Adaptive Flare Suppression Algorithm for High-Precision AUV Landing Recovery Based on Transfer Learning and R-Shaped Beacon Array

This paper proposes a high-precision Adaptive Flare Suppression Guidance Algorithm (AFSGA) to address the challenge of pose estimation in Autonomous Underwater Vehicle (AUV) landed recovery caused by non-uniform light fields, lens flare, and impurity interference during the terminal stage. By integrating Gaussian Mixture Models (GMM) and transfer learning techniques, the proposed algorithm overcomes the dependency of conventional Perspective-n-Point algorithms on ideal optical conditions, enabling robust six-degree-of-freedom pose estimation in complex underwater environments. First, a dynamic flare suppression mechanism extracts image regions of interest via GMM and adaptively adjusts local thresholds to mitigate flare effects. Second, a transfer learning-enhanced Otsu algorithm is introduced, requiring minimal annotated data to generalize across scenarios and significantly reducing deployment costs. Finally, an R-shaped beacon array paired with a geometry-based feature matching strategy ensures pose estimation robustness under arbitrary rotational motions. Controlled pool experiments demonstrate that AFSGA achieves stable attitude estimation at 2 Hz on embedded platforms, with root mean square errors (RMSE) of 0.623° and 0.543° for roll and pitch angles, respectively. Compared to deep learning baselines (e.g., DeepLabV3 model with a MobileNetV3 backbone), AFSGA improves segmentation accuracy by 28.9% and reduces computational latency by 23.1%. These results validate the algorithm’s capability to provide high-precision, low-cost, and real-time guidance for AUV recovery in optically challenging underwater environments.