TAPO: Task-Decoupled Alignment and Pseudo-Label Optimization for Cross-Scene Inshore SAR Ship Detection

Abstract

In complex inshore regions, ships are often densely berthed and arbitrarily oriented, while docks, shorelines, coastal facilities, and strong scatterers produce substantial background clutter and target-like scattering responses, leading to foreground-background confusion and localization ambiguity. In practical inshore synthetic aperture radar (SAR) ship detection, training and testing images often originate from different coastal scenes. Such cross-scene domain shifts further exacerbate these challenges, resulting in degraded detection performance and reduced pseudo-label reliability during self-training. To address these issues, this study proposes a task-decoupled alignment and pseudo-label optimization framework (TAPO) for cross-scene inshore SAR ship detection with oriented bounding boxes. First, the Task-Decoupled Imbalanced Feature Alignment module (TD-IFA) separately aligns classification- and regression-related region-of-interest features, allowing foreground-background discrimination and oriented-box localization to be adapted according to their different cross-scene shifts. Imbalanced source-target alignment weights are further introduced to reduce the influence of noisy target-scene proposals. Second, the Dynamic Uncertainty-Driven Pseudo-Label Optimization module (DUD-PLO) improves self-training reliability by selecting pseudo-labels that are both confident and stable. Dynamic confidence filtering combines a fixed threshold with an adaptive fallback threshold to suppress low-confidence noisy candidates while reducing the risk of obtaining empty pseudo-label sets in difficult target scenes. Cross-view consensus retains candidates that remain consistently matched across non-geometric perturbation views, while rotated-box geometric stability, measured by rotated intersection over union (rotated IoU), and uncertainty weighting further suppress pseudo boxes with unstable locations, scales, or angles. In addition, a Kullback–Leibler divergence (KLD)-based auxiliary loss provides soft geometric consistency for oriented boxes during training. Experiments on a self-constructed cross-scene inshore SAR ship dataset show that TAPO achieves a mean average precision (mAP) of 0.439, improving over the Source-only oriented-box baseline by 0.102 and achieving the best overall performance among the compared methods.