DyAqua-YOLO: a high-precision real-time underwater object detection model based on dynamic adaptive architecture

With the rapid development of marine resource exploitation and the increasing demand for underwater robot inspection, achieving reliable target perception in turbid, low-illumination, and spectrally limited underwater environments has become a key challenge that urgently needs to be addressed in the field of computer vision. This paper proposes DyAqua-YOLO, a dynamic adaptive model specifically designed to address the critical challenges of low-contrast blurred targets and pervasive small-object detection in complex underwater optical environments. Central to our approach are three core innovations: 1) The Dynamic Scale Sequence Feature Fusion (DySSFF) module, which replaces static upsampling with a dynamic grid generator to preserve spatial details of blurred and small targets; 2) The DyC3k2 module, which introduces dynamic kernel weighting into the reparameterization process, enabling adaptive feature extraction for degraded underwater images; 3) A unified Focaler-Wise Normalized Wasserstein Distance (FWNWD) loss, not a mere combination but a hierarchical framework where WIoU provides gradient modulation, Focaler-IoU handles hard-easy sample bias, and NWD ensures small-object sensitivity, working in concert to resolve optimization conflicts. On the DUO dataset containing 74,515 instances, the DyAqua-YOLO model achieves mAP@0.5 of 91.8% and mAP@0.5:0.95 of 72.2%, demonstrating outstanding accuracy. Compared to the baseline (YOLO11n), these metrics have improved by 3.9% and 3.7%, respectively. On the OrangePi AIpro platform (8TOPS NPU, 16GB RAM), the enhanced model achieves an inference speed of 21 FPS, striking an optimal balance between accuracy and efficiency. Ablation experiments show that the DyC3k2 module increases mAP@0.5 by 1.2% and mAP@0.5:0.95 by 1.7% compared to the YOLO11 baseline model, while reducing FLOPs by 3.2%, thereby enhancing model accuracy and optimizing computational efficiency. The FWNWD loss function improves the recall of small targets by 3.6% compared to the CIoU loss function, effectively balancing the optimization conflicts between hard examples and small targets and improving localization accuracy. This research provides a new approach for high-precision real-time detection in underwater embedded devices, and its dynamic and adjustable architecture has broad applicability guiding value for application in other scenarios with similar challenges.