欢迎访问《中国农学通报》，
中国农学通报

中国农学通报 ›› 2025, Vol. 41 ›› Issue (34): 157-164.doi: 10.11924/j.issn.1000-6850.casb2025-0426

• 农业信息·科技教育 • 上一篇    

基于深度学习的雪茄茄衣智能分级模型构建与优选

杜超凡1(), 汪睿琪2, 吴天翊2, 沈翠玉1, 沈福龙1, 赖日君1, 林晓路1, 马旭东1, 谢小芳2,3()   

  1. 1 福建省烟草公司龙岩市公司，福建龙岩 364000
    2 福建农林大学生命科学学院，福州 350002
    3 福建农林大学福建省作物设计育种重点实验室，福州 350002
  • 收稿日期:2025-05-30 修回日期:2025-11-20 出版日期:2025-12-04 发布日期:2025-12-04
  • 通讯作者:
    谢小芳，女，1978年出生，福建龙岩人，副教授，博士，研究方向：植物遗传学与基因组学。通信地址：350002 福建省福州市仓山区上下店路15号 福建农林大学，E-mail：。
  • 作者简介:

    杜超凡，男，1978年出生，福建龙岩人，农艺师，学士，研究方向：烟叶栽培。通信地址：364000 龙岩市烟草公司新罗区龙岩大道288号，Tel：0597-2999908，E-mail：。

  • 基金资助:
    福建省烟草公司龙岩市公司科技计划项目“雪茄烟和烤烟分级智能化检测技术研究”(LK-2022Y06); 福建省烟草公司龙岩市公司科技计划项目“雪茄烟发酵过程的微生物作用机制及其工艺调控研究”(LK-2022Y02); 中国烟草总公司科技计划项目(110202201028(LS-12))

Construction and Screening of Intelligent Grading Model of Cigar Leaves Based on Deep Learning

DU Chaofan1(), WANG Ruiqi2, WU Tianyi2, SHEN Cuiyu1, SHEN Fulong1, LAI Rijun1, LIN Xiaolu1, MA Xudong1, XIE Xiaofang2,3()   

  1. 1 Fujian Tobacco Company Longyan Branch, Longyan, Fujian 364000
    2 College of Life Sciences, Fujian Agriculture and Forestry University, Fuzhou 350002
    3 Fujian Key Laboratory of Crop Breeding by Design, Fujian Agriculture and Forestry University, Fuzhou 350002
  • Received:2025-05-30 Revised:2025-11-20 Published:2025-12-04 Online:2025-12-04

RichHTML

1

PDF (PC)

6

摘要：

本研究旨在针对雪茄烟叶等级判定环节，改善当前国内因缺乏成熟智能分级方法而主要依赖人工操作，进而导致效率低下且标准不统一的问题，以保障雪茄烟叶产品质量。以福建龙岩主栽品种‘FX-01’为研究对象，采集9个常用收购等级的8637张图像数据，采用Swin、ViT、ResNet、Beit和ConvNext 5种主流深度学习模型，分别针对上、中、下部叶构建智能分级模型。结果表明，所有模型均满足日常响应速度需求，其中ConvNext和ViT模型在中部叶测试集上的表现最优，平均准确率达93.3%。这些结果验证了基于图像的深度学习技术在雪茄茄衣智能分级中的可行性，可为后续系统改进和移动端部署提供技术支持与理论依据，同时为雪茄生产的自动化和标准化提供参考。

关键词: 雪茄, 分级, 图像识别, 深度学习, 模型构建, 智能分级模型, ConvNext, 视觉Transformer(ViT)

Abstract:

This study aims to address the challenges in the cigar leaf grading process in China, where the lack of mature intelligent grading methods has led to a reliance on manual operations, resulting in inefficiency and inconsistent standards. The goal is to ensure the quality of cigar leaf products. The ‘FX-01’ variety, the main cultivar in Longyan, Fujian, was used as the research material, and a dataset of 8637 images covering nine commonly used acquisition grades was collected. Five state-of-the-art deep learning models (Swin, ViT, ResNet, Beit and ConvNext) were employed to develop intelligent grading models for upper, middle, and lower leaves, respectively. The results showed that all models met the requirements for daily response speed, with the ConvNext and ViT models achieving the best performance on the middle leaf test set, with an average accuracy of 93.3%. These findings demonstrate the feasibility of deep learning-based image technology in the intelligent grading of cigar wrapper leaves and provide technical support and theoretical guidance for further system improvement and mobile deployment, laying a foundation for the automation and standardization of cigar production.

Key words: cigar, grading, image recognition, deep learning, model construction, intelligent grading model, ConvNext, vision Transformer (ViT)