SvelteF3M: toward a new accuracy–efficiency trade-off for real-time underwater detection for edge deployment

Underwater object detection for real-world robotic systems requires both strong robustness against severe visual degradation and efficient inference under resource-constrained edge deployment settings. However, existing methods often improve detection accuracy at the cost of increased computational complexity, making them difficult to deploy in practice. In this paper, we propose SvelteF3M-YOLO, a lightweight underwater object detector that jointly addresses degradation-aware representation and deployment efficiency within a unified framework. The model integrates a streamlined Frequency-Fused Feature Module (F3M) to enhance high-frequency structural cues while suppressing low-frequency interference, together with a hardware-friendly lightweight neck (SvelteNeck) to enable efficient multi-scale feature propagation. Extensive experiments on three representative underwater datasets, including SCoralDet, TrashCan-Instance, and URPC2019, show that SvelteF3M-YOLO maintains competitive mAP 50–95 with reduced parameters and GFLOPs. Real-time deployment experiments on an NVIDIA Jetson Orin Nano Super 8GB further show that the proposed model achieves a mean latency of 4.52 ms and a throughput of 241.63 FPS under TensorRT FP16 inference, compared with 4.56 ms and 239.96 FPS for the YOLO11n baseline. Meanwhile, on the TrashCan-Instance dataset, SvelteF3M-YOLO improves mAP 50–95 from 0.680 to 0.695 over YOLO11n. These results indicate that the proposed framework improves underwater detection accuracy while maintaining real-time edge deployment efficiency, providing a practical solution for resource-constrained underwater perception.