Underwater Object Detection in the Deep Learning Era: Techniques, Benchmarks and Challenges

Abstract

The proposed system can be used to detect underwater objects, which is essential for autonomous underwater exploration, biodiversity monitoring, and the fulfilment of the United Nations Sustainable Development Goal (SDG) 14. Even though deep learning models trained on terrestrial imagery work well, severe visual degradation stemming from light scattering, colour absorption, turbidity, and changing illumination in underwater environments fundamentally limits their applicability. This structured literature review compares the CNN-based models (YOLOv8, YOLONAS, SSD, RetinaNet, Faster R-CNN, Mask R-CNN), transformer-based models (Swin Transformer, DETR, DINO) and hybrid CNN-Transformer architectures, and evaluates their results in a unified four-dimensional framework: detection accuracy (mAP), environmental robustness under compound underwater disturbances, computational efficiency (FPS, power consumption), and edge deployability on AUV hardware. The comparative analysis shows that various architectural paradigms have various advantages and disadvantages in the underwater environment. Transformer-based architectures tend to achieve better detection performance and robustness in visually degraded environments, thanks to stronger contextual modelling. The YOLO-based architectures have shown a good trade-off between detection accuracy, computational efficiency, and deployment suitability, making them attractive for real-time AUV applications. Two-stage detectors are well suited for precision-oriented inspection and segmentation applications, while hybrid CNN-Transformer networks offer greater robustness in turbid, visually complex underwater environments. Model selection must be environment- and mission-specific. Four research gaps are identified: architecture design, data engineering, deployment optimisation, and robustness evaluation, each with specific unsolved problems, including the absence of cross-paradigm benchmarks, compounded dataset biases, and a lack of compound-condition robustness testing. This structured literature review provides condition-specific model selection guidelines aligned with SDG 14 and identifies future research priorities including selfsupervised learning, sensor fusion, and lightweight AUV-optimised architectures.