Multiscale Adaptively Spatial Feature Fusion Network for Spacecraft Component Recognition

Spacecraft component recognition is crucial for tasks such as on-orbit maintenance and space docking, aiming to identify and categorize different parts of a spacecraft. Semantic segmentation, known for its excellence in instance-level recognition, precise boundary delineation, and enhancement of automation capabilities, is well-suited for this task. However, applying existing semantic segmentation methods to spacecraft component recognition still encounters issues with false detections, missed detections, and unclear boundaries of spacecraft components. In order to address these issues, we propose a multiscale adaptively spatial feature fusion network (MASFFN) for spacecraft component recognition. The MASFFN comprises a spatial attention-aware encoder (SAE) and a multiscale adaptively spatial feature fusion-based decoder (Multi-ASFFD). First, the spatial attention-aware feature fusion module within the SAE integrates spatial attention-aware features, mid-level semantic features, and input features to enhance the extraction of component characteristics, thus improving the accuracy in capturing size, shape, and texture information. Second, the multi-scale adaptively spatial feature fusion module within the Multi-ASFFD cascades four adaptively spatial feature fusion blocks to fuse low-level, middle-level, and high-level features at various scales to enrich the semantic information for different spacecraft components. Finally, a compound loss function comprising the cross-entropy and boundary losses is presented to guide the MASFFN better focus on the unclear component edge. The proposed method has been validated on the UESD and URSO datasets, and the experimental results demonstrate the superiority of MASFFN over existing spacecraft component recognition methods.