Abstract: Weed is one of the most important factors affecting sugarcane growth, in order to recognize different sugarcane weeds, a deep learning‑based weed detection method for sugarcane was proposed. Firstly, images were collected with weeds that were common and harmful to sugarcane growth in the Guangxi region, and the images were then enhanced with operations such as panning, flipping, cropping, shrinking, contrast and brightness adjustment and denoising. Following that, the dataset was tested with the YOLOv8n detection model and compared to the YOLOv3—tiny, YOLOv4—tiny, YOLOv5n, and Yolov7—tiny detection models. The experimental results show that the YOLOv8n detection model has a 98.3% precision, a 96.8% recall, and a 98.2% mAP. Precision is improved by 26.9%, 12.7%, 4.2%, and 9%, respectively, when compared to the current mainstream lightweight target detection algorithms YOLOv3—tiny, YOLOv4—tiny, YOLOv5s, and YOLOv7—tiny. The recall is improved by 25%, 24.5%, 6.1%, and 10.6%, respectively, and the mAP is improved by 25.9%, 19.4%, 3.6%, and 7.8%. Meanwhile, the YOLOv8n detection model can recognize sugarcane weeds with high accuracy in dense, occluded, interspersed weeds, small targets, and dim environments.

Key words: sugarcane, weed identification, YOLOv8 algorithm, object detection

摘要： 杂草是影响甘蔗生长的重要因素之一，为实现对不同甘蔗杂草的识别，提出一种基于深度学习的甘蔗杂草检测方法。以广西地区常见且对甘蔗生长危害较大的杂草为对象进行图片采集，并对采集的图片进行平移、翻转、裁剪、缩小、对比度和光亮调整及去噪等操作增强数据。利用YOLOv8n检测模型对数据集进行试验，并与YOLOv3—tiny、YOLOv4—tiny、YOLOv5n和Yolov7—tiny进行对比。试验结果表明，YOLOv8n检测模型的精确率为98.3%，召回率为96.8%，mAP为98.2%。与目前主流的轻量化目标检测算法YOLOv3—tiny、YOLOv4—tiny、YOLOv5s和YOLOv7—tiny对比，精确率分别提高26.9%、12.7%、4.2%和9%；召回率分别提高25%、24.5%、6.1%和10.6%；mAP分别提高25.9%、19.4%、3.6%、7.8%。同时，在密集、遮挡、杂草交错、小目标、昏暗环境的情况下YOLOv8n检测模型对甘蔗杂草能实现高精度识别，具有较强的鲁棒性。

关键词: 甘蔗, 杂草识别, YOLOv8n算法, 目标检测