AdaptiveFocus-DETR: Enhanced Real-Time Detection with Adaptive Feature Focus

Underwater object detection is crucial for marine exploration, environmental monitoring, and mission execution by autonomous underwater vehicle (AUV). To address challenges such as low visibility, complex backgrounds, and morphological diversity of targets—which lead to missed detections, false positives, and low detection accuracy—this paper proposes AdaptiveFocus-DETR, a novel model incorporating three key innovations: The proposed model first introduces a feature extraction module, Dynamic Snake Convolution Block, significantly enhanced the feature extraction capability for underwater targets. Secondly, the encoder combines the Adaptive Sparse Self-Attention with the original AIFI module, enabling the model to focus on discriminative feature regions while suppressing noise in complex underwater environments. Finally, a hybrid loss function that improves bounding box localization accuracy in challenging scenarios by integrating MPDIoU with FocalerIoU, effectively addressing misalignment issues in cluttered or occluded scenes. Experiments on the URPC2020 dataset demonstrate that AdaptiveFocus-DETR achieves 0.7%, 1%, and 1.9% improvements in F1 score, mAP@50, and mAP@50:95. Furthermore, it outperforms mainstream detection frameworks in mAP@50. These results underscore the efficacy of AdaptiveFocus-DETR in advancing underwater object detection, offering a robust solution for real-world applications in aquatic environments.