Answers are grounded in this paper and cite their sources.
Underwater trash detection is challenging due to wavelength-dependent attenuation, low contrast, and complex seafloor backgrounds, which undermine detectors relying solely on RGB input. This paper proposes MSFD-YOLO, a dual-branch mid-fusion detection framework that couples RGB appearance with a Morphometric Structural Cue Map (MSCM)—a six-layer RGB-derived structural prior, deterministically computed from the RGB image itself via classical morphometric operators, that encodes base intensity, gradient energy, orientation, curvature, neighborhood irregularity, and residual detail without introducing any additional sensing modality. Built upon YOLOv11, the framework employs a symmetric two-stream backbone with mid-level cross-branch aggregation at the P4 and P5 stages and introduces two redesigned modules: C3K2-MDE for multi-dilation receptive field expansion and C2-DFA for spatially adaptive deformable attention, together with a DIoU-based regression loss. Structured ℓ 2 -norm pruning with selective layer protection is applied after training to ensure deployment efficiency. Experiments on the TrashCan dataset show that MSFD-YOLO outperforms the RGB-only YOLOv11n baseline by 7.2% in precision, 5.9% in recall, 3.8% in m A P 50 , and 3.7% in m A P 50:95 . Compared with 19 state-of-the-art detectors spanning recent YOLO variants, transformer-based detectors, and task-specific underwater models, MSFD-YOLO achieves the best m A P 50 (68.0%) and m A P 50:95 (50.2%) with only 4.20 M parameters, 6.46 GFLOPs, achieving 149 FPS network-only throughput on an RTX 4060 Ti GPU. Edge deployment and recovery-value analyses together provide complementary evidence of the framework’s practical feasibility for ROV/AUV-based cleanup operations.
No takes yet. Share an insight, caveat, or question.
Liu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: