基于高光谱图像技术结合深度学习算法的萝卜种子品种鉴别

Classification of Radish Seeds Using Hyperspectral Imaging and Deep Learning Method

提出一种基于可见-近红外光谱技术的无损检测方法,以期实现对萝卜种子品种的鉴别。通过光谱成像系统采集6类常见萝卜种子的高光谱图像,并利用HSI软件提取光谱数据。使用Savitzky Golay(SG)平滑与多元散射校正(multiple scattering correction,MSC)叠加对光谱数据进行预处理以消除高频随机误差。采用堆叠自动编码器(stacked autoencoder,SAE)、连续投影算法(successive projections algorithm,SPA)和变量迭代空间收缩算法(variable iterative space shrinkage approach,VISSA)进行数据降维。利用Softmax与支持向量机(support vector machine,SVM)算法对全光谱和选取的特征光谱数据建立分类模型。结果表明:SAE-Softmax模型的分类效果最优,其训练集和预测集准确率分别达99.72%和96.22%。因此,利用可见-近红外光谱技术与深度学习算法结合的方法对萝卜种子的品种鉴别是可行的。该研究为种子品种无损检测分析提供参考。

Based on the VIS-NIR hyperspectral imaging technique,a rapid and nondestructive method was investigated for discriminating varieties of radish seeds.After removing noise band,the hyperspectral imaging system with spectrum range of480.46-1001.6 nm was used to collect six varieties of radish seeds containing 411 bands of hyperspectral images.Savitzky Golay(SG)smooth and multiple scattering correction(MSC)were used to eliminate high frequency superposition of random error.The stack autoencoder(SAE),the successive projections algorithm(SPA)and the variable iterative space shrinkage approach(VISSA)were used to reduce dimensionality of hyperspectral data of radish seeds.Softmax and the support vector machine(SVM)classification model were applied to identify radish seeds samples after dimensionality reduction.Experiment results showed that optimal model was the SAE-Softmax model,and accuracy of training set and accuracy of prediction set by the algorithm were reached 99.72%and 96.22%,respectively.The study demonstrated that VIS-NIR hyperspectral imaging technique was potential for nondestructive classification of radish seed.It was feasible and efficient to apply classification model into seed varieties nondestructive testing analysis.