Improved Segformer-based Extraction of Building Information from Remote Sensing Images

To tackle the challenges of insufficient feature extraction, unclear segmentation of straight contours, and low segmentation accuracy in semantic segmentation of remote sensing buildings using deep learning models, this paper proposes improvements to the Segformer model and designs an algorithm specifically for extracting building information from remote sensing images, named PD-SegFormer. In this approach, the mit_b1 is employed as the encoder, and two Pag feature fusion modules are utilized at the decoder to merge features of different scales extracted by the encoder, selectively learning the most important features from different scales. Then, the DEC module is introduced to enhance the model's learning ability on building boundaries and strengthen its feature extraction capability. Experimental results show that the improved PD-SegFormer achieves Intersection over Union (IoU) and F1 scores of 78.47% and 87.94%, respectively, representing an improvement of 1.38% and 0.87% over the original model. The comprehensive performance of the improved model surpasses that of the original model and outperforms most segmentation algorithms.