Satellite Component Segmentation Based on Symmetry Loss and Edge Perception

Component segmentation is a crucial task in satellite recognition and tracking, focusing on separating the target from the background and segmenting its internal components for detailed structural analysis. Nevertheless, this task faces significant challenges due to viewpoint variations, illumination changes, and noise interference. To address these challenges, we first propose an improved multi-task network based on U-Net for three-class segmentation, binary segmentation, and edge detection of satellite. The proposed multi-task segmentation framework enables the network to perform fine-grained segmentation of complex satellite structures under varying illumination and viewpoints, thus enhancing the model's robustness. In the edge detection task, we introduce an edge-aware network that extracts multi-scale edge features using a multi-scale convolutional network, which improves the separability of different components. Additionally, we propose a symmetry regularization loss that leverages the satellite's structural symmetry to guide segmentation and improve performance. Experiments demonstrate the efficacy of our method for the satellite component segmentation task.