Edge Feature-Guided Optimization of VM-UNet Network: Towards Building Segmentation in High-Resolution Remote Sensing Images

Due to the highly complex nature of urban architectural backgrounds, the process of segmenting architectural images often encounters problems such as over-segmentation or under-segmentation. Therefore, achieving precise segmentation of architectural images poses a considerable challenge. To address these challenges, we propose an Edge-guided VM-UNet (EVM-UNet) and introduce an Orientation Convolution Edge Feature Extraction Branch (OCEFEB) to leverage edge information effectively. Furthermore, the model incorporates a Visual State Space (VSS) block to manage the contextual information of image features and employs an Edge Semantic Fusion Module (EF-SF Module) to integrate edge features with the output from the VSS block for enhanced segmentation accuracy Finally, comparative experiments were conducted on the “Instance-Segmentation-Building-Dataset-of-China” dataset. The results show that EVM-UNet surpasses the baseline model on the test set, achieving an average Intersection over Union (MIoU) of 0.6650 compared to 0.6464 for the baseline. This indicates that the proposed model exhibits strong performance in semantic segmentation tasks involving buildings.