Abstract: Accurate and robust space-based infrared maritime small target detection plays a vital role in wide-area surveillance, target reconnaissance, and situational awareness. Due to long-range imaging from spaceborne platforms and complex downward-looking backgrounds, maritime infrared targets are extremely small, often occupying only a few pixels, and exhibit weak energy and sparse spatial distributions; background clutter such as cloud edges and coastlines can resemble true targets, leading to false alarms and missed detections. To address these challenges, we propose a Unified Edge-Structure Perception Network (UESP-Net) for space-based infrared maritime small target detection. First, we design an edge-aware enhancement module that explicitly models edge contour features through learnable edge convolutions initialized from multi-directional gradient templates, strengthening the perception of target geometric structures. Second, we introduce a multi-scale context-aware module and a semantic-guided feature refinement module to model global structural dependencies and align cross-layer features. Finally, we design an edge-aware joint loss that improves the preservation of target geometric structures via edge saliency enhancement and directional consistency constraints. Experiments on two public real-world datasets, NUDT-SIRST-Sea and Maritime-SIRST, show that UESP-Net outperforms existing approaches. On Maritime-SIRST, it improves the detection probability (Pd) by 3.31% in absolute terms and reduces the false-alarm rate (Fa) by 27.63% in relative terms over the best-performing baseline AGPCNet. On NUDT-SIRST-Sea, it attains a Pd of 77.80%, surpassing the best competing method MTU-Net by 2.64%, with a recall of 71.55% and an F1 score of 74.63%.

Key words: space-based infrared detection, infrared small target detection, edge-structure joint perception, complex background suppression, semantic segmentation.