高光谱图像与稀疏核典型相关分析冷鲜羊肉新鲜度无损检测

羊肉新鲜度受多种因素影响，通常由多个指标来综合评价，常规试验操作复杂不适合在线检测。高光谱成像数据能够反映羊肉新鲜度变化过程中多种成分的变化信息，但是光谱特征提取与评价模型的建立对最终结果影响较大。为了研究高光谱成像与多指标的快速检测羊肉新鲜度的可行性，提出一种稀疏核典型相关分析方法，借助实验室测定的多个标准值，研究多指标的羊肉新鲜度无损检测。采集了70个代表各级新鲜程度的羊肉样本400~1 000 nm高光谱图像，采用实验室方法测定了挥发性盐基氮(TVB-N)和菌落总数(TAC)标准值，选择感兴趣区域(ROIs)提取代表性光谱图像，利用所提出的特征提取方法提取光谱特征信息，并按照3:1划分校正集和预测集，利用三层神经网络进行分类识别试验。结果表明，新鲜度等级分类总体精度(OA)为0.939 3，Kappa系数为0.906 0，均方根误差(RMSEC)为0.297。研究表明，所提出的多指标光谱特征提取方法可用于快速无损检测羊肉新鲜程度，为采用高光谱成像综合多个新鲜度检测指标，改善由于单一检测指标造成评价模型的适用性和鲁棒性提供了基础。

Study on Detection of Chilled Mutton Freshness Based on Hyperspectral Imaging Technique and Sparse Kernel Canonical Correlation Analysis

The mutton freshness is affected by many factors, which is usually evaluated by a number of indicators, and its routine detection is complicated and not suitable for online detection. Hyperspectral imaging data can reflect the changes of components in the process of mutton freshness changing, but the establishment of spectral feature extraction and evaluation model has a great influence on the final result. In order to study the feasibility of rapid detection of mutton freshness with hyperspectral imaging technique and multi-parameter indicators, this paper proposed a sparse kernel canonical correlation analysis method, and researches comprehensive evaluation of mutton freshness on multi-parameter using laboratory standard values. In this study, 400~1 000 nm hyperspectral images were collected from 70 mutton samples, and the standard values of total volatile basic nitrogen (TVB-N) and total aerobic plate count (TAC) were determined with laboratory methods. The representative spectra of mutton samples were extracted and obtained after selection of the region of interests (ROIs). The spectral feature information is extracted by using the feature extraction method proposed in this paper. The samples of calibration set and the prediction set are divided at the ratio of 3:1. The experiment of classification and recognition using three layer neural network shows that the overall accuracy (OA) is 0.939 3, the Kappa coefficient is 0.906 0, and the root mean square error (RMSEC) is 0.297. The research shows that the multi-parameter spectral feature extraction method proposed in this paper can be used to detect the freshness of mutton quickly and nondestructively. This paper provides a basis for improving the applicability and robustness of the evaluation model due to the single detection indicator by using the hyperspectral imaging technique to synthesize the spectral information of several freshness indexes.