In this study, we investigated the application of three convolutional neural network models YOLOv5, YOLOR, and YOLOv7 for precisely detecting individual radish plants, radish rows, and weeds. A comprehensive dataset was created, capturing diverse conditions and annotated for three target classes: radish, radish-line, and weed. Through extensive experimentation involving 39 combinations of model types, batch sizes (2, 4, 8), and learning rates (0.1, 0.01, 0.001), we determined that the YOLOv5-x model with a batch size of 4 and a learning rate of 0.01 offers superior performance. This configuration achieved a remarkable 99% accuracy for the radish class, 98% for radish-line, and 91% for weed, as confirmed by confusion matrices. Further analysis using the F1-score, Precision-Recall (PR) curves, and training progress plots underscored the model's robustness, particularly its high mAP_0.5:0.95 score. Despite the Weed class posing greater detection challenges, likely due to its lower representation in the dataset, the YOLOv5-x outperformed YOLOR-D6 and YOLOv7-D6 in critical metrics after 300 epochs. This research not only highlights the efficacy of YOLOv5-x in agricultural applications but also suggests potential enhancements in data annotation and model training strategies to further improve weed detection. Our findings provide significant insights for developing automated, high-precision plant-weed detection systems, contributing to more efficient and sustainable agricultural practices.

中文翻译：

---

利用计算机视觉推进精准农业：用于杂草和作物识别的 YOLO 模型的比较研究


在本研究中，我们研究了三种卷积神经网络模型 YOLOv5 、 YOLOR 和 YOLOv7 在精确检测单个萝卜植株、萝卜行和杂草中的应用。创建了一个全面的数据集，捕捉了不同的条件，并针对三个目标类别进行了注释：萝卜、萝卜线和杂草。通过涉及模型类型、批量大小（2、4、8）和学习率（0.1、0.01、0.001）的 39 种组合的广泛实验，我们确定批量大小为 4 且学习率为 0.01 的 YOLOv5-x 模型提供了卓越的性能。这种配置对萝卜类实现了 99% 的准确率，对萝卜线的准确率达到了 98%，对杂草的准确率达到了 91%，混淆矩阵证实了这一点。使用 F1 分数、精度-召回率 （PR） 曲线和训练进度图的进一步分析强调了该模型的稳健性，尤其是其较高的 mAP_0.5：0.95 分数。尽管 Weed 类带来了更大的检测挑战，可能是因为它在数据集中的代表性较低，但在 300 个 epoch 后，YOLOv5-x 在关键指标上的表现优于 YOLOR-D6 和 YOLOv7-D6。这项研究不仅强调了 YOLOv5-x 在农业应用中的功效，还提出了数据注释和模型训练策略的潜在增强功能，以进一步改进杂草检测。我们的研究结果为开发自动化、高精度的植物杂草检测系统提供了重要的见解，有助于更高效和可持续的农业实践。