摘要
本研究使用近红外高光谱成像技术获取苹果的高光谱数据,对苹果糖度、酸度指标进行无损检测.针对高光谱数据量大、信息冗余多的特点,分别采用标准化(Standardization,SS)、标准正态变换(Standard normal vari-ate,SNV)、最小二乘平滑滤波(Savitzky-Golay smoothing filtering,SG)和多元散射校正(Multiplicative scatter correc-tion,MSC)对苹果的光谱数据进行预处理.针对高光谱图像波段多的特点,分别采用连续投影(Successive projec-tions algorithm,SPA)算法、竞争性自适应加权重(Competitive adaptive reweighted sampling,CARS)算法和随机蛙跳(Random frog,RF)算法选取苹果的特征波长.对提取的特征波长分别用支持向量机(Support vector machine,SVM)模型、卷积神经网络(Convolutional neural networks,CNN)模型和基于深度学习的定量光谱数据分析(DeepSpectra)模型对苹果的糖酸比进行预测.结果表明,基于深度学习的定量光谱数据分析(DeepSpectra)模型预测的正确率达到93.70％,有较高的精确度,可以较好地对苹果糖酸比进行预测.本研究将高光谱成像技术与基于深度学习的定量光谱数据分析模型相结合,实现了无损检测苹果糖酸比.
Abstract
The hyperspectral data of apples were obtained by using near-infrared hyperspectral imaging technology,and the indexes of sugar content and acidity were detected nondestructively.For the characteristics of large amount of hyper-spectral data and information redundancy,standardization(SS),standard normal variate(SNV),Savitzky-Golay smoot-hing filtering(SG)and multiplicative scatter correction(MSC)were used to preprocess the spectra of apples.According to the characteristic of hyperspectral images with many bands,successive projections algorithm(SPA),competitive adaptive reweighted sampling(CARS)algorithm and random frog(RF)algorithm were used to select the characteristic wavelengths of apples.Support vector machine(SVM)model,convolutional neural networks(CNN)model and quantitative spectral data analysis based on deep learning(DeepSpectra)model were used to predict the sugar-acid ratio of apples.The results showed that the prediction ac-curacy of DeepSpectra model was 93.70％,which had high accuracy and could be used to predict the sugar-acid ratio of apples.In this study,hyperspectral imaging technology and DeepSpectra model were combined to realize the non-destructive detection of the sugar-acid ratio of apples.
基金项目
浙江省"尖兵""领雁"重点科技计划项目(2022C04039)
维普
万方数据