Research on Underwater Vision Enhancement and Detection Algorithms for Sea Cucumber Harvesting Robots Oriented Towards Edge Computing

Abstract During autonomous underwater operations, sea cucumber harvesting robots rely on edge computing platforms for their vision modules; the multiple constraints of platform computational capacity, power consumption, and multi-task parallelism impose demanding requirements on the resource efficiency and operational stability of visual algorithms.Targeting the Jetson Orin Nano edge platform, this study proposes a lightweight visual enhancement and detection algorithm. To address the color cast and low contrast of underwater images, an improved Auto Level image enhancement algorithm is designed that effectively restores color balance and image details at low computational cost; at the network level, building upon the YOLOv11 framework, the ADown efficient downsampling module and the lightweight GhostConv module are introduced, substantially reducing the parameter count and computational complexity of the model. The experiment was carried out using a private dataset built from 622 sea cucumber images collected by an underwater robot platform in real coastal farming waters. Experimental results show that the optimized model achieves an inference speed of 33.12 FPS on the Jetson Orin Nano, an 8.4% improvement over the unenhanced YOLOv11n baseline; GFLOPs are reduced to 4.82, a 25.2% reduction, and mAP50-95 reaches 63.54%, a decrease of 1.64 percentage points relative to the baseline. Within an acceptable range of accuracy loss, the approach realizes efficient real-time detection of sea cucumber targets.