% and 70.89% on the WHU and Massachusetts datasets, respectively, exhibiting the best performance compared to popular semantic segmentation models.