FH-YOLO: a cross-stage stabilized framework for underwater benthic object detection

Underwater object detection is challenged by light attenuation, scattering, and severe appearance degradation, which introduce instability across different stages of the detection pipeline. To address this issue, we propose FH-YOLO, a cross-stage stabilized framework designed to improve detection robustness under degraded underwater conditions. The proposed method integrates three key components: Hybrid Data Augmentation (HDA), which simulates stochastic underwater degradation to expand the input distribution; Bi-level Routing Attention (BRA), which refines multi-scale feature representations by suppressing noise-amplified responses; and Wise-IoU v3, which stabilizes regression by adaptively reweighting gradient contributions. Evaluated on the Underwater Benthic Dataset (UBD), FH-YOLO achieves 87.1% [email protected], outperforming YOLOv11 and several representative detectors. Both quantitative and qualitative results demonstrate that the proposed cross-stage coordination significantly improves detection accuracy and robustness in complex underwater environments.