AF-DETR: A Dual-Stream Frequency Enhancement and Blur-AdaptiveTransformer for Underwater Object Detection

Underwater object detection is essential for marine biodiversity monitoring and ecosystem assessment. However, underwater detection tasks remain challenging due to complex environments, insufficient illumination, scattering, low contrast, and target blurring, which collectively degrade image quality and reduce detection robustness. To address these issues, this paper proposes AF-DETR, a frequency-enhanced detection model tailored for underwater scenes. First, a frequency-domain separation and dual-stream feature perception module is introduced to independently process structural and textural components in the frequency domain, thereby restoring fine details and visual realism in degraded images. Second, a spectral-aware and blur-adaptive multiscale feature extraction backbone is designed, integrating frequency-domain priors with contextual perception to strengthen feature representation for blurred and scale-varying objects. Third, a feature refinement and enhancement mechanism is incorporated, combining joint channel–spatial attention with head–feature interaction and a scale-guided attention module, effectively improving multiscale localization and recognition accuracy. Experimental results on URPC2020 dataset demonstrate that AF-DETR achieves an mAP@0.5 of 81.3. On DUO dataset, it attains an mAP@0.5 of 88.6, validating the effectiveness and superiority of the proposed approach.