Underwater Object Detection Algorithms Using Deep Learning Techniques

Underwater object detection has emerged as a pivotal field for marine exploration, environmental monitoring and underwater robotics. The unique challenges of aquatic environments—such as light scattering, colour distortion, turbidity and the presence of small or occluded targets—demand specialised deep learning approaches. Contemporary algorithms typically integrate advanced convolutional neural network architectures with pre- and post-processing strategies to enhance detection accuracy and computational efficiency. Pre-processing steps often include image enhancement, dehazing and colour correction to mitigate water-induced visibility degradation. Within the network, multi-scale feature extraction and attention mechanisms allow detectors to focus on salient marine objects across varying sizes and depths. State-of-the-art models employ lightweight backbones to enable real-time performance on embedded platforms, while hybrid frameworks leverage fusion of acoustic and optical data for improved detection in turbid conditions. Benchmark evaluations on datasets such as DUO, Brackish and underwater extensions of VOC demonstrate substantial gains, with mean average precision values exceeding 80% in controlled settings. Practical applications span autonomous underwater vehicles for seabed mapping, real-time monitoring of coral reef health and detection of marine species for conservation. Future directions include unsupervised domain adaptation to bridge the gap between synthetic and real images, and the integration of temporal information from video streams to enhance robustness against dynamic underwater scenes.

Research from Nature Portfolio

No recent Nature Portfolio content available.

Research from all publishers

A 2024 study adapted the YOLOv4 architecture for a seven-class marine object dataset, incorporating specialised colour-space transformations and targeted data augmentations. This approach achieved robust detection of fish, corals and underwater structures, demonstrating that tailored pre-processing can substantially improve performance under varying lighting and turbidity. Another 2024 contribution introduced Dynamic YOLO, which combines deformable convolutional layers with a unified multi-scale attentional feature fusion framework to address small-target detection. By disentangling classification and localisation tasks in the detection head, this model outperformed prior methods on the DUO benchmark, yielding significant gains in average precision for small objects. Foundational work from 2021 proposed a lightweight detector built on MobileNet v2 and depth-wise separable convolutions, coupled with a multi-scale attentional feature fusion module. This design balanced accuracy and speed, achieving above 81% mAP on PASCAL VOC-derived underwater datasets with real-time frame rates, while compressing model size by over 80% relative to conventional YOLOv4. Collectively, these studies illustrate a trend towards modular architectures integrating attention, deformable kernels and efficient backbones to meet the stringent requirements of underwater deployment.

Underwater Object Detection Algorithms Using Deep Learning Techniques publication trend

The graph below shows the total number of articles in underwater object detection algorithms using deep learning techniques across all publications each year (not limited to Nature Index journals).

Technical terms

Convolutional Neural Network (CNN): A class of deep learning models that apply convolutional filters to extract hierarchical features from images.

You Only Look Once (YOLO): A real-time object detection framework that predicts bounding boxes and class probabilities in a single forward pass.

Deformable Convolution: An adaptation of standard convolution that allows the sampling grid to shift, improving alignment with object shapes.

Attention Mechanism: A network component that weights feature responses to focus on the most informative regions of an image.

Mean Average Precision (mAP): A standard metric that summarises detection accuracy by averaging precision values across recall levels and object classes.

Depth-wise Separable Convolution: A factorised convolution operation that reduces parameters and computational cost by separating spatial and channel-wise filtering.

References

1. Marine Object Detection using YOLOv4 Adapted Convolutional Neural Network. Decision Making Advances (2024).
2. Dynamic YOLO for small underwater object detection. Artificial Intelligence Review (2024).
3. Lightweight Underwater Object Detection Based on YOLO v4 and Multi-Scale Attentional Feature Fusion. Remote Sensing (2021).