Evaluating multi-source training strategies at the data, pixel, object, and training levels demonstrates that synthetic data effectiveness is not inherent but conditional on how it is generated, structured, and integrated relative to the target architecture and available real data, providing practical guidance for scaling UOD in marine monitoring applications.
Abstract
Underwater Object Detection (UOD) is essential for marine monitoring, ecological assessments, and autonomous underwater operations, yet progress is significantly hindered by the scarcity, cost, and complexity of collecting and annotating real-world underwater imagery. Synthetic datasets generated from high-fidelity 3D environments offer a promising alternative, but the domain gap caused by unrealistic visual conditions and incomplete feature distributions limits their applicability. This study systematically evaluates multi-source training strategies at the data, pixel, object, and training levels, comparing two deep learning architectures, Faster R-CNN and YOLOv11, and evaluating performance using mean Average Precision (mAP) across five-fold cross-validation and an external test set. Results show that synthetic data is most beneficial in data-scarce settings, preventing model collapse in two-stage detectors (TSD) and improving overall mAP by up to 35.4% in single-shot detectors (SSD). When integrating synthetic data into the full real-world dataset, realism-enhancing pre-processing is most beneficial for SSD, improving performance by 8.0% over the real-data baseline, whereas hybrid composition provides larger gains for TSD (9.9% over the baseline), highlighting the importance of object-level realism and spatial consistency in region-based detectors. Curriculum learning yields more modest but consistent improvements for both detectors (TSD: 2.7%; SSD: 4.7%). Overall, the findings demonstrate that synthetic data effectiveness is not inherent but conditional on how it is generated, structured, and integrated relative to the target architecture and available real data, providing practical guidance for scaling UOD in marine monitoring applications.
Underwater target detection is hindered by light scattering, low contrast, blur, occlusion, and background clutter, which particularly affect small objects. To address these limitations, we propose RCG-YOLO, a YOLOv8n-based detector that improves feature preservation, multi-scale feature extraction, and cross-scale feature fusion. Specifically, the proposed model introduces a residual down-sampling (ResDown) module to preserve fine spatial information during downsampling, a gated attention and multi-scale extractor (GAME) module to strengthen multi-scale feature extraction and positional sensitivity, and a cross-scale gated fusion (CGF) module to selectively fuse shallow spatial features and deep semantic features. Experiments on three public underwater datasets shows consistent improvements over YOLOv8n. RCG-YOLO achieves mAP50 scores of 85.2 ± 0.03%, 86.5 ± 0.02%, and 85.0 ± 0.03% on UTDAC2020, DUO, and RUOD datasets, corresponding to absolute mAP50 gains of 3.4%, 3.9%, and 1.7% over the baseline, respectively. Ablation studies confirm that ResDown, GAME, and CGF each contribute to the final detection performance. Although RCG-YOLO maintains real-time inference on an NVIDIA RTX 4090, its higher GFLOPs and lower frames per second compared with YOLOv8n indicate that the method prioritizes accuracy over maximum inference efficiency. These results indicate that RCG-YOLO provides a practical accuracy improvement over YOLOv8n for underwater object detection.
Samuel Atta Antwi, Zhiyu Zhou, Haiyan Wang· Engineering Research Express· 0 citations
Detecting objects underwater remains a challenging problem due to light absorption, scattering effects, low visibility, color distortion, and object occlusion in complex marine environments. Recently, deep learning models based on YOLO have demonstrated promising real-time detection capabilities. However, their performance degrades in dense underwater scenes with overlapping objects. In this work, we propose YOLOv8NX, an improved underwater object detection framework. An X-Attention mechanism is integrated into the YOLOv8 architecture to enhance spatial feature discrimination and localization accuracy. The proposed model was evaluated on an open-source underwater object detection dataset containing 5,542 annotated images using YOLOv5, YOLOv7, YOLOv8, and YOLOv11. All models were trained and tested under identical experimental settings using Precision, Recall, mAP@0.5, and mAP@0.5:0.95 metrics. The experimental results demonstrate that YOLOv8NX achieved the highest precision of 83.77%, resulting in cleaner and more distinguishable bounding boxes in dense underwater scenes with overlapping objects. The proposed model demonstrated improved localization reliability and spatial separation in challenging underwater conditions, even though YOLOv8 had higher recall and mAP values. The results indicate that the proposed attention- enhanced framework can support more robust real-time underwater biodiversity monitoring and intelligent marine observation systems.
Sanjey S, Janarthan B, Sam V George· 2026 6th International Confe...· 0 citations
Sonar object detection is constrained by scarce instance-level annotations, high acquisition costs, and long-tailed category distributions in real underwater environments. To address these limitations, this paper proposes a generative data augmentation framework based on the Stable Diffusion Model (SDM) for synthesizing sonar images together with target bounding boxes. The framework first fine-tunes SDM with an instance-level slice cropping strategy to strengthen the alignment between text prompts and local acoustic target structures. It then introduces a cross-modal sparse localization module (CMSL), which uses denoising features and text priors to infer 2D bounding boxes for generated samples. Synthetic long-tail samples are mixed with real URPC2022 data under a fixed-ratio saturation compensation strategy and evaluated through UTD-SCnet fine-tuning. The results show that instance-level cropping provides the best generation quality among the tested strategies (FID = 31.42, DR = 0.815, CCSCR = 0.075), and that a 60% synthetic-data injection ratio yields the best detection performance. These findings indicate that diffusion-based augmentation can provide a practical, semi-automated supplement for long-tail sonar detection, while excessive synthetic data may introduce domain-shift effects.
Reliable underwater object detection (UOD) is critical for applications including marine ecosystem monitoring, aquaculture management, and subsea infrastructure inspection. However, underwater imagery is severely affected by wavelength-dependent attenuation, scattering, turbidity, low illumination, and sensor noise, which collectively degrade visual quality and hinder detection accuracy. Recent advances in deep learning, particularly within the YOLO family of one-stage detectors, have enabled real-time detection with improved robustness under challenging conditions. Despite rapid progress, a clear synthesis of how the latest YOLO variants address underwater-specific degradations remains limited. This study presents a structured review of underwater object detection approaches based on YOLOv8 to YOLOv12, covering both optical imaging and side-scan sonar modalities. A systematic literature review was conducted following PRISMA-based criteria, resulting in the analysis of recent primary studies with emphasis on architectural design, training strategies, datasets, and evaluation protocols. The reviewed methods are organized into a taxonomy distinguishing detector-level innovations from enhancement-assisted pipelines, including physics-guided preprocessing, learning-based restoration, and joint enhancement–detection frameworks. Comparative analysis indicates that improvements in feature fusion, attention mechanisms, lightweight backbone design, and small-object detection strategies significantly enhance performance in degraded underwater conditions. However, persistent challenges remain, particularly in cross-domain generalization, reproducibility, and limited reporting of deployment metrics. Emerging research directions include task-driven enhancement, physics-informed data augmentation, and multi-modal learning approaches. This review provides a consolidated perspective and practical guidance for developing robust and deployable underwater detection systems.