高光谱技术融合纹理信息的羊肉总酚浓度快速检测

Fusion of Visible Near-Infrared (VNIR) Hyperspectral Imaging and Texture Feature for Prediction of Total Phenolics Content in Tan Mutton

为了研究可见-近红外(Vis-NIR)高光谱成像对滩羊肉中总酚浓度(TPC)快速检测的可行性,基于光谱信息融合图像纹理特征建立TPC含量的预测模型,实现滩羊肉中TPC含量的可视化表达。将样本集根据3∶1的比例划分成校正集和预测集,采用多元散射校正(MSC)、基线校准(Baseline)、去趋势(De-trending)、卷积平滑(S-G)、标准正态变量变换(SNV)、归一化(Normalize)等校正方法去除原始光谱中不良散射等干扰信息。通过竞争性自适应加权抽样(CARS)、引导软收缩(BOSS)、区间变量迭代空间收缩法(iVISSA)和变量组成集群分析-迭代保留信息变量(VCPA-IRIV)提取TPC浓度的代表性特征光谱。采用灰度共生矩阵(GLCM)算法依次提取肉样第1主成分图像中纹理特征。基于特征光谱及图谱融合信息建立滩羊肉中TPC含量的偏最小二乘回归(PLSR)与最小二乘支持向量机(LSSVM)预测模型并进行对比分析。结果表明,(1)利用De-trending+SNV预处理后的光谱数据建立的PLSR预测模型性能较好,其R_(C)^(2)=0.8749,R_(P)^(2)=0.7932;(2)采用CARS,BOSS,iVISSA和VCPA-IRIV分别提取出了23,35,57和43个特征波长,占全光谱的18.4%,28%,45.6%和16.8%;(3)采用BOSS法提取的关键性波长建立的LSSVM模型性能较好,其R_(C)^(2)=0.8513,R_(P)^(2)=0.7459,RMSEC=0.1168和RMSEP=0.1550;(4)与基于特征波长建立的预测模型相比,BOSS-ASM-ENT-CON-LSSVM模型为滩羊肉中TPC浓度的最佳图谱融合预测模型(R_(C)^(2)=0.8500,R_(P)^(2)=0.7709,RMSEC=0.1160,RMSEP=0.1447);(5)利用BOSS-PLSR简化模型将TPC浓度反演到样本的高光谱图像上,通过色彩直观化形式展现出来,实现TPC含量的可视化表达。

The visible near-infrared(Vis-NIR)hyperspectral imaging technology was used to rapidly detect Tan mutton’s total phenol concentration(TPC)content.The prediction mode and visualization of TPC content in Tan mutton were built and realized based on spectral information in combination with texture features.Firstly,the calibration set and prediction set were divided by 3∶1,and then multiplicative scatter correction(MSC),Baseline,De-trending,savitzky-golay(S-G),and Standard normal variate transformation(SNV),and Normalize were used for model optimization.Secondly,feature bands were obtained by competitive adaptive reweighted sampling(CARS),bootstrapping soft shrinkage(BOSS),interval variable iterative space shrinkage approach(iVISSA)and variable combination population analysis coupled with iteratively retained informative variables(VCPA-IRIV),respectively.Then textural feature variables for the first principal component image were extracted by gray-level co-occurrence matrix(GLCM),respectively.Finally,partial least squares regression(PLSR)and least-squares support vector machines(LSSVM)models were built and optimized to predict TPC content.The results showed that:(1)The PLSR model yielded promising results after De-trending-SNV preprocessing,and R_(P)^(2) and R^(2)_(C) were 0.7932 and 0.8749;(2)The 23,35,57 and 43 characteristic bands based on the original spectral were extracted by CARS,BOSS,iVISSA and VCPA-IRIV methods,respectively,accounting for 18.4%,28%,45.6%and 16.8%of the total bands;(3)The simplified BOSS-LSSVM model yielded good results in assessing TPC content(R^(2)_(C) vs.R_(P)^(2)=0.8513 vs.0.7459,RMSEC vs.RMSEP=0.1168 vs.0.1550);(4)Compared with predictive models based on characteristic wavelengths,the simply model BOSS-ASM-ENT-CON-LSSVM despited good results(R^(2)_(C)=0.8500,R_(P)^(2)=0.7709,RMSEC=0.1160,RMSEP=0.1447);(5)The simplified BOSS-PLSR model was displayed on the sample image in the form of pseudo-color to realize the visualization expression of TPC content.