高光谱成像的豆腐形成过程中组分含量变化检测

Detection of Component Content Changes During Tofu Formation Based on Hyperspectral Imaging Technology

豆腐作为我国传统食品,其生产已有两千多年的生产历史,但目前市场上的豆腐还是以经验式小作坊生产为主,难以保证豆腐质量和品质的均一性。水分和蛋白质含量是影响豆腐品质的重要因素,然而,水分与蛋白质的传统检测方法过程繁琐,耗时、费力,无法及时指导生产。因此,探索豆腐制备过程中水分和蛋白质分布的快速、无损、定量描述方法,可为精确调控豆腐制备工艺提供科学依据。应用高光谱成像技术结合化学计量学方法检测豆腐形成过程中豆浆、热浆、凝胶、豆腐四种不同状态下水分与蛋白质含量变化并实现其含量分布可视化。采集每种状态下120个样品在432~963 nm波段范围内的高光谱图像,利用ENVI软件选取感兴趣区域并计算样品的平均光谱数据。采用卷积平滑(savitzky-golay, SG)结合多元散射校正(multiplicative scatter correction, MSC)对原始光谱进行预处理,消除光谱噪声影响。用预处理后的光谱数据建立偏最小二乘回归(PLSR)和主成分回归(PCR)定量模型,比较发现豆浆、热浆、凝胶、豆腐样品的PCR模型对与水分和蛋白质的预测结果均低于PLSR模型。选用PLSR模型作为最优模型,采用连续投影算法(SPA)筛选豆浆、热浆、凝胶、豆腐样品的特征波长,分别选取13, 9, 8和9个特征波长建立基于特征波长下的PLSR模型。结果表明:与全波段下的PLSR模型相比基于特征波长建立的SPA+PLSR模型的预测效果更好,对水分的预测模型RP达到0.84~0.96,蛋白质的预测模型达到0.92~0.97。基于预测效果更好的SPA+PLSR模型计算豆浆、热浆、凝胶、豆腐图像中每个像素点的水分与蛋白质含量,将样品中的水分与蛋白质分布用不同的颜色直观显示,实现水分与蛋白质在不同状态下的分布。验证了高光谱技术对豆腐形成中水分与蛋白质含量检测的可行性,解决传统检测方法的缺陷,为豆腐生产的工业化和智能化提供理论依据。

Tofu was a traditional food in China in two thousand years.However,it is still mainly produced by individual workshop,and the security and uniformity of quality is difficult to guarantee.Water and protein content are important factors which affecting the quality of tofu.The traditional detection method for water and protein is complicated,time consuming and laborious.The detection results are often later than the production process which unable to guide tofu production in time.Therefore,it is necessary to develop a new method for quantitatively describing homogeneity of water and protein distribution in Tofu.It is also the a scientific basis for accurately regulating the production process of tofu.In this study,hyperspectral imaging technique combined with chemometrics method were used to detect the changes of water and protein content and distribution under four different conditions:soybean milk,hot soybean milk,gel and tofu.The hyperspectral image of 120 samples in each state was collected in the wavelength range of 432 to 963 nm.Use ENVI software to select the region of interest and calculate the average spectral data of the sample.The original spectrum was pre-processed by convolution smoothing(Savitzky-Golay,SG)as well as multiplicative scatter correction(MSC)to eliminate the influence of spectral noise.A partial least squares regression(PLSR)and principal component regression(PCR)quantitative model were established for the pre-processed spectral data.The prediction results of water and protein by using PCR model are lower than the PLSR model,therefore,the PLSR model is selected as the optimal model.The characteristic wavelengths of soybean milk,hot soybean milk,gel and tofu samples were selected by continuous projection algorithm(SPA),13,90,8 and 9 characteristic wavelengths to establish the PLSR model based on the characteristic wavelengths.The results show that the SPA+PLSR model based on the characteristic wavelengths is better than the PLSR model under the full-band model.The prediction model R p for for water is 0.84~0.96,and the prediction model R p for protein is 0.92~0.97.Based on the SPA+PLSR model with better prediction effect,the water and protein contents of different states pixel in the image of soybean milk,hot soybean milk,gel and tofu,were calculated and the water and protein distributions in the sample were visualized with different colors.The feasibility of hyperspectral technology for detecting water and protein contents in tofu was verified,which could be used to solve the defects of traditional detection methods,as well as provided a theoretical basis for the industrialization and intelligence of tofu production.