欢迎访问《农学学报》，
农学学报

农学学报 ›› 2026, Vol. 16 ›› Issue (1): 57-64.doi: 10.11923/j.issn.2095-4050.cjas2024-0180

• 林学 园艺 园林 食用菌 • 上一篇    下一篇

基于Resnet与支持向量机融合识别网络的鲜烟叶部位识别

李昌根1(), 李珂2, 赵东方3, 张帅3, 孟祥宇2, 林勇4, 魏硕1, 王廷贤4()   

  1. 1 河南农业大学烟草学院，郑州 450046
    2 河南中烟工业有限责任公司，郑州 450016
    3 重庆市烟草公司酉阳分公司，重庆 409800
    4 福建省烟草公司南平市公司，福建南平 353000
  • 收稿日期:2024-09-10 修回日期:2025-02-26 出版日期:2026-01-20 发布日期:2026-01-15
  • 通讯作者:
    王廷贤，男，1986年出生，河南舞钢人，助理农艺师，学士，主要从事烟叶生产管理工作。通信地址：353000 南平市滨江中路389号，Tel：0599-8832034，E-mail：。
  • 作者简介:

    李昌根，男，2000年出生，河南洛阳人，在读研究生，研究方向为烟草种植和烟叶加工工艺。通信地址：450046 河南郑州市郑东新区龙子湖高校园区15号 河南农业大学烟草学院，Tel：0371-56552221，E-mail：。

  • 基金资助:
    福建省烟草公司南平市公司资助项目“南平烟区翠碧一号烘烤工艺数字化基础研究”(NYK2023-03-03); 中国烟草总公司科技重点研发项目“基于图像精准识别的烟叶智能烘烤关键技术研究与应用”(110202102007)

Position Identification of Fresh Tobacco Leaf Based on Resnet and Support Vector Machine Fusion Recognition Network

LI Changgen1(), LI Ke2, ZHAO Dongfang3, ZHANG Shuai3, MENG Xiangyu2, LIN Yong4, WEI Shuo1, WANG Tingxian4()   

  1. 1 College of Tobacco Science, Henan Agricultural University, Zhengzhou 450046
    2 China Tobacco Henan Industrial Co., Ltd., Zhengzhou 450016
    3 Youyang Branch of Chongqing Tobacco Company, Chongqing 409800
    4 Nanping Branch of Fujian Provincial Tobacco Company, Nanping, Fujian 353000
  • Received:2024-09-10 Revised:2025-02-26 Online:2026-01-20 Published:2026-01-15

RichHTML

1

PDF (PC)

10

摘要：

为实现鲜烟叶采收部位的数字化识别，满足智能烘烤中鲜烟叶部位快速无损识别的需求，构建了一种融合Resnet-50和支持向量机(SVM)的鲜烟叶部位识别网络模型(R-SVM)。在预训练的Resnet-50网络模型提取鲜烟叶图像不同卷积层（第1、10、22、40、49层）的特征基础上，结合不同的池化方法[平均池化(AVP)、全局平均池化(GAP)和空间金字塔池化(SPP)]和降维算法[主成分分析(PCA)和ReliefF]，分别训练支持向量机(SVM)并筛选出不同的鲜烟叶部位识别模型，再通过不同的模型融合策略（硬投票、软投票、Stacking方法）得到最终鲜烟叶部位识别模型。结果表明，不同池化方法对模型性能的影响各异，在低层卷积层中，经SPP池化后模型性能显著提高，模型准确率提高10%以上，而其对高层卷积层获取的特征训练得到的模型性能影响较小；PCA降维能够有效提升不同卷积层下识别模型的性能；不同卷积层中的第40层输出模型在测试集上的准确率最高(92.12%)，利用Stacking融合方法得到的模型性能最佳，在测试集上的准确率为96.83%。本研究建立的鲜烟叶部位融合识别模型能够实现烟叶部位的准确无损识别。

关键词: 残差网络, 支持向量机, 模型融合, 鲜烟叶部位, 分类

Abstract:

This study aims to achieve digital recognition of fresh tobacco leaf harvesting positions and meet the demand for rapid, non-destructive identification in intelligent curing. A hybrid network model (R-SVM) integrating Resnet-50 and support vector machine (SVM) is proposed for fresh tobacco leaf position recognition. Based on the features of different convolutional layers (layers 1, 10, 22, 40, 49) of fresh tobacco leaf images extracted by the pre-trained Resnet-50 network model, combined with different pooling methods [average pooling (AVP), global average pooling (GAP) and spatial pyramid pooling (SPP)] and dimensionality reduction algorithms [principal component analysis (PCA) and ReliefF], support vector machines (SVM) were trained respectively and different recognition models of fresh tobacco leaf harvesting positions were screened out, and then different model fusion strategies (hard voting, soft voting, Stacking method) were used to obtain the final recognition model of fresh tobacco leaf position. The results indicated that different pooling methods exhibited distinct impacts on model performance. In low-level convolution layers, SPP pooling significantly improved model accuracy by over 10%, while its effect was minimal on models trained using features from high-level convolution layers. PCA dimensionality reduction effectively enhanced recognition performance across all convolutional layers. The 40th layer output model in different convolution layers had the highest accuracy rate on the test set, which was 92.12%. The model obtained by the Stacking fusion method had the best performance, and the accuracy rate on the test set was 96.83%. The fusion recognition model for fresh tobacco leaf position established in this study can achieve accurate and non-destructive identification of tobacco leaf positions.

Key words: residual network, support vector machine, model fusion, fresh tobacco leaves position, classification