高光谱成像技术对牛肉水分含量及分布的快速检测

Rapid detection of beef moisture content and distribution by hyperspectral imaging

利用可见/近红外高光谱成像技术对牛肉水分含量及分布进行快速检测。采用可见/近红外高光谱成像系统(400-1 000 nm)采集150个黄牛肉样本的高光谱图像,利用ENVI软件提取样本感兴趣区域(ROI)并计算平均光谱值;对原始光谱数据进行预处理并利用连续投影算法(SPA)、竞争性自适应重加权算法(CARS)和无信息变量消除算法(UVE)进行特征波长提取,建立基于不同特征波长的偏最小二乘回归(PLSR)模型,进而优选牛肉水分含量预测的最优模型。通过蒙特卡罗交叉验证法剔除26个异常样本值;经卷积平滑(Smoothing-SG)法预处理后的原始光谱数据所建PLSR模型效果较好,其校正集决定系数(R^2c)与预测集决定系数(R^2_p)分别为0.817、0.850;利用CARS、SPA、UVE法分别优选出12、27、27个特征波长;对比基于全波段光谱与特征波段光谱所建PLSR牛肉水分预测模型的优劣,结果显示基于CARS-PLSR法建立的牛肉水分预测模型效果最好,其R^2_c、R^2_p值分别为0.814、0.750,校正集均方根误差(RMSEC)与预测集均方根误差(RMSEP)分别为0.477、0.555;最后,利用CARS-PLSR模型计算牛肉样本每个像素点的水分含量并利用伪彩色图对牛肉样本水分分布进行可视化分析,进而实现牛肉水分含量的快速检测及分布的可视化表达。该研究结果可为黄牛肉水分含量的快速检测提供理论支撑。

The visible/near-infrared hyperspectral imaging technique was used to rapidly detect the moisture content and distribution of beef.Hyperspectral images of 150 yellow cattle samples were collected using a visible near-infrared hyperspectral imaging system(400-1000 nm),and the region of interest(ROI) of the samples was extracted using ENVI4.8 software and the average spectral values were calculated;The raw spectral data is preprocessed and the feature wavelength extraction is performed by using continuous projection algorithm(SPA),competitive competitive reweighting(CARS) and non-information variable elimination algorithm(UVE),and the characteristic wavelengths are extracted.Partial Least Squares Regression(PLSR) model was preferably the best predictive model.A total of 26 abnormal samples were eliminated by Monte Carlo cross-validation;the PLSR model constructed by spectral data pre-processing by convolution smoothing(SG) method was relatively good,with R^2_c of 0.817 and R^2_p of 0.850;using CARS and SPA The UVE method preferably has 12,27,and 27 characteristic wavelengths;The PLSR model established by the full-band spectrum and the extracted characteristic band spectrum is compared.The results show that the CARS-PLSR model based on hyperspectral imaging technology has the best effect,and the R^2_c,R^2_p values are 0.814,0.750,respectively.RMSEC and RMSEP values of 0.477 and 0.389,respectively;The CARS-PLSR model was selected to calculate the moisture content of each pixel of the beef sample.The pseudo-color map was used to visualize the moisture content distribution of the beef sample,and the non-destructive detection of the moisture content of the beef and the visual expression of the distribution were realized.Detection provides theoretical support.