高光谱图像包含了大量的光谱信息和图像信息, 采用高光谱成像技术对牛肉品种进行识别。 获取可见-近红外(400～1 000 nm)光谱范围内的安格斯牛、 利木赞牛、 秦川牛、 西门塔尔牛、 荷斯坦奶牛五个品种共252个牛肉样本的高光谱图像。 在ENVI软件中对高光谱图像进行阈值分割并构建掩膜图像, 获取样本的感兴趣区域（ROI）, 并结合伪彩色图对牛肉样本的反射率指数进行可视化表达; 采用Kennard-Stone(KS)法对样本集进行划分以提高模型的预测性能; 对原始光谱采用卷积平滑(SG) 、 区域归一化(Area normalize)、 基线校正(Baseline)、 一阶导数（FD）、 标准正态变量变换(SNV)及多元散射校正（MSC）等6种方法进行预处理; 采用竞争性自适应重加权算法（CARS）提取特征波长。 然后利用颜色矩对不同牛肉样本的颜色特征进行提取; 对原始光谱图像进行主成分分析, 结合灰度共生矩阵（GLCM）算法, 提取主要纹理特征。 最后结合偏最小二乘判别（PLS-DA）算法建立牛肉样本基于特征波长、 颜色特征以及纹理特征的识别模型。 KS法将牛肉样本划分为校正集190个, 预测集62个; 将未经预处理的光谱数据与经过6种不用预处理的光谱数据进行建模分析, 结果发现经FD法处理后的光谱数据所建模型的识别率最高; 结合CARS法对经FD法预处理后的光谱数据进行特征波长提取, 共提取出22个波长; 利用颜色矩和GLCM算法分别提取出每个牛肉样本的9个颜色特征、 48个纹理特征。 将特征波长数据与颜色、 纹理特征信息进行融合建模, 结果表明, 基于特征光谱+纹理特征的模型识别效果最佳, 其校正集与预测集识别率分别为98.42%和93.55%, 均高于特征光谱数据模型识别率, 说明融合纹理特征后使样本分类信息的表达更加全面; 融合颜色特征后模型的校正集识别率均有所增加, 但预测集识别率稍逊, 颜色特征虽携带了部分有效信息, 但这些信息与牛肉样本的相关性不大。 因此, 寻找与牛肉样本相关性更大的颜色特征是提高模型识别率的重要途径之一。 该研究结果为牛肉品种的快速无损识别提供了一定的参考。

In this study, beef variety was identified by hyperspectral imaging technology which contains abundant spectral and spatial information in an object. Firstly, hyperspectral images of beef samples in the visible and near infrared (400-1000 nm) regions were acquired by the hyperspectral imaging system which contain 252 samples of five varieties of Angus, Limuzan, Qinchuan, Simmental, and Holstein cows. The binary mask image was successfully determined with a certain threshold from ENVI, and ROI (Region of Interest) of beefsample was determined by using the binary mask image. The visual distribution map of reflectance index in beef sample was plotted by pseudo-color map. Samples were dividedby using KS method, which is to improve the prediction performance of the model; The spectral pretreatment method wasutilized, such as SG, Area normalize, Bseline, FD, SNV, MSC and so on; Feature wavelengths were extracted by using competitive adaptive weighting algorithm (CARS). The color characteristics were represented by used color moment for different beef sample images; Principal component analysis was performed on the original hyperspectral image. The image textural information was described by extracting main texture features by the gray level co-occurrence matrix (GLCM) algorithm of the beef sample. Then spectral data from CARS, color feature and texture feature (from three principle component images) were utilized to develop different partial least squares discrimination (PLS-DA)models to identify beef samples respectively. The samples were divided into calibration set and prediction set by KS method, and calibration samples was 190, and prediction samples was 62; The spectral pretreatment was studied by the 7 methods. The results showed that the model effect of FD methods pretreatment was the best; A total of 22 characteristic wavelengths were extracted by the CARS method for spectral data using FD method; A total of 9 color features were extracted by color moments, and the GLCM algorithm was used to extract 48 texture features of each beef sample. Fusion models of spectral data, color feature, texture feature were established to identify beef samples. The results showed that, the model based on spectral data combined texture feature was the best with the correction set and prediction set recognition rate of 98.42% and 93.55%, respectively, which were higher than the recognition rate of feature spectral data. The texture feature made the expression of classification information more comprehensive. The recognition rate of the model correction set was increased by increasing color features, but the recognition rate of the prediction set was relatively poor. This meant the color features had some valid information, but the correlation between color features and the beef sample was not well, so the recognition rate of prediction set was reduced. Therefore, it is an important way to find color features that are more relevant to beef samples which could improve the recognition rate of models. This study provided valuable information for rapid destructive beef samples.