Fusion of Near-Infrared Hyperspectral Imaging (NIR-HSI) and Texture Feature for Discrimination of Lingwu Long Jujube With Different Bruise Grades
JING Yi-xuan1, WU Di2, LIU Gui-shan2*, HE Jian-guo2*, YANG Shi-hu2, MA Ping2, SUN Yuan-yuan2
1. School of Physics and Electronic-Electrical Engineering, Ningxia University, Yinchuan 750021, China
2. School of Food and Wine, Ningxia University, Yinchuan 750021, China
摘要： 利用近红外高光谱成像技术（NIR-HSI），将采集的光谱信息融合图像纹理信息建立分类模型，实现灵武长枣瘀伤等级的判别。采用瘀伤装置获得瘀伤等级为Ⅰ、Ⅱ、Ⅲ、Ⅳ、Ⅴ级的200个长枣样品，并按3∶1的比例划分校正集和预测集。采集不同瘀伤等级长枣的近红外高光谱图像，使用ENVI软件提取感兴趣区域（region of interest，ROI）并计算平均光谱值。为消除无用信息的干扰，采用正交信号修正（OSC）、基线校准（Baseline）、多元散射校正（MSC）、移动平均（MA）、卷积平滑（S-G）和去趋势（De-trending）对原始光谱进行预处理并建立偏最小二乘判别分析（PLS-DA）模型；基于最优预处理方法所得的全波段数据采用变量组合集群分析法（VCPA）、无信息消除变量算法（UVE）、竞争性自适应加权抽样算法（CARS）、区间变量迭代空间收缩法（iVISSA）和连续投影算法（SPA）提取特征波长后建立PLS-DA模型；将高光谱图像进行掩膜，利用主成分分析（PCA）获取贡献率最高的主成分图像，并在该图像上采用灰度共生矩阵（GLCM）提取纹理参数，包括能量（ASM）、熵（ENT）、对比度（CON）和相关性（COR），建立图谱融合的PLS-DA模型。结果表明，原始光谱数据建立的PLS-DA模型，校正集和验证集准确率分别为89%和86%；原始光谱经De-trending预处理后的PLS-DA模型效果最优，校正集和预测集准确率均为90%，较原始光谱模型分别提高了1%和4%；基于SPA选择特征波长后建立的PLS-DA模型的校正集和预测集准确率均为90%；De-trending-SPA-COR-PLS-DA图谱融合模型效果最优，模型校正集和预测集准确率均为92%。因此，利用近红外高光谱成像技术融合纹理信息可实现不同瘀伤等级灵武长枣的快速无损判别。
关键词：灵武长枣；高光谱；图谱信息融合；瘀伤等级判别
Abstract：Near-infrared hyperspectral imaging technology (NIR-HSI) was used to collect spectral image texture information to realize the discrimination of different bruise grades of Lingwu jujubes. 200 long jujube samples with bruising gradesⅠ, Ⅱ, Ⅲ, Ⅳ and Ⅴ were obtained by the bruising device, and the calibration set and prediction set were divided according to the ratio of 3∶1. Hyperspectral images of jujubes with different bruising grades were collected by NIR-HSI, the region of interest (ROI) was extracted using ENVI software and the average spectral value was calculated. Orthogonal signal correction (OSC), baseline, multiplicative scatter correction (MSC), moving average (MA), savitzky-golay (S-G) and de-trending were used to preprocess the original spectra and a partial least squares-discriminate analysis (PLS-DA) model was established; Variable combination population analysis (VCPA), uninformative variable elimination (UVE), competitive adaptive reweighted sampling (CARS) interval variable iterative space shrinkage approach (iVISSA) and successive projections algorithm (SPA) was used to extract the characteristic wavelengths based on the spectral data obtained by the optimal pretreatment method, and then the PLS-DA model was built. The hyperspectral image was subjected to masking and principal component analysis (PCA). Then the gray-level co-occurrence matrix (GLCM) was used to extract the texture parameters of the image with the highest principal component contribution rate, includingparameters of the angular second moment (ASM), entropy (ENT), contrast (CON), and correlation (COR), the PLS-DA models of data fusionwereestablished.The results showed that in the PLS-DA model of the original spectrum, the accuracies of the calibration set and prediction set were 89% and 86%; The PLS-DA model of the original spectrum based on de-trending preprocessing was the best, the accuracies of the calibration set and prediction set were both 90%, which were 1% and 4% higher than the original spectrum model, respectively. The De-trending-SPA-PLS-DA model based on the characteristic wavelength obtained the best results. The accuracies of the calibration set and prediction set were both 90%, which remained the same results as the optimal preprocessing model; The De-trending-SPA-COR fusion model obtained the best performance with 92% accuracy on both the calibration set and prediction set, which were 2% and 2% higher than the optimal spectral data model, respectively. Therefore, NIR-HSI,in combination with texture information, could realize rapid and non-destructive discrimination of Lingwu jujubes with different bruise grades.
