Answers are grounded in this paper and cite their sources.
Due to severe color distortion in underwater optical imagery and the inherently weak features of small scale targets, underwater object detection remains highly challenging. Although conventional YOLO detectors achieve competitive performance in general scenarios, they often fail to effectively model the correlations between small, weak targets and complex underwater backgrounds. This limitation frequently results in missed detections and false alarms. To address these issues, this article proposes Hyper-YOLOn-MDS, an efficient and accurate detection framework tailored for article small object scenarios. Specifically, three modules that balance complexity and effectiveness are designed. First, the multiscale parameter-shared dilated pyramid captures adaptive receptive fields through a cascaded structure, flexibly capturing effective sensory fields and enabling robust multiscale representation. Second, the dual-branch adaptive downsampling method achieves adaptive downsampling by leveraging saliency operators and a dual-path architecture, effectively enhancing the representation of weak features and small-scale targets. Third, the shallow-enhanced cross-stage omni fusion enhances feature representation by injecting shallow P2 features into the neck, achieving an effective multiscale feature fusion of low-level details and high-level semantics through cross stage interactions and dual domain modeling. Extensive experiments are conducted on two publicly available underwater small object datasets, DUO and URPC 2020. The proposed framework achieves mAP50 scores of 86.87% and 74.65%, respectively, demonstrating clear improvements over baseline YOLO models. Furthermore, comparisons with other mainstream detection methods show that Hyper-YOLOn-MDS maintains a favorable balance between accuracy, lightweight design, and deployment efficiency. Visualization results further confirm its robustness and generalization capability in complex underwater environments.
No takes yet. Share an insight, caveat, or question.
Yang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: