结合高光谱图像的光谱和纹理信息预测羊肉可溶性蛋白和GSH含量

可溶性蛋白和谷胱甘肽（GSH）是羊肉重要的生理生化指标，是衡量机体抗氧化能力大小的重要因素，传统检测方法程序复杂，检测费时。为此应用可见-近红外（400～1 000 nm）高光谱成像技术实现可羊肉可溶性蛋白和还原性谷胱甘肽（GSH）含量无损、快速检测。首先，对采集的180个羊肉样本的原始光谱信息采用4种方法进行预处理，再运用竞争自适应加权算法（CARS）、区间变量迭代空间收缩算法-迭代和保留信息变量法（iVISSA-IRIV）进行特征波段的提取。同时使用灰度共生矩阵法（GLCM）提取贡献率最高的主成分图像的纹理信息。最后将优选出的预处理方法和特征波长信息作为光谱信息和光谱-纹理融合信息分别结合多元线性回归（MLR）、最小二乘支持向量机（LS-SVM）模型建立羊肉可溶性蛋白和谷胱甘肽含量的预测模型。结果显示未经预处理的原始光谱建立的羊肉可溶性蛋白含量PLSR模型效果最佳，其R_c和R_p分别为0.875 7和0.854 7；采用SNV法预处理后光谱建立的羊肉GSH含量PLSR模型效果最佳，其R_c和R_p分别为0.804 8和0.826 5。利用iVISSA-IRIV共筛选出31个特征波长，建立的羊肉可溶性蛋白LS-SVM模型的R_c和R_p最优，分别为0.914 6和0.881 8；同时利用iVISSA-IRIV筛选出29个特征波长，建立的羊肉GSH-MLR模型的R_c和R_p最优，分别为0.844 6和0.870 5。最终经光谱特征信息和图谱信息融合模型对比发现，建立iVISSA-IRIV-LS-SVM模型对羊肉可溶性蛋白预测效果最佳，其R_c和R_p分别为0.914 6和0.881 8；利用SNV-iVISSA-IRIV法提取的光谱特征信息与纹理信息融合建立的MLR模型为预测羊肉GSH含量的最优模型，其R_c和R_p分别为0.849 5和0.890 4。利用最优iVISSA-IRIV-LS-SVM和iVISSA-IRIV-MLR模型和成像处理方法，结合伪色彩图像直观的表示羊肉样本的可溶性蛋白和GSH含量的空间分布情况。研究结果表明利用高光谱图像的光谱和纹理信息能够用来预测羊肉可溶性蛋白和GSH含量。

Combination of Spectral and Textural Informations of Hyperspectral Imaging for Predictions of Soluble Protein and GSH Contentsin Mutton

Soluble protein and glutathione(GSH)are important physiological and biochemical indicators of mutton,which are also significant in measuring the body’s antioxidant capacity.However,the traditional detection methods are complicated and time-consuming.This report applied visible-near-infrared(400~1000 nm)hyperspectral imaging technology to achieve nondestructive and rapid detection of soluble protein and glutathione(GSH)content in mutton.Four methods are used to preprocess the original spectral information of the collected 180 mutton samples,and then use the competitive adaptive weighting algorithm(CARS),the wavelength space iterative shrinkage algorithm-iteration and retained information variable method(iVISSA-IRIV)method for characteristics band extraction.At the same time,the gray level co-occurrence matrix method(GLCM)is used to extract the texture information of the principal component image with the highest contribution rate.Finally,the optimized preprocessing method and the characteristic wavelength information are combined with multiple linear regression(MLR),and least squares support vector machine(LS-SVM)prediction models respectively,as spectral information and spectral-texture fusion information,to establish the prediction models of soluble protein and glutathione content of mutton.The results illustrate that the PLSR model of mutton soluble protein content established by the original spectrum without pretreatment has the best effect,and its R_c and R_p are 0.8757 and 0.8547,respectively;the PLSR model of mutton GSH content established by the spectra after pretreatment with SNV method work best,with R_c and R_p of 0.8048 and 0.8265,respectively.A total of 31 characteristic wavelengths were screened using iVISSA-IRIV,and the R_c and R_p of the established mutton soluble protein LS-SVM model were 0.9146 and 0.8818 respectively,which are the best.The meanwhile,29 characteristic wavelengths were screened using iVISSA-IRIV,and the R_c and R_p of the established mutton GSH-MLR model were optimal,0.8446 and 0.8705,respectively.The comparison of the spectral feature information and the fusion model of the map information revealed that the establishment of the iVISSA-IRIV-LS-SVM model was the best for the prediction of soluble protein in mutton,with R_c and R_p of 0.9146 and 0.8818,respectively.The MLR model established by fusion of the spectral feature information extracted by SNV-iVISSA-IRIV method with the texture information is the optimal model for predicting the GSH content of mutton,and its R_c and R_p are 0.8495 and 0.8904,respectively.The optimal iVISSA-IRIV-LS-SVM and iVISSA-IRIV-MLR models and imaging processing methods visually represented the spatial distribution of soluble protein and GSH contents of mutton samples in combination with pseudo-color images.The current study demonstrated that the spectral and textural information from hyperspectral images could predict soluble protein and GSH content of mutton.