高光谱结合RF和1D-CNN对纸质快递文件袋的分类研究

Hyperspectral Study on the Classification of Paper Express Document Bags Combined with RF and 1D-CNN

建立一种对纸质快递文件袋快速无损的分类方法。利用高光谱对63个纸质快递文件袋样品进行检验,采用平滑、降噪和多元散射校正方式以消除高光谱的背景干扰。根据样品谱图的显著区别对样品进行分类,63个纸质快递文件袋样品可被分成3类。同时利用竞争性自适应重加权算法(Competitive Adaptive Reweighted Sampling,CARS)和随机森林算法(Random Forest,RF)对样品的特征光谱数据进行提取,并建立一维卷积神经网络(1D-CNN)模型,分别结合两种算法提取的特征波长对原始分类标签进行分析验证,并对未知样品进行预测识别。CARS-1D-CNN和RF-1D-CNN的模型识别准确率分别是84.21%、94.73%。通过比较发现,RF-1D-CNN可以实现对纸质快递文件袋更加有效的分类。该方法简单快速,样品用量少且无损样品,可为快递文件袋类的物证鉴定提供科学依据。

A fast and non-destructive classification method for paper express document bags was established.63 paper express document bag samples were analyzed by hyperspectral method.Smoothing,noise reduction,and multiple scattering correction were used to eliminate background interference from hyperspectral analysis.According to the significant differences in the sample spectra,the 63 paper express document bag samples were classified into three categories.Meanwhile,competitive adaptive reweighting sampling algorithm