基于连续投影算法-遗传算法-BP神经网络的可见/近红外光谱木材识别

  目的  基于可见/近红外光谱技术，以10种木材为研究对象，探索不同预处理和特征提取方法下BP神经网络识别木材的效果。  方法  利用美国ASD公司生产的LabSpec 5000光谱仪采集10种木材的光谱图，分别进行移动平均法处理、移动平均法+多元散射校正(MSC)、移动平均法＋标准正态变量变换(SNV)、Savitzky-Golay卷积平滑算法(S-G滤波器)、S-G滤波器+MSC和S-G滤波器+SNV的预处理，运用主成分分析法(PCA)、连续投影算法(SPA)、SPA和遗传算法(GA)联合分别进行特征提取，将提取的特征结合BP神经网络进行木材识别试验。  结果  以SPA和GA联合提取光谱特征时，移动平均法+SNV的预处理效果最佳，以吸收峰为起始波段(W_initial=1 445 nm)、吸收峰个数为特征个数(N_tot=9)时，识别率较高，特征个数大部分减少为SPA提取特征值个数的一半左右。BP神经网络的平均识别速度提升明显。10种木材的平均识别率为98.0%，其中7种木材的识别率达到了100.0%。  结论  在移动平均法+SNV的预处理下，SPA和GA联合提取光谱图的特征，既可提高BP神经网络识别木材的正确率，又可提升识别速度。图3表6参23

Visible/near infrared spectrum wood identification based on SPA-GA-BP neural network

  Objective  The purpose of this study is to explore the effect of BP neural network identification under different pretreatment and feature extraction methods based on visible/near infrared spectroscopy technology, with 10 wood species as objects.  Method  The LabSpec 5000 spectrometer produced by American ASD company was used to collect the spectrograms of 10 species of wood, which were pretreated by moving average method, moving average method + multiplicative scattering correction(MSC), average method+standard normal variable transformation (SNV), Savitzky-Golay convolution smoothing algorithm (S-G filter), S-G filter+MSC and S-G filter+SNV. Meanwhile, principal component analysis(PCA), successive projections algorithm(SPA), and SPA combined with genetic algorithm(GA) were used for feature extraction respectively. The extracted features were combined with BP neural network for wood identification test.  Result  When SPA and GA were combined to extract spectral features, the moving average+SNV method had the best preprocessing effect. When absorption peak was used as the initial waveband (W_initial=1 445 nm) and the number of absorption peaks (N_tot=9) as the number of features, the identification rate was high, and the number of features mostly decreased to about 1/2 of the number of feature values extracted by SPA. The average identification speed of BP neural network significantly increased. The average identification rate of the 10 wood species was 98.0%, and the identification rate of 7 of them reached 100.0%.  Conclusion  Under the pretreatment of moving average method+SNV, the combined use of SPA and GA in spectral feature extraction can improve not only the accuracy of wood identification by BP neural network, but also the identification speed. Ch, 3 fig. 6 tab. 23 ref.]