Research On Nondestructive Determination Of Tea Quality Based On Machine Vision And Spectroscopy Techniques

Tea is among the three most favorite beverages in the world, and China is the third-largest tea exporter in the world. There are nearly 100 million people who work in tea production and related activity in China, and tea output is main supply of farmers'income. In 2004, the total output value of tea and its related industries in China alone had reached more than 30 billion. So, it is very important to develop advanced techniques in tea production and processing which can greatly improve tea quality and increase farmer income. With the improvement of life quality, people's demand for tea with high quality is getting stronger and stronger. While the prevailing method for evaluating tea quality is through organoleptic way at home and abroad, the results are subjected to many factors, such as external environment, internal emotion and physiology, what's more, such evaluation is time-consuming, making it disagree with the demand of modern production and trade of tea. Concerning the above situation, this paper put forward a nondestructive and fast method for evaluation of tea quality based on machine vision and spectroscopy techniques, and conducted researches in three aspects:evaluation of external attributes of tea, determination of internal compositions, and diagnosis of tea plant information furthermore. The main contents and conclusions were showed as below:(1) Evaluation of external quality of tea by adoption of machine vision technique.A machine vision system was developed based on a MS3100 multi-spectral camera (Duncan co.). To obtain steady machine vision system, a contrast experiment was conducted with two different light sources respectively. And the results showed that the sun light could provide a more uniform illumination, without asymmetry and shadow in system based on single tungsten halogen light source. Elimination of shadow could greatly enhance the stability and precision of the model for tea and similar geometrical objects.This paper put forward a method for classification of tea variety and grade based on images of tea in different distribution statuses for the first time. Two classification models were developed according to images of tea granule without touch and those of stacked tea granule respectively. The study drew two conclusions:firstly, the classification rate of the model based on images of stacked tea was much higher than the other; secondly, the images of tea in stack could not only reflect geometrical attributes of tea granule, but also the macroscopical textural attributes, meanwhile, the feasibility to classify tea in stack status could greatly enhance the classification efficiency, and reflect the quality in whole, avoiding the localization in sample selection at classification based on image with untouched tea granule.Feature extraction is the pivotal technique for image recognition. According the attributes of tea, this paper extracted 18 morphologic features and 15 textural features, and the variety classification accuracy rate of the model developed accordingly reached 93.8%. Energy of 540nm monochromatic image, entropy of image filtered by standard deviation and correlation of 670nm monochromatic image were the three most important features contributing to the discrimination. Then grade classification model was developed with accuracy rate of 87.5%, with energy of 800 nm monochromatic image the most important feature.(2) Four methods for determination of tea compositions, such as polyphenols, amino acid and theine were developed and contrasted. Several spectral modes such as ASD Vis/NIR spectroscopy, NEXUS Fourier near infrared spectroscopy (FT-NIR), JASCO Fourier infrared reflectance spectroscopy (FT-IR-RS) and JASCO fourier infrared transmission spectroscopy (FT-IR-TS) were adopted for study the internal compositions of tea, Except NEXUS FT-NIR and JASCO FT-IR-TS were based on ground tea powder, the other two modes were all nondestructive method with intact tea.For determination of tea polyphenols, the precision of JASCO FT-IR-RS mode was the highest, with the prediction correlation value of the model reached 0.908, which was developed on characteristic wavelengths (1632cm-1-1768cm-1). The performance of ASD Vis/NIR was the second, prediction correlation of model with full wavelengths was 0.897. The model based on JASCO FT-IR-TS was developed with prediction correlation of 0.896, and the characteristic wavelengths were 924cm-1-1792cm-1. The model based on NEXUS FT-NIR mode obtained the lowest precision with prediction correlation of 0.865, and this model was developed based on narrow characteristic wavelengths of 6140cm-1-7140cm-1 and 5000cm-1-5960cm-1.For determination of amino acid, the prediction correlation of the model developed on JASCO FT-IR-TS mode reached 0.914, best performance of the four, and the characteristic wavelengths were detected as 870cm-1-1278cm-1. For the model of JASCO FT-IR-RS mode, prediction correlation was 0.899, and characteristic wavelengths were 450cm-1-1895cm-1. The correlation values of models based on ASD Vis/NIR mode and NEXUS FT-NIR mode reached 0.892 and 0.813 respectively.As for theine, wonderful results were obtained. Especially the model based on JASCO FT-IR-TS mode, prediction correlation reached 0.995, this model was developed based on three narrow characteristic wavelengths of 1784cm-1-2048cm-1,2191cm-1 and 2345cm-1. NEXUS FT-IR model got similar results with prediction correlation of 0.993, with characteristic wavelengths of 4302cm-1-5750cm-1. And the prediction correlation of ASD Vis/NIR model reached 0.96. However, JASCO FT-IR-RS model performed poorly, with prediction correlation of 0.442.The contrast of the above results showed the difference of the four methods, and also indicated that it is feasible to conduct nondestructive determination of polyphenols, amino acid and theine based on spectroscopy techniques.The research also explored methods for nondestructive determination of tea moisture based on Vis/NIR spectroscopy, built model for determination of dried tea, fresh leaves and tea in processing simultaneous. Wavelet transform and support vector machine were adopted to build models, and wonderful determination model was developed with r= 0.987. So it can be concluded that Vis/NIR spectroscopy can be used to measure moisture of tea in real-time, nondestructive and fast way. Compared with other researches based on ground tea powder, this determination method was completely nondestructive with intact tea.(3) Furthermore, for quick diagnosis of tea plant information, researches of the relationships among Vis/NIR spectroscopy and tea plant characteristics (including variety, nitrogen stress and relative chlorophyll content) were conducted.And the results showed that the three varieties tea plants in different physiological ages, different culture condition and different field management (fertilization, pesticide) could be discriminated based Vis/NIR spectroscopy in field.Through studying the relationship between Vis/NIR spectroscopy and nitrogen stress of 5 varieties tea plant, strong correlation was found, and the prediction correlation value of the model reached 0.885.For chrolophyll index,11 tea plant varieties were taken for establishing relationship between Vis/NIR spectroscopy and relative chlorophyll content of leaves, and field experiment and lab experiment all obtained wonderful performances with high prediction correlation (r=0.922 for field experiment, r=0.913 for lab experiment), and this was the first time to contrast the two ways and find the former a better performance.