DAF-DETR: Dual-Attention and Dual-Fusion DETR for Underwater Object Detection

Underwater object detection plays a crucial role in marine resource exploration and ecological conservation. However, it suffers from significant challenges due to severe image blurring and low contrast, which substantially degrade the detection performance. To overcome these limitations, we propose DAF-DETR, a novel Detection Transformer framework with Dual Attention and Dual Fusion Modules (DFMs). DAF-DETR first introduces an Aggregated Attention Mechanism to enhance the ResNet residual blocks, which boosts both global context awareness and local detail extraction in the backbone network. Second, it incorporates a Deformable Attention-based Feature Interaction (DAFI) module to improve the discriminability between object and background features in low-contrast underwater images. Finally, a DFM, which integrates Global-to-Local Spatial Aggregation (GLSA) with Haar Wavelet-based Downsampling (HWD), is employed to effectively alleviate the adverse effects of severe underwater blurring. Extensive experiments on the DUO and Underwater Object Detection Dataset (UODD) datasets validate the effectiveness and robustness of the proposed DAF-DETR framework, demonstrating significant improvements over existing methods.