鱼类产品的新鲜程度研究一直是重要的课题, 其中挥发性盐基总氮(TVB-N)是一项重要指标, 该指标已列入我国食品卫生标准, 一般在低温条件下, 鱼类挥发性盐基氮的量达到30 mg/100 g时, 即认为是肉质变质的标志。 传统的物理检测方法不能够实现定量检测, 化学检测法则耗时长, 且需要专业人员进行破坏式检测。 为了克服传统光谱检测技术无法检测分析外部空间属性的缺点, 该实验采用波长范围在900~1 700 nm高光谱成像结合化学计量法实现了三文鱼的TVB-N含量检测。 首先对从市场买的新鲜三文鱼按照背面和反面(腹部)进行分割处理, 背面和反面(腹部)再10等分, 每条三文鱼制作成20个样本, 一共100个样本, 其中75个样本用于校正集, 25个样本用于预测集。 然后用高光谱成像系统采集三文鱼鱼样本的光谱数据, 再通过蒸馏法测定三文鱼TVB-N的含量, 并建立其理化值样本, 然后分别采用最小二乘支持向量机(LS-SVM)和偏最小二乘(PLS)模型对100个样本光谱全波长数据进行三文鱼TVB-N建模分析。 LS-SVM模型和PLS模型预测决定系数( R²)分别为0.918和0.907, 预测均方根误差(RMSEP)分别为2.312%和2.751%。 为了进一步提高运算效率和优化模型, 对全谱数据利用连续投影算法(SPA)提取到8个特征波长(956, 1 013, 1 152, 1 210, 1 286, 1 301, 1 397和1 464 nm), 基于8个特征波长分别建立SPA-LS-SVM和SPA-PLS模型, 模型预测决定系数( R²)分别为0.903和0.901, RMSEP分别为2.761%和2.801%, SPA-LS-SVM模型的结果优于SPA-PLS。 最后SPA-LS-SVM模型因其可靠性和有效性而被选择为最适合TVB-N预测模型, 基于图像处理编程技术将高光谱图像中的每个像素转换成相应的TVB-N值并以不同颜色表示, 实现了三文鱼肉TVB-N含量的可视化, 可以很形象的表达三文鱼的TVB-N的含量分布情况。 实验说明, 可利用高光谱成像技术预测三文鱼的TVB-N含量预测, 这为水产品的自动加工和分类奠定了基础, 渔业可以从高光谱技术中获益。

For fish products, the study of freshness has always been an important topic. Among them, the total volatile base nitrogen (TVB-N) is an important indicator. This indicator has been listed in China food hygiene standards. Generally, under low temperature conditions, when the amount of volatile base nitrogen in fish reaches 30 mg/100 g, it is considered a sign of meat deterioration. Traditional physical detection methods cannot achieve quantitative detection, andchemical testing methods are time-consuming and require professionals to perform destructive testing. In order to overcome the shortcomings of traditional spectral detection techniques that can not detect and analyze external space properties, this paper uses a wavelength range of 900~1 700 nm. Hyperspectral imaging technology combined with stoichiometry that has achieved the detection of TVB-N content in salmon. First, the fresh salmon bought from the market is divided into back and abdomen, and the back and abdomen are divided into 10 equal parts, each salmon is made into 20 samples, a total of 100 samples, 75 of which are used for calibration set, and 25 samples are used for prediction set, then use the hyperspectral imaging system to collect the spectral data of the salmon fish sample, next determine the content of salmon TVB-N by distillation, and establish its physical and chemical value samples, after that use the least square support vector machine (LS-SVM) and partial least squares (PLS) model performs salmon TVB-N modeling analysis on 100 sample spectral full wavelength data. The prediction coefficient of determination ( R²) of the LS-SVM model and the PLS model are 0.918 and 0.907, respectively, and the root mean square error (RMSEP) of the prediction is 2.312% and 2.751%, respectively. In order to further improve the computational efficiency and optimize the model, 8 characteristic wavelengths (956, 1 013, 1 152, 1 210, 1 286, 1 301, 1 397, 1 464 nm) are extracted from the full spectrum data using the successive projections algorithm (SPA). For the LS-SVM and SPA-PLS models, the model prediction coefficient of determination ( R²) is 0.903 and 0.901, and the RMSEP is 2.761% and 2.801%, respectively. The results of the SPA-LS-SVM model are better than those of the SPA-PL. Finally, the SPA-LS-SVM model was selected as the most suitable TVB-N prediction model due to its reliability and effectiveness. Based on image processing programming technology, each pixel in the hyperspectral image was converted into a corresponding TVB-N value and used different colors to indicate the visual distribution of the TVB-N content of salmon meat, which can vividly express the distribution of TVB-N content of salmon. Experiments show that hyperspectral imaging technology can be used to predict the content of salmon TVB-N, which lays the foundation for the automatic processing and classification of aquatic products. Fisheries can benefit from hyperspectral technology.