TCINet: Transformer–CNN Interaction Network for Space Noncooperative Object Saliency Detection

Space Non-Cooperative Object Detection (SNCOD) is an essential part of Space Situation Awareness (SSA). The localization and segmentation capabilities of the Salient Object Detection (SOD) method have been proven effective in other common visual tasks, but its application is poorly studied in SNCOD tasks. This paper first conducts extensive comparative experiments using state-of-the-art SOD methods on SwissCube, Satellite and Speed+ datasets to investigate the generalization ability of the SOD method in SNCOD, providing supports for subsequent SNCOD research. Then, based on the above generalization analysis, a Transformer-Convolutional Neural Network (CNN) Interaction Network (TCINet) equipped with a Transformer-CNN Trilateral Backbone (TCTB) and a Loop View and Detail compensation Module (LVDM) is proposed for space non-cooperative object saliency detection. The TCTB incorporates the Pyramid Transformer-in-Transformer (PTNT), the Bilateral Multi-scale Network (BMN), and the Multi-level Cross-Complementary Attention Module (MCAM) to improve the multi-scale adaptability of CNN and guide the semantic modeling of CNN, committed to coping with the scale diversity of space non-cooperative objects. The LVDM chains different dilated convolutions, Top-Down Path (TDP), Multi-level Feature-guided Enhancement (MFE) block, and Multi-scale Detail Enhancement (MDE) block to construct different decision paths. Further, the LVDM sets Switch-Decision Path (SDP) to assign scale-specific view perception and fine-grained enhancement to different decision paths, which decodes the feature set into precise saliency maps through setting chained recurrent training. Finally, comprehensive comparison experiments on three space object benchmark datasets demonstrate that the proposed method achieves superior performance comparable to extant state-of-the-art SOD methods with fewer parameters and computational burden.