BioVis-Det: scale–adaptive and noise–resilient underwater detection via dynamic aggregation and orthogonal calibration

Underwater object detection encounters inherent difficulties such as turbidity, scattering, dense small targets, and diverse morphologies, which severely compromise the robustness of generic methods. Biological vision has evolved to possess efficient perceptual mechanisms in similarly complex environments, providing a natural source of inspiration. This paper addresses two core challenges: balancing fine-grained perception of small objects with inference efficiency, and mitigating noise and geometric degradation in cross-scale fusion. Drawing on the hierarchical perception and neuromodulatory mechanisms of the biological visual pathway (retina → LGN → V1), we propose BioVis-Det (Biological Vision Detection), an improved algorithm based on YOLO11 that takes these core biological visual mechanisms as a high-level conceptual guide for architectural design. Three bio-inspired modules are integrated into the algorithm: SDHR (Scale-Adaptive Detection Head Restructuring) mimics retinal foveal heterogeneous sampling and “use it or lose it” pruning to enhance small-object recall while preserving efficiency; DNFA (Dynamic Noise-Filtering Aggregation) emulates LGN neural gating and gain control to realize confidence-driven cross-scale fusion and noise suppression; OGSC (Orthogonal-Global Spatial Calibration) simulates V1 orientation selectivity and lateral inhibition to strengthen geometric representation under refraction and deformation. Extensive experiments on URPC, RUOD, and COCO datasets demonstrate that BioVis-Det achieves mAP50 improvements of 3.4%, 2.6%, and 2.5% over the YOLO11 baseline, respectively. With a negligible increase in parameters, the proposed algorithm effectively reconciles detection accuracy and deployment feasibility on resource-constrained underwater devices.