基于不同预处理高光谱信息的鸡肉滴水损失率快速预测研究

本文旨在通过挖掘不同预处理高光谱(900~1700 nm)信息构建鸡肉滴水损失率的快速预测模型。首先采集每个鸡肉样本高光谱图像并提取图像感兴趣区域内的平均光谱信息,经基线校正(BC)、标准正态变量校正(SNV)、多元散射校正(MSC)、高斯滤波平滑(GFS)、归一化校正(NC)等五种光谱不同预处理,利用偏最小二乘回归(Partial Least Squares Regression,PLSR)算法构建光谱信息与鸡肉滴水损失率之间的定量关系。然后分别基于回归系数法(Regression Coefficient,RC)、连续投影算法(Successive Projections Algorithm,SPA)和逐步回归算法(Stepwise)筛选出对模型精度影响较大的最优波长优化全波段PLS模型。结果显示,基于BC光谱的全波段PLSR模型(BC-PLSR)预测鸡肉滴水损失率效果更好(r_P=0.95,RMSEP=0.29%,RPD=3.07,ΔE=0.0024%)。利用Stepwise法从BC光谱中选取的14个最优波长(900.6、903.8、905.5、907.1、917.0、997.7、1162.2、1272.4、1354.8、1369.6、1410.8、1425.6、1584.1和1695.1 nm)建立的SW-BC-PLSR模型(r_P=0.97,RMSEP=0.24%,RPD=3.82,ΔE=0.0012%)和多元线性回归(Multiple Linear Regression,MLR)模型SW-BC-MLR(r_P=0.97、RMSEP=0.22%、RPD=4.19,ΔE=0.0036%)预测鸡肉滴水损失率效果均良好。本试验表明,基于近红外高光谱信息可潜在实现鸡肉滴水损失率的快速预测。

Fast Prediction of Drip Loss Rate in Chicken Meat Based on Different Pretreated Hyperspectral Information

This paper was conducted to construct a fast model for predicting drip loss rate of chicken by mining different preprocessed hyperspectral information(900~1700 nm). First,the hyperspectral images of each chicken sample were collected and the spectral information within the region of interest of the images were averaged and extracted. The mean spectral data were preprocessed by baseline correction(BC),standard normal variable correction(SNV),multiplicative scatter correction(MSC) and Gaussian filter smoothing(GFS) and normalization correction(NC),respectively. Partial least squares regression(PLSR) were used to explore the quantitative relationship between the spectral information and the drip loss rate of chicken samples. Then the regression coefficient method(RC),successive projections algorithm(SPA) and stepwise regression(Stepwise) were applied to select optimal wavelengths carrying most information for the full band PLSR models. The results showed that the BC-PLSR model built with full BC spectra had better predictive performance(r_P=0.95,RMSEP=0.29%,RPD=3.07,ΔE=0.0024%). Both the SW-BC-PLSR model(r_P=0.97,RMSEP=0.24%,RPD=3.82,ΔE=0.0012%) and SW-BC-MLR model(r_P=0.97,RMSEP=0.22%,RPD=4.19,ΔE=0.0036%) built with 14 optimal wavelengths(900.6,903.8,905.5,907.1,917.0,997.7,1162.2,1272.4,1354.8,1369.6,1410.8,1425.6,1584.1 and 1695.1 nm) selected by stepwise from full BC spectra had similar good accuracy for drip loss prediction. This experiment showed that it could be potentially realized for the rapid prediction of drip loss rate in chicken based on near-infrared hyperspectral data.