PMF-YOLO: a parameter-efficient multi-scale detector for underwater optical imagery under scattering and nonuniform illumination

Underwater optical images suffer from absorption, scattering, and nonuniform illumination, causing veiling light, contrast loss, and blur that weaken small-target boundaries. We propose PMF-YOLO, a parameter-efficient multi-scale detector based on YOLOv8, to address cross-scale imbalance in underwater feature pyramids: high-resolution details are attenuated while low-resolution semantics can be biased by haze-like backgrounds. PMF-YOLO integrates PMFFPN with channel control and a small-object head, CCSA for boundary/saliency recalibration, and ODSSBlock for efficient global modeling of spatially correlated degradation fields. On URPC2020 Test-A/Test-B and DUO, PMF-YOLO achieves 78.1%, 72.1%, and 86.3% mAP@0.5, exceeding YOLOv8n by 7.9, 6.1, and 3.2 points, respectively, with substantially fewer parameters than stronger detectors.