基于PPF投影算法和高光谱技术的卷烟牌号识别模型  

Cigarette brand identification model based on PPF projection algorithm and hyperspectral technology

作　　者：朱亚昆 梅吉帆 郭文孟 苏子淇 李嘉康 徐大勇[1] 堵劲松[1] 李辉 ZHU Yakun;MEI Jifan;GUO Wenmeng;SU Ziqi;LI Jiakang;XU Dayong;DU Jinsong;LI Hui(Zhengzhou Tobacco Research Institute of CNTC,Zhengzhou 450001,China;Technology Center,China Tobacco Chongqing Industrial Co.,Ltd.,Chongqing 400000,China;Xiangcheng Branch of Xuchang Tobacco Company,Xuchang 461700,China)

机构地区：[1]中国烟草总公司郑州烟草研究院,河南郑州450001 [2]重庆中烟工业有限责任公司技术中心,重庆400000 [3]许昌市烟草公司襄城县分公司,河南许昌461700

出　　处：《轻工学报》2024年第4期118-126,共9页Journal of Light Industry

基　　金：中国烟草总公司重点研发项目(110202202011)。

摘　　要：针对卷烟成分复杂、牌号识别难度大的问题,提出了一种基于PPF投影算法和高光谱技术的卷烟牌号识别模型。首先,采用PPF投影算法对原始光谱及经多元散射校正(MSC)、一阶微分(1^(st)D)、二阶微分(2^(nd)D)处理后的数据进行光谱相似度表征;然后,通过连续投影算法(SPA)和遗传算法(GA)进行特征波长选取,以利于后续分类性能的提升;最后,分别采用支持向量机(SVM)、极限梯度提升(XGBoost)算法建立牌号识别模型,并进行性能评价。结果表明:在高光谱技术结合PPF投影算法中,2ndD为最佳波段处理方法,能够显著增强光谱可分性;基于2ndD-SPA-SVM建立的识别模型为最佳模型,训练集、测试集的总体分类精度(OA)分别为94.58%、92.50%;不同牌号卷烟配方烟丝的叶丝和膨胀丝占比差异越大,类别之间的相似度越小,识别准确率越高。该模型可为不同牌号的卷烟提供一种新的高效快速、准确无损的分类判别方法,为高光谱技术在卷烟品牌维护、叶组配方设计等中的应用提供参考。Addressing the complexity of cigarette ingredients and the challenges in brand identification,a cigarette brand identification model based on Projection Pursuit based on Fishers Criterion(PPF)and hyperspectral technology was proposed.Firstly,the PPF projection algorithm was employed to conduct spectral similarity analysis on the original spectra as well as those processed with Multiplicative Scatter Correction(MSC),first-order derivative(1stD),and second-order derivative(2ndD).Secondly,feature wavelength selection was carried out using Successive Projections Algorithm(SPA)and Genetic Algorithm(GA)to enhance the classification performance in subsequent steps.Finally,cigarette brand identification models were established using Support Vector Machine(SVM)and Extreme Gradient Boosting(XGBoost)algorithms,and the personality of the model was evaluated.The results indicated that,in the combination of hyperspectral technology and the PPF projection algorithm,the second-order derivative(2ndD)served as the optimal band processing method,significantly enhancing the spectral separability.The recognition model established based on 2ndD-SPA-SVM emerged as the best model,achieving an overall classification accuracy of 94.58%and 92.50%for the training and test sets respectively.Moreover,a greater difference in the proportion of leaf silk and expanded silk among different brand formulations led to a lower similarity value between categories and a higher recognition accuracy.This model can provide a new efficient,fast,accurate,and non-destructive classification and discrimination method for cigarettes of different grades.It offered theoretical support for the application of hyperspectral technology in cigarette brand maintenance,tobacco blend formulation design,and other scenarios.

关 键 词：PPF投影算法 高光谱技术 卷烟牌号识别 相似度表征 

分 类 号：TS452[农业科学—烟草工业] TP181[自动化与计算机技术—控制理论与控制工程]