Infrared-visible space debris perception and detection with light-boosting convolution and cross-modal attention fusion

Introduction Space debris poses an increasing threat to the safety of spacecraft because reliable perception under complex illumination remains challenging for autonomous on-orbit operations. Existing visible-light methods are highly susceptible to low-light degradation, whereas infrared images preserve thermal information but lack sufficient structural details. Methods To address these limitations, we propose the Space Debris Perception and Detection Network (SPDNet), an end-to-end infrared-visible framework that jointly performs image enhancement, cross-modal feature fusion, and object detection. Specifically, a Light-Boosting Convolution module restores degraded visible features under weak illumination, a Cross-Modal Attention (CMA) module progressively exchanges complementary infrared and visible information during hierarchical feature extraction, and a Multi-Scale Adaptive Fusion (MSAF) module adaptively integrates multi-scale semantic features to improve both overall detection and small-object recognition. Experiments are conducted on a paired infrared-visible space debris dataset constructed from publicly available visible images and geometrically aligned simulated infrared imagery. Results The proposed SPDNet achieves an average gradient (AG) of 88.9, spatial frequency (SF) of 87.286, mutual information (MI) of 5.946, and Chen-Blum metric (QCB) of 0.941 for image fusion. For debris detection, SPDNet obtains 78.48% mAP, 98.74% mAP50, and 40.27% APsmall, outperforming representative image fusion methods under a unified detection framework. Discussion These results demonstrate that jointly optimized cross-modal feature enhancement and adaptive multi-scale fusion provide an effective solution for autonomous space debris perception under challenging illumination conditions.