为探究高光谱成像(400~1 000 nm)对羊肉中饱和脂肪酸(SFA)含量检测的可行性, 提出一种基于特征光谱信息和图像纹理特征融合的SFA含量预测模型, 实现对羊肉中SFA含量的快速检测及分布可视化。 利用分段阈值法构建掩膜图像, 获取羊肉样本感兴趣区域(ROI), 结合SPXY法对样本集进行划分并对相关光谱信息进行预处理, 分别采用连续投影算法(SPA)、 变量组合集群分析法(VCPA)和 β权重系数法提取特征光谱; 通过获取羊肉样本主成分图像, 结合灰度共生矩阵(GLCM)算法提取图像纹理信息; 分别对特征光谱、 图像信息及图谱融合信息建立的偏最小二乘回归(PLSR)与最小二乘支持向量机(LS-SVM)预测模型进行对比分析。 利用5种不同对原始光谱数据进行预处理, 经SNV法预处理后的光谱其校正集与预测集相关系数分别为0.921和0.875, 较原始光谱分别增加了0.001和0.04, 均方根误差模型分别为0.244和0.268, 较原始光谱模型分别减少了0.003和0.06; 对SNV法预处理后的光谱数据进行特征波长提取, SPA法、 VCPA法及 β权重系数法分别提取出12, 10和9个特征波长; 获取羊肉样本的前5个主成分图像, 选择所含信息量最多的第一主成分图像进行纹理特征提取, 依次提取0, 45°, 90°和135°方向下的能量、 熵、 同质性和相关性共4个主要纹理特征。 利用SPA法提取的特征波长建立的PLSR与LS-SVM模型性能较好, PLSR模型校正集与预测集相关系数分别为0.884 9和0.880 7, 均方根误差分别为0.300 1和0.260 6; LS-SVM模型校正集与预测集相关系数分别为0.898 7和0.892 6, 均方根误差分别为0.276 7和0.247 6; 图谱信息融合模型中, PLSR模型校正集与预测集相关系数分别为0.907 1和0.907 8, 较特征光谱模型分别增加了0.02和0.03, 均方根误差分别为0.326 9和0.299 2, 较特征光谱模型分别增加了0.03和0.04; LS-SVM模型校正集与预测集相关系数分别为0.920 6和0.894 6, 较特征光谱模型分别增加了0.02和0.002, 均方根误差分别为0.251 9和0.245 8, 较特征光谱模型分别减少了0.02和0.002。 光谱预处理中经SNV法处理后的光谱所建模型性能优于其他预处理方法; 采用SPA法提取的12个特征波长简化了光谱模型, 提高了模型性能, 特征光谱建模的最优方法为SPA-LS-SVM; 图谱信息融合模型较特征光谱模型, 模型相关系数增加较少, 表明图像纹理信息虽携带了部分有效信息, 但这些信息与羊肉中SFA含量之间的相关性有待进一步研究。 基于图谱信息融合模型的预测性能最优, 其次为光谱信息模型。 择优选取SPA-PLSR模型计算羊肉样本中每个像素点的SFA含量, 利用伪彩色图直观表示了羊肉样本中SFA的含量分布。 实现对羊肉样本SFA含量的无损检测及分布可视化表达。

In order to explore the feasibility of detection of saturated fatty acids (SFA) in muttons by using hyperspectral imaging techniques (400-1000 nm), this paper proposed a prediction model based on the fusion of characteristic spectral information and image texture features, realizing the rapid detection and distribution visualization of SFA content in mutton. Firstly, the binary mask image was successfully determined by the segmentation of a certain threshold, and Region of Interest (ROI) in the sample of mutton was determined by binary mask image. SPXY methods were used for dividing the sample set, preprocessing of correlation spectral information. And continuous projection algorithm SPA,VCPA and β weight were used to select wavelength of the spectrum. The image textural information was described by taking the principal component image and the gray level co-occurrence matrix (GLCM) algorithm of the mutton samples. The partial least squares regression (PLSR) and the least squares support vector machine (LS-SVM) prediction model built based on the characteristic wavelength, textural information, textural combined with characteristic wavelength were compared and analyzed, respectively. Preprocessing of original spectral data using five methods without pretreatment. The correlation coefficients of calibration set and prediction set were 0.921 and 0.875, respectively. Compared with the original spectrum, the correlation coefficients of calibration set and prediction set were increased by 0.001 and 0.04, and the root mean square errors were 0.244 and 0.268, respectively. Compared with the original spectrum, the correlation coefficients of calibration set and prediction set were reduced by 0.003 and 0.06 respectively. This paper extracted characteristic wavelengths of the spectral from the pre-processed date using SNV, SPA, VCPA and β coefficient methods extracted 12, 10 and 9 characteristic wavelengths, respectively. Five principal component images were selected based on PCA, and four textural feature variables (energy, entropy, homogeneity and correlation) were extracted by the first principal component image, with which the most information in the 0, 45°, 90°, and 135° directions, respectively. The performance of PLSR and LS-SVM models based on characteristic wavelengths extracted by SPA method was better. The correlation coefficients of PLSR model correction set and prediction set were 0.8849 and 0.8807, and the root mean square errors were 0.300 1 and 0.260 6, respectively. The correlation coefficients of LS-SVM model correction set and prediction set were 0.898 7 and 0.892 6, and the root mean square errors were 0.276 7 and 0.247 6, respectively. In the atlas information fusion model, the correlation coefficients of correction set and prediction set of PLSR model were 0.907 1 and 0.907 8 respectively, which were 0.02 and 0.03 higher than that of characteristic spectral model, and the root mean square errors were 0.3269 and 0.2992, respectively, which were 0.03 and 0.04 higher than that of characteristic spectral model; The correlation coefficients of LS-SVM model calibration set and prediction set were 0.920 6 and 0.894 6, respectively, which were 0.02 and 0.002 higher than that of characteristic spectral model, and the root mean square errors were 0.251 9 and 0.245 8, respectively, which were 0.02 and 0.002 less than that of characteristic spectral model. Compared with other pretreatment methods, the performance of the model constructed by the SNV was better than others; The 12 characteristic wavelengths were extracted by SPA method to simplify the spectral dimension and improve the performance of the model. The optimal method of characteristic spectral modeling was SPA-LS-SVM. Compared with the characteristic spectral model, the correlation coefficient of the model increased less, which indicated that the image texture information carried less effective information, and the correlation between these information and saturated fatty acid content in Mutton needed to be further studied. The prediction performance based on the textural combined with characteristic wavelength information fusion model was the best, and the texture information model was the worst. Thus, the SFA content of could be calculated by SPA-PLSR model, and the visual distribution map of SFA content in mutton samples was plotted by using pseudo-color drawing.