Stable Incremental Underwater Object Detection via Adaptive Representation Routing and Topology-Preserved Replay

Incremental object detection (IOD) is critical for autonomous underwater perception, where detectors deployed on long-duration underwater platforms must continuously adapt to evolving marine environments while retaining previously learned recognition and localization capabilities. However, underwater IOD is particularly challenging because visual degradation, small-object ambiguity, background dominance, rare-class dilution, and non-stationary data distributions jointly cause structure instability in incremental representations. Existing response-based distillation methods, such as elastic response distillation (ERD), mainly preserve output-level responses and often rely on rigid backbone updating and static optimization strategies, making them insufficient for maintaining hierarchical representation stability under degraded and imbalanced underwater observations. To address these limitations, we propose a structure-stable underwater IOD framework that jointly regulates representation update, replay topology, and optimization dynamics within a unified stability–plasticity formulation. Specifically, Structure-Adaptive Residual Routing (SARR) replaces binary layer freezing with adaptive residual paths, task-aware gradient routing, and semantic-sensitive update gates, enabling parameter-efficient incremental representation routing. Topology-Aware Semantic Replay (TASR) maintains class prototypes, teacher-guided relation matrices, and decision-boundary anchor samples to preserve the neighborhood topology of rare old classes in the feature manifold. Uncertainty-Aware Plasticity–Stability Feedback (UPSF) dynamically adjusts classification and localization distillation strengths according to old-class forgetting risk, new-class learning difficulty, and localization uncertainty. Extensive experiments on UTDAC2020 and DUO demonstrate that the proposed method outperforms representative distillation- and transformer-based IOD baselines in most incremental settings. In particular, our method achieves 30.7% AP on UTDAC2020 under the 2 + 2 setting and narrows the gap to full-data training, validating the effectiveness of structure-stable representation evolution for robust incremental underwater object detection.