基于计算机视觉技术的茶叶品质随机森林感官评价方法研究
|
摘要
为弥补茶叶品质感官审评存在的缺陷,利用计算机视觉技术对茶叶品质进行快速无损评价研究。以碧螺春绿茶为对象,依据专家感官审评结果,将茶样分成4个等级;采用中值滤波及拉普拉斯算子对茶样图像进行预处理,并提取预处理后的茶样图像的颜色特征和纹理特征以表征茶叶图像的外形特征,利用随机森林算法对茶叶外形特征属性进行重要性排序;筛选出重要性较大的特征及随机森林算法中最优的决策树棵数建立感官评价模型,并与建立的支持向量机(SVM)模型性能相比较。结果表明:色调均值、色调标准差、绿体均值、平均灰度级、饱和度均值、红体均值、饱和度标准差、亮度均值、一致性等9个特征属性的重要性较大,且与感官审评特征描述结果相一致;当采用优选出的9个重要性较大的特征及决策数棵数为500时,建立的模型性能最优,模型总体判别率为95.75%,Kappa系数为0.933,OOB误差为5%,较SVM模型分别提高了3.5%,0.066,优选的9个重要性较大的图像特征与感官审评特征描述相一致。研究表明:利用随机森林方法筛选出对茶叶外形特征属性贡献最大的少数几个特征建立模型,模型性能就能达到很好的识别效果,模型得到简化,同时模型精度和稳定性都高于其他方法。
In order to make up some flaws of sensory evaluation of tea quality,computer vision technology as a fast and nondestructive method was used to evaluate tea quality in this paper.Biluochun green tea samples were studied,and the tea samples were divided into four grades based on the evaluation results by experts for tea samples.The median filter and Laplace operator were used to preprocess the images of tea samples,and tea appearance features such as color features and texture features were extracted from the preprocessed images.Random forest(RF)method was used to analyze the significance of tea appearance features.The most important features and the optimal amounts of tree pruning of decision tree were investigated to develop the sensory evaluation model of tea quality.And the performance of RF model was compared to that of SVM model.The results show that nine more important features such as hue mean,hue standard deviation,greed channel mean,mean grey,saturation mean,red channel mean,saturation standard deviation,vision mean and uniformity were selected,and the result was consistent with the sensory evaluation profile;the optimal model was obtained when 9most importance feature variables were selected and tree pruning of decision tree were 500.The overall recognition rate of the model was 95.75%,Kappa coefficient was 0.933,and OOB error was 5%.Compared with SVM model,the average recognition rate and Kappa coefficient of RF algorithm were improved 3.5% and 0.066,respectively.The 9significant image features selected were consistent with the feature description of tea sensory evaluation terms by NY/T 863—2004.The study indicated that the developed model with a few significant appearance feature variables selected by RF method has high accuracy,and the model is simplified without lowering accuracy.The precision and stability of RF model are superior to that of SVM.
In order to make up some flaws of sensory evaluation of tea quality,computer vision technology as a fast and nondestructive method was used to evaluate tea quality in this paper.Biluochun green tea samples were studied,and the tea samples were divided into four grades based on the evaluation results by experts for tea samples.The median filter and Laplace operator were used to preprocess the images of tea samples,and tea appearance features such as color features and texture features were extracted from the preprocessed images.Random forest(RF)method was used to analyze the significance of tea appearance features.The most important features and the optimal amounts of tree pruning of decision tree were investigated to develop the sensory evaluation model of tea quality.And the performance of RF model was compared to that of SVM model.The results show that nine more important features such as hue mean,hue standard deviation,greed channel mean,mean grey,saturation mean,red channel mean,saturation standard deviation,vision mean and uniformity were selected,and the result was consistent with the sensory evaluation profile;the optimal model was obtained when 9most importance feature variables were selected and tree pruning of decision tree were 500.The overall recognition rate of the model was 95.75%,Kappa coefficient was 0.933,and OOB error was 5%.Compared with SVM model,the average recognition rate and Kappa coefficient of RF algorithm were improved 3.5% and 0.066,respectively.The 9significant image features selected were consistent with the feature description of tea sensory evaluation terms by NY/T 863—2004.The study indicated that the developed model with a few significant appearance feature variables selected by RF method has high accuracy,and the model is simplified without lowering accuracy.The precision and stability of RF model are superior to that of SVM.
引文
[1] Wang S,Phillips P,Liu A,et al.Fundamenta Informaticae,2017,151(1-4):325.
