Improved Lightweight Saliency Model Based on Neural Network for Noncooperative Spacecraft Detection

With the expansion of space exploration and activities, the space environment becomes increasingly complex, and thus, the importance of noncooperative spacecraft detection (NCSD) becomes more significant for enhancing space situational awareness capabilities. However, the implementation of NCSD in space faces two challenges: the resource constraints of spaceborne embedded systems and the object scale variations caused by complex imaging environments in space. Accordingly, an improved lightweight saliency model is proposed in this article for NCSD in the space environment. In this model, a lightweight feature extractor that aims to extract multilevel multiscale features is designed as the backbone network using the in-layer multiscale block and the intrafrequency multiscale block. Integrity channel attention feature fusion modules are then applied to enhance the channel discrimination of multilevel features. Furthermore, a part-whole verification module is introduced to strengthen the learned integrity feature by measuring the consistency between the target parts and the complete target region. Extensive contrast experiments were conducted on the SwissCube dataset, and the results clearly demonstrate that the proposed lightweight saliency model achieves relatively ideal segmentation effect of noncooperative object spacecraft and competitive performance in terms of a wide range of metrics with only 1.99M parameters.