Authors

Abstract

Corn seedling and weed detection during the 2–5-leaf stage is essential for precision weeding, but similar morphology, field background interference, and the accuracy-efficiency trade-off limit the practical use of lightweight detectors. To address these problems, this study proposes Corn-Weed-ENAS, an improved YOLOv8-based detection model with joint structure and loss-function search. A differentiated ENAS search space was constructed for the backbone, neck, detection head, and loss function to automatically identify a compact architecture suitable for corn seedling-weed detection. A field image dataset of 2–5-leaf corn seedlings and weeds were collected in Heilongjiang Province, and comparative and ablation experiments were conducted against mainstream detectors. The proposed model achieved 97.3% mAP@0.5 with 2.5M parameters and 6.7 GFLOPs, showing a favorable balance between detection accuracy and lightweight deployment. To further verify its agricultural engineering applicability, an indoor simulated operation test was conducted on a Jetson Orin Nano Super platform using 20 pots of corn seedlings and weeds and a moving test vehicle at 1.2 m/s. In the virtual spraying signal verification, YOLOv8n produced five missed detections and one false detection, whereas Corn-Weed-ENAS produced three missed detections and no false detection. These results indicate that Corn-Weed-ENAS can provide more reliable weed localization and virtual trigger outputs for selective spraying decision support.

Abstract in Chinese

玉米 2~5 叶期苗草识别是精准除草的重要基础，但玉米幼苗与杂草形态相近、田间背景干扰复杂，且轻量化检测模型常存在精度与效率难以兼顾的问题。针对上述问题，本文提出一种结合结构搜索与损失函数搜索的 YOLOv8 改进模型 Corn-Weed-ENAS。该方法面向骨干网络、颈部网络、检测头和损失函数构建差异化 ENAS 搜索空间，自动获得适用于玉米苗草检测任务的紧凑型网络结构。本文采集黑龙江省 2~5 叶期玉米幼苗与杂草图像数据集，并与主流目标检测模型开展对比和消融试验。结果表明，所提出模型在 2.5M 参数量和 6.7 GFLOPs 条件下取得 97.3% mAP@0.5，实现了检测精度与轻量化部署之间的较好平衡。为进一步验证其农业工程应用适应性，本文基于 Jetson Orin Nano Super 平台开展了室内模拟作业试验，采用 20 盆玉米苗与杂草盆栽以及 1.2 m/s 移动试验车进行虚拟喷药信号验证。结果显示，YOLOv8n 出现 5 次漏检和 1 次误检，而 Corn-Weed-ENAS 仅出现 3 次漏检且无误检。结果说明，Corn-Weed-ENAS 能够为选择性喷药决策支持提供更可靠的杂草定位与虚拟触发输出。