[2] Borah S,Hines E L,Bhuyan M.Journal of Food Engineering,2007,79(2):629.
[3] Laddi A,Sharma S,Kumar A,et al.Journal of Food Engineering,2013,115(2):226.
[4] Dong C,Zhu H,Zhao J,et al.Journal of Zhejiang University-SCIENCE B(Biomedicine&Biotechnology),2017,18(6):544.
[5] ZHANG Zhi-feng,LI Shi-hai,TANG Yi-ming,et al(张志峰,李世海,汤一明,等).Journal of Light Industry(轻工学报),2016,31(6):103.
[6] YU Hong,WU Rui-mei,AI Shi-rong,et al(余洪,吴瑞梅,艾施荣,等).Laser Journal(激光杂志),2017,38(1):51.
[7] XU Guo-xiang,YANG Zhen-jian(徐国祥,杨振建).The Journal of Quantitative&Technical Economics(数量经济技术经济研究),2011,28(2):135.
[8] Verma A,Dong X.Journal of Biomedical Science and Engineering,2016,9(5):259.
[9] LIU Ya-dong,CUI Ri-xian(刘亚东,崔日鲜).Spectroscopy and Spectral Analysis(光谱学与光谱分析),2015,35(12):3480.
[10] Fatemeh N,Tomasz S,Dorota K,et al.International Journal of Speech Technology,2017,20(2):239.
[11] Li Z,Wang R,Ye Z,et al.Fisheries Science,2017,83(1):1.
[12] WU Rui-mei,WU Yan-hong,AI Shi-rong,et al(吴瑞梅,吴彦红,艾施荣,等).Acta Agriculturae Universitis Jiangxiensis(江西农业大学学报),2013,35(2):413.
[13] Wang L,Chen H,Yan H,et al.International Wound Journal,2015,12(5):590.
[14] LI Xin-hai(李欣海).Chinese Journal of Applied Entomology(应用昆虫学报),2013,50(4):1190.
[15] Zliobaite I,Bifet A,Read J,et al.Machine Learning,2015,98(3):455.
[16] XU Shu-liang,WANG Jun-hong(徐树良,王俊红).Computer Science(计算机科学),2016,43(12):173.
[2] Borah S,Hines E L,Bhuyan M.Journal of Food Engineering,2007,79(2):629.
[3] Laddi A,Sharma S,Kumar A,et al.Journal of Food Engineering,2013,115(2):226.
[4] Dong C,Zhu H,Zhao J,et al.Journal of Zhejiang University-SCIENCE B(Biomedicine&Biotechnology),2017,18(6):544.
[5] ZHANG Zhi-feng,LI Shi-hai,TANG Yi-ming,et al(张志峰,李世海,汤一明,等).Journal of Light Industry(轻工学报),2016,31(6):103.
[6] YU Hong,WU Rui-mei,AI Shi-rong,et al(余洪,吴瑞梅,艾施荣,等).Laser Journal(激光杂志),2017,38(1):51.
[7] XU Guo-xiang,YANG Zhen-jian(徐国祥,杨振建).The Journal of Quantitative&Technical Economics(数量经济技术经济研究),2011,28(2):135.
[8] Verma A,Dong X.Journal of Biomedical Science and Engineering,2016,9(5):259.
[9] LIU Ya-dong,CUI Ri-xian(刘亚东,崔日鲜).Spectroscopy and Spectral Analysis(光谱学与光谱分析),2015,35(12):3480.
[10] Fatemeh N,Tomasz S,Dorota K,et al.International Journal of Speech Technology,2017,20(2):239.
[11] Li Z,Wang R,Ye Z,et al.Fisheries Science,2017,83(1):1.
[12] WU Rui-mei,WU Yan-hong,AI Shi-rong,et al(吴瑞梅,吴彦红,艾施荣,等).Acta Agriculturae Universitis Jiangxiensis(江西农业大学学报),2013,35(2):413.
[13] Wang L,Chen H,Yan H,et al.International Wound Journal,2015,12(5):590.
[14] LI Xin-hai(李欣海).Chinese Journal of Applied Entomology(应用昆虫学报),2013,50(4):1190.
[15] Zliobaite I,Bifet A,Read J,et al.Machine Learning,2015,98(3):455.
[16] XU Shu-liang,WANG Jun-hong(徐树良,王俊红).Computer Science(计算机科学),2016,43(12):173.