Adaptive multi-scale attention networks fusion in YOLOv12-LSE: a lightweight framework for efficient and robust underwater object detection

Underwater object detection is difficult for existing models to balance accuracy and efficiency due to problems such as complex optical interference, blurred features of small targets, and the limited computing power of edge devices. To address the limitations of YOLOv12 in underwater scenarios, including insufficient multi-scale feature modeling (mAP50: 0.652) and high computational complexity (21.2 GFLOPs), we propose YOLOv12-LSE, a lightweight framework that achieves the collaborative optimization of underwater detection performance and efficiency through three innovations: First, the C2PSA_LSKA module is designed, integrating channel-space joint attention and lightweight large-core decomposition strategies to enhance the multi-scale object detection capability and suppress edge blur interference. Second, the Slim Neck architecture is constructed based on gradient sparse convolution and cross-level feature aggregation to reduce computational redundancy and retain shallow detail features. Finally, efficient channel attention is introduced into the detection head to improve the feature discrimination of low-contrast targets through adaptive channel selection. Experiments on the DUO dataset show that the mAP50 of YOLOv12-LSE reaches 0.681, significantly improving by 4.45% compared to the baseline model YOLOv12. The precision increases by 6.3%, while the GFLOPs decrease by 7.08% to 19.7, the lowest in the same series. The ablation experiment verified the collaborative optimization effect of each module, and the visualization results further proved its robustness in low-illumination, dense targets, and edge-blurred scenes. This study breaks through the bottleneck of the “precision-efficiency” trade-off with a lightweight design, providing a high-precision real-time detection solution for underwater mobile devices (such as AUVs) and promoting the deep integration of the attention mechanism and edge computing.