基于高光谱成像技术的配方烟丝组分判别

Components Discrimination for Formula Tobacco Based on Hyperspectral Imaging

应用近红外(1 000~2 200 nm)高光谱成像技术开展了面对像素、面对样本的配方烟丝4种组分(叶丝、梗丝、薄片丝、膨胀丝)的判别研究。以样本高光谱图像的所有像素点光谱数据进行面对像素的组分判别;以样本所有像素点的平均光谱数据进行面对样本的组分判别。采用二阶导数法结合萨维茨基-戈莱平滑(SG)滤波对光谱数据进行预处理。通过面对像素数据的主成分分析,证实了基于面对像素的高光谱数据进行组分判别的可行性,以前5主成分建立的支持向量机模型很好地实现了叶丝与梗丝、叶丝与薄片丝的判别任务。建立了面对样本的4组分的K近邻和支持向量机判别模型,通过连续投影算法和二阶导数法进行特征波长选择,筛选出具有高判别准确率的波段,组分判别率达86.97%。

Near-infrared(1000-2200 nm)hyperspectral imaging technique was applied to the discrimination of components in formula tobacco,including cut lamina,cut stem,expanded tobacco and reconstituted tobacco.Two approaches,named pixel-wise and object-wise,were investigated to conduct this research.The pixel-wise components discrimination study was based on the spectral data of all pixels of the hyperspectral images of samples.Second derivative coupled with Savitzky-Golay(SG)algorithm was applied as preprocessing method for original spectral data.Through principal component analysis of the pixel data,the feasibility for component discrimination of the pixel hyperspectral data was confirmed.The established support vector machine(SVM)model based on first five components’data showed its excellent character in discriminating cut lamina and cut stem,cut lamina and reconstituted tobacco,obtaining intuitive discrimination results.The K-nearest neighbor and support vector machine discriminant model for the four components of samples was established.The characteristic wavelength was selected by the continuous projection algorithm and the second derivative method,and the band with high discrimination accuracy was selected,with a component discrimination rate reached 86.97%.