基于CWT和GRNN的可见-近红外漫反射光谱检测樱桃糖度的研究

Determination of Sugar Content in Cherry Fruit by Visible-Near Infrared Diffuse Reflectance Spectroscopy Based on CWT and GRNN

联合使用连续小波变换(continuous wavelet transform,CWT)和广义回归神经网络(generalized regression neural networks,GRNN)建立用于测定樱桃中糖含量的CWT-GRNN预测校正模型。利用CWT提取樱桃样本数据中反映含糖量的关键光谱特征,在CWT域中选择3个具有代表性的尺度,并在每个尺度下根据樱桃样本的可见-近红外光谱的特征将其划分为4个特征区间,从而构造12个特征输入到GRNN,GRNN的光滑因子取为0.0001。CWT-GRNN模型对20个预测样本集中的樱桃含糖量的预测相对误差在2%以内。结果表明,可见-近红外光谱技术可以快速、准确和无损地测定樱桃中的含糖量,本研究提出的方法可以用于果蔬产业的品质管理与控制。

CWT - GRNN model was constructed to predict sugar content in cherry fruit by combining continuous wavelet transform （CWT） with generalized regression neural networks （GRNN）. CWT was used to extract the key features, which were related to the sugar content of cherry. Three scales in the CWT domain were selected to efficiently extract the features of cherry fruit, and four feature spaces were divided according to the features of visible-near infrared （VIS-NIR） diffuse reflectance spectroscopy. Thus a feature vector, which contains twelve parameters, was input to the GRNN and the smooth factor of the GRNN was set as 0.0001. Twenty cherry samples were used to verify the performance of the CWT - GRNN model. Experimental results showed that the relative error of predicted samples was below 2%. CWT - GRNN model could be used to quickly, accurately and non-destructively predict the sugar content in cherry fruit. Also, the proposed method could be applied in control and evaluation in fruit and vegetable industry.