Key words：Lingwu long jujube; Hyperspectral imaging; Fusion of spectra and image information; Discrimination of bruise grades
景怡萱，吴 迪，刘贵珊，何建国，杨世虎，马 萍，孙媛媛. 图谱数据融合的灵武长枣瘀伤等级判别[J]. 光谱学与光谱分析, 2023, 43(08): 2644-2648.
JING Yi-xuan, WU Di, LIU Gui-shan, HE Jian-guo, YANG Shi-hu, MA Ping, SUN Yuan-yuan. Fusion of Near-Infrared Hyperspectral Imaging (NIR-HSI) and Texture Feature for Discrimination of Lingwu Long Jujube With Different Bruise Grades. SPECTROSCOPY AND SPECTRAL ANALYSIS, 2023, 43(08): 2644-2648.
[1] LI Yan, LIU Gui-shan, LI Yue, et al(李 燕，刘贵珊，李 月，等). Food and Fermentation Industry(食品与发酵工业), 2022, 48(4): 145.
[2] GAO Ying-wang, GENG Jin-feng, RAO Xiu-qin, et al(高迎旺，耿金凤，饶秀琴，等). Food Science(食品科学), 2017, 38(15): 277.
[3] Lu Y Z, Saeys W, Kim M，et al. Postharvest Biology and Technology, 2020, 170: 111318.
[4] Chandrasekaran I, Panigrahi S S, Ravikanth L，et al. Food Analytical Methods, 2019, 12(11): 2438.
[5] Xiong J T, Lin R, Bu R B, et al. Sensors, 2018, 18(3): 700.
[6] Khodabakhshian R, Emadi B. International Journal of Food Properties, 2017, 20(sup3): S3149.
[7] Fan S X, Zhang B H, Li J B, et al. Postharvest Biology and Technology, 2016, 121: 51.
[8] Yuan R R, Guo M, Li C Y, et al. Postharvest Biology and Technology, 2022, 185: 111810.
[9] Fan N Y, Ma X, Liu G S, et al. Journal of Food Composition and Analysis, 2021, 103: 104110.
[10] Zhang J J, Ma Y H, Liu G S, et al. Food Control, 2022, 135: 108815.
[11] Cheng Lijuan, Liu Guishan, He Jianguo, et al. Food Chemistry, 2021, 342: 128351.
[12] Yang G T, He Y C, Li X, et al. ACS Omega, 2022, 7(5): 3889.
[13] Yang L, Gao H Q, Meng L W, et al. Food Chemistry, 2021, 334: 127614.
[14] Kim G, Lee H, Baek I, et al. Sensors and Actuators B: Chemical, 2022, 352(1): 130997.
[15] Sun Y, Pessane I, Pan L Q, et al. LWT-Food Science and Technology, 2021, 141: 110863.
[16] YUAN Rui-rui, LIU Gui-shan, HE Jian-guo, et al(袁瑞瑞，刘贵珊，何建国，等). Spectroscopy and Spectral Analysis(光谱学与光谱分析), 2021, 41(4): 1182.
