UFPN-YOLO: an enhanced lightweight object detector with underwater feature pyramid network

Deep learning–based underwater object detection has become increasingly important as demand grows for marine resource exploration, and autonomous environmental monitoring. Nevertheless, the unique characteristics of underwater environments such as image blurriness caused by light scattering and attenuation, the wide variation in object sizes and shapes, and the limited computational capacities of edge-computing devices, present significant challenges for robust, and accurate detection. Underwater feature pyramid network (UFPN)-YOLO, a lightweight underwater object detection network based on YOLOv11n is proposed in this paper. At the core of the model is an UFPN that integrates a skip feature adaptive integration unit and an information supplement pathway to strengthen feature perception, particularly for blurred and small objects common in underwater imagery. Furthermore, a dual-path efficient downsampling module is introduced to enhance multi-scale feature representation. In the backbone network, a deformable convolution-based C3k2-DCN module is employed, together with a C2-single-head self-attention (SHSA) module that integrates a SHSA mechanism. These components enhance the detection for irregularly shaped objects while reducing computational redundancy. Experiments on the RUOD dataset demonstrate that UFPN-YOLO achieves 86.9% mAP₅₀ and 66.3% mAP_50–95, while using only 2.5 M parameters. On the UTDAC2020 dataset, it achieves 85.3% mAP₅₀ and 64.1% mAP_50–95. Such performance substantially surpasses that of existing YOLO variants and other state-of-the-art underwater object detection models.