Underwater object detection is confronted with critical challenges such as extreme scale variations, the interplay between fine-grained and global features, high miss-detection rates for small targets, and severe background interference. This paper presents AACF-Net, a novel detection network featuring three core innovations including a CD multi-kernel residual module that leverages heterogeneous 3 \(\times \) 3+5 \(\times \) 5 convolutions to simultaneously capture local details and global morphology, an ACF-Net adaptive fusion mechanism with Sigmoid-based dynamic weight allocation, and a BiFPN-ASF collaborative mechanism that integrates bidirectional cross-scale fusion with channel-spatial attention suppression. A P2 high-resolution feature layer is further incorporated to quadruple the feature map resolution, thereby fundamentally mitigating information loss in small-object detection. Experiments on the URPC2020 dataset demonstrate that AACF-Net achieves 72.5% mAP50 and 45.8% mAP50-95 with merely 3.3M parameters and 385.8 FPS, outperforming YOLOv11 by 4.2% in mAP50 and reducing the false detection rate by 6.6% in complex scenes. Comprehensive ablation studies verify the contribution of each module, while cross-dataset evaluations on the RUOD and COCO2017 datasets confirm the strong generalization ability of the proposed network.
