Toward high-accuracy underwater object detection in optically challenging environments via dynamic adaptive optimization.

Achieving reliable object perception in underwater environments characterized by turbidity, low-light, and spectral distortion remains a critical challenge for vision-based applications. To address this, we propose a novel, to our knowledge, dynamic adaptive object detection model. Our core innovation lies in a collaborative optimization framework that integrates a dynamic feature fusion module (utilizing DySample for upsampling) and a dynamic convolution module (DyC2f) to adaptively enhance multi-scale feature representation. Furthermore, we design a tailored loss function (FWNWD) that combines a WIoU dynamic focusing mechanism, a Focaler-IoU weighting strategy, and the normalized Wasserstein distance (NWD) to significantly improve localization robustness, particularly for blurry and small objects. Extensive evaluations on the large-scale DUO dataset (containing 74,515 instances) demonstrate the effectiveness of our approach. The model achieves a state-of-the-art mAP@0.5 of 91.7% with only 6.6G FLOPs, and runs at 23 FPS on an embedded platform (Orange Pi Aipro), successfully balancing high accuracy with real-time efficiency. Ablation studies confirm the contribution of each component, notably showing a 2.3% increase in recall for small objects attributable to the FWNWD loss. This work provides a practical and efficient solution for high-precision underwater perception, and its dynamically tunable architecture offers valuable insights for vision systems in other complex, degraded environments.