Authors

Abstract

To address the problems of low efficiency, strong subjectivity, and high cost associated with traditional tobacco leaf detection methods, an identification model suitable for tobacco leaves was developed to achieve rapid and non-destructive detection and to support the standardization of tobacco production. In this study, the Convolutional Block Attention Module (CBAM) was improved to enhance the feature extraction capability of tobacco leaves and highlight key feature information. MobileNetV3 and a Deformable Convolution Network were integrated to optimize the model structure, thereby reducing the number of parameters and computational complexity. Based on these improvements, a tobacco leaf detection model was constructed. Experimental results showed that the proposed algorithm achieved an accuracy of 97.01% with a loss value of 0.09, outperforming the Multi-Marker Similarity Assessment method and the Gorilla Troop Optimization Algorithm. The constructed model achieved an accuracy of 94.65%, a recall of 91.24%, and an F1 score of 93.54%. The model contains 9.36 M parameters and has a size of 50.69 MB, demonstrating better performance compared with the reference models. The results indicate that the improved tobacco leaf detection model can significantly enhance detection efficiency and accuracy. This study provides a useful approach for precise tobacco leaf detection in complex field environments and contributes to the development of modern intelligent agriculture.

Abstract in Chinese

烟叶检测在现代农业发展中发挥着重要作用。准确检测烟叶可以降低劳动力成本，减少资源浪费。然而，传统的检测方法往往存在效率低、主观性强、运行成本高等问题。因此，本研究改进了卷积块注意力模块，优化烟叶特征提取能力，突出关键特征信息，并结合 MobileNetV3 和可变形卷积网络优化模型结构，降低参数和计算负荷。最后，构建了烟叶检测模型。实验结果表明，所提出的算法准确率达到 97.01%，损失值为 0.09，优于多标记相似性评估算法和猩猩部队优化算法。所构建模型的准确率为 94.65%，召回率为 91.24%，F1 得分为 93.54%。其参数和模型大小分别为 9.36 M 和 50.69 MB，表现优于对比模型。这些结果表明，改进后的烟叶检测模型可以提高检测效率和准确性。这项研究有助于在复杂的田间环境中精确构建烟叶检测模型，促进现代农业的发展。