SvelteNeck by EHConv: A Cross-Generational Lightweight Neck for Real-Time Object Detection

Efficient object detection is vital for Remotely Operated Vehicles (ROVs) performing marine debris cleanup, yet existing lightweight designs frequently encounter efficiency bottlenecks when adapted to deeper neural networks. This research identifies a critical “Inverted Bottleneck” anomaly in the Slim-Neck architecture on the YOLO11 backbone, where deep-layer Memory Access Cost (MAC) abnormally spikes. To address this, we propose SvelteNeck-YOLO. By incorporating the proposed EHSCSP module and EHConv operator, the model systematically eliminates computational redundancies. Empirical validation on the TrashCan and URPC2019 datasets demonstrates that the model resolves the memory wall issue, achieving a state-of-the-art trade-off with only 5.8 GFLOPs. Specifically, it delivers a 34% relative reduction in computational load compared to specialized underwater models while maintaining a superior Recall of 0.859. Consequently, SvelteNeck-YOLO establishes a robust, cross-generational solution, optimizing the Pareto frontier between inference speed and detection sensitivity for resource-constrained underwater edge computing.