English

Journal of Chinese Agricultural Mechanization

Journal of Chinese Agricultural Mechanization ›› 2022, Vol. 43 ›› Issue (1): 158-166.DOI: 10.13733/j.jcam.issn.20955553.2022.01.023

Previous Articles     Next Articles

Fluecured tobacco leaf grading method based on deep learning and multiscale feature fusion

Lu Mengyao, Zhou Qiang, Jiang Shuwen, Wang Cong, Chen dong, Chen Tianen.    

  • Online:2022-01-15 Published:2022-02-17

基于深度学习与多尺度特征融合的烤烟烟叶分级方法

鲁梦瑶1,周强2,姜舒文1, 3,王聪1,陈栋1, 3,陈天恩1, 3   

  1. 1. 国家农业信息化工程技术研究中心,北京市,100097; 2. 安徽皖南烟叶有限责任公司,安徽宣城,242000;
    3. 农芯(南京)智慧农业研究院,南京市,211800
  • 基金资助:
    安徽皖南烟叶有限责任公司科技重点项目(20180563006)

Abstract: In order to realize the rapid and accurate recognition of fluecured tobacco grade, reduce the influence of subjective factors on the grading results during manual grading, and improve the accuracy and consistency of tobacco grading, a tobacco image grading method based on multiscale feature fusion of RGB image and deep learning was proposed. The attention mechanism (SENet) was introduced to enhance the importance of different channel features. Meanwhile, FPN (Feature Pyramid Network) was used to fuse the extracted tobacco leaf features at different levels from shallow to deep to realize the expression of tobacco leaf multiscale features. The front and the back images of a total of 6 068 fluecured tobacco in southern Anhui were collected for modeling and analysis. The results indicated that the tobacco grading accuracy of the proposed method was 5.21% higher than that of the classical CNN (Convolutional Neural Network). The final grading accuracy of the model was 80.14% on the new batch of 7 grades of tobacco leaves, and the adjacent grading accuracy was 91.50%. Therefore, using RGB images combined with deep learning technology can accurately identify the grade of fluecured tobacco leaves, which provides a new method for the evaluation of fluecured tobacco grades in purchasing.

Key words: tobacco grading, deep learning, image classification, feature fusion, feature pyramid network, SE Net

摘要: 为实现烤烟等级的快速准确识别,降低人工分级中主观因素对分级结果的影响,提高烟叶分级的准确性和一致性,提出一种基于烤烟RGB图像和深度学习的多尺度特征融合的烟叶图像等级分类方法,采用ResNet50提取烟叶图像特征,并引入基于注意力机制的SE模块(压缩激发模块),增强不同通道特征的重要程度;同时,采用FPN(特征金字塔网络)对提取的由浅及深不同层级的烟叶特征进行融合,以实现烟叶多尺度特征的表达。采集皖南地区6 068个烤烟的正面和背面图像用于建模和分析。结果表明,提出的烟叶分级方法的分级正确率比经典CNN(卷积神经网络)高出5.21%,分级模型在新批次7个等级烟叶上的分级正确率为80.14%,相邻等级的分级正确率为91.50%。因此,采用RGB图像结合深度学习技术可实现烤烟烟叶等级的良好识别,可为烤烟烟叶收购等级评价提供一种新方法。

关键词: 烟叶分级, 深度学习, 图像分类, 特征融合, 特征金字塔网络, SE模块

CLC Number: