基于ResNet的智能烟叶分级系统  

Intelligent Tobacco Leaf Grading System Using ResNet

作　　者：胡建欣 卢敏瑞 钟永健[3] 王辉[3] 俞贝楠 曹思源 刘英 沈会良[1] HU Jianxin;LU Minrui;ZHONG Yongjian;WANG Hui;YU Beinan;CAO Siyuan;LIU Ying;SHEN Huiliang(College of Information Science and Electric Engineering,Zhejiang University,Hangzhou Zhejiang 310027,China;Fujian Wuyi Tobacco Leaf Co.,Ltd.,Shaowu Fujian 354000,China;Technology Center,China Tobacco Zhejiang Industrial Co.Ltd.,Hangzhou Zhejiang 310008,China)

机构地区：[1]浙江大学信息与电子工程学院,浙江杭州310027 [2]福建武夷烟叶有限公司,福建邵武354000 [3]浙江中烟工业有限责任公司技术中心,浙江杭州310008

出　　处：《传感技术学报》2025年第1期96-103,共8页Chinese Journal of Sensors and Actuators

基　　金：国家自然科学基金项目(61871346)。

摘　　要：在卷烟生产中,为确保品质稳定,需要对烤烟分级。不同等级烟叶外观差异小,分级难度大。为充分利用烟叶有效信息,提高自动化分级准确率,设计了基于ResNet18的多源图像智能烟叶分级系统。该系统同时采集反射及透射图像,反射图像可提供颜色及纹理信息,透射图像可提供厚度及脉络形状信息。采用Phase Correlation配准和拉普拉斯金字塔融合局部图像,得到烟叶的完整透射图像,采用IC-LK配准反射及透射图像,提升模型特征提取能力。以ResNet18为基础,设计了多源图像特征融合分级模型,通过引入透射图像,分级准确率可从88.0%提升至90.22%,表明了所设计系统及烟叶分级方法的有效性。In cigarette production,in order to ensure stable quality,it is necessary to grade the flue-cured tobacco.However,subtle differ-ences in the appearance of different grades of tobacco make it hard to achieve accurate grading results.To make full use of the effective in-formation of tobacco leaves and improve the accuracy of automatic grading,a multi-source image intelligent tobacco leaf grading system based on ResNet18 is designed.This system simultaneously acquires the reflection and transmission images of a tobacco leaf,the reflection image provides color and texture information while the transmission image provides the information of thickness and vein shape.Local ima-ges are aligned and fused by using phase correlation and Laplace pyramid,respectively,to obtain the complete transmission images.IC-LK is adopted to align the reflection and transmission images to improve the feature extraction ability.Further,a multi-source image feature fu-sion grading model is designed based on ResNet18.By introducing transmission images,the proposed model can improve the grading accu-racy from 88.0%to 90.22%,verifying the effectiveness of the proposed system and tobacco grading method.

关 键 词：烟叶分级 多源图像 多模态特征提取 系统设计 ResNet 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]