A Lightweight Edge Computing Framework for Real‐Time Weed Detection and Laser Control in Maize Fields Using HRE ‐ YOLO

ABSTRACT

Laser weeding is an emerging non‐chemical technology for precision agriculture, but its practical field deployment requires a real‐time vision system capable of accurately detecting weeds, localizing target points, and driving site‐specific laser execution under complex crop‐field conditions. This study proposes an edge‐deployable vision‐guided laser weeding framework for maize fields based on HRE‐YOLO. The framework integrates visual perception, weed detection, target point localization, laser positioning, and execution control into a closed‐loop system for real‐time laser weeding. To improve detection robustness in complex field backgrounds, the proposed HRE‐YOLO model incorporates feature interaction enhancement, global feature correlation modeling, and small‐target feature enhancement modules, which strengthen the representation of small and partially occluded weeds while maintaining a lightweight network structure. Based on the detection results, laser positioning and execution strategies were designed to protect maize seedlings, adapt to different weed density levels, and reduce unnecessary laser energy consumption. A maize‐field laser weeding platform was constructed and validated through field detection and laser treatment experiments. Experimental results showed that HRE‐YOLO achieved a detection precision of 90.82% and an mAP@0.5 of 93.30%, outperforming the original model by 4.15 and 5.25 percentage points, respectively, while using only 3.1 M parameters. The complete laser weeding system achieved a weed control rate of 91.53%, demonstrating its potential for real‐time, accurate, and energy‐efficient weed control in maize fields. The results indicate that the proposed framework provides a feasible edge‐intelligent solution for precision laser weeding and establishes a foundation for further optimization toward practical agricultural applications.