Statistics-enhanced cross-domain perception network for spaceborne infrared tiny ship detection

Abstract

Spaceborne thermal infrared remote sensing is indispensable for maritime surveillance and ship management, yet detecting tiny ships in infrared imagery remains challenging due to small object sizes, weak textures, and strong structured clutter. To address these challenges, we propose a statistics-enhanced cross-domain perception network (SCPNet), a lightweight one-stage detector tailored to spaceborne thermal infrared tiny ship detection. The main technical innovation lies in a statistics-guided perception design that uses second-order feature statistics to retain polarity-insensitive ship deviations, cross-dimensional channel–spatial interaction to coordinate weak localization cues, and strip-wise frequency-aware contextual modeling to suppress structured maritime clutter. To support systematic evaluation, we also construct the spaceborne thermal infrared ship detection dataset, a large-scale benchmark derived from the Sustainable Development Science Satellite-1 (SDGSAT-1) thermal infrared spectrometer, containing 7660 image tiles and 12,073 annotated ship instances. On the constructed dataset, ablation experiments show that the proposed modules improve average precision from 56.3 percent to 59.6 percent over the baseline, with only 0.18 million additional parameters and 0.2 billion additional floating-point operations. On two additional public ship detection datasets, the proposed detector further exceeds the strongest compared baselines by 1.5 and 1.6 percentage points in average precision, respectively. These results show that the proposed detector improves weak infrared tiny-ship representation with modest computational cost and provides a reproducible baseline for all-day maritime ship detection. The implementation is publicly available at: https://github.com/vazheaven/SCPNet .