Authors

Abstract

Classification of barley plants plays a crucial role in understanding barley varietal diversity and breeding. Traditional classification methods rely on expert experience and require significant manual effort. With the rise of deep learning based on machine vision technologies, particularly the emergence of transfer learning, the issue of model overfitting on small datasets has been mitigated, leading to enhanced generalization capabilities. This study employs a self-constructed barley plant image dataset to compare five state-of-the-art deep learning models, while analyzing the impacts of various factors - including image resolution and training-test split ratio - on classification accuracy. The results indicate that the DenseNet model achieves the best classification performance at an input resolution of 512×512 pixels, with an accuracy of 96.02%. Increasing the proportion of training data further improved performance, with the 80%:20% training-test split ratio yielding optimal results across all five models. Transfer learning models outperform training from scratch, with EfficientNet-v2 achieving the highest accuracy of 98.86%. Additionally, gradient-weighted class activation mapping (Grad-CAM) was utilized to generate heatmap visualizations of the decision-making processes in each transfer learning model. By applying deep learning for barley plant classification and selecting the optimal model, this research provides a reliable technical solution for barley variety identification and classification.

Abstract in Chinese

大麦植株分类对于理解大麦品种多样性和育种研究具有关键作用。传统分类方法依赖专家经验且需要大量人工操作。随着基于深度学习的机器视觉技术的兴起，尤其是迁移学习技术的出现，减轻了在小数据集上模型过拟合的现象，提高了模型的泛化能力。本研究采用自建的大麦植株图像数据集，对比了五种先进的深度学习模型，并分析了不同图像分辨率、不同训练集-测试集比例等因素对分类准确率的影响。研究结果表明：DenseNet模型在512×512像素输入分辨率下表现最优，分类准确率达96.02%；增加训练数据比例能提升模型性能，80%: 20%的训练-测试划分比例在五种模型中均取得最佳效果；使用迁移学习模型显著优于从头训练，其中EfficientNet-v2模型以98.86%的准确率表现最佳。此外，研究还采用Grad-CAM技术对五种迁移学习模型的预测过程进行了热力图可视化分析。本研究通过深度学习技术实现大麦植株分类并优选最佳模型，为大麦品种鉴定与分类提供了可靠的技术方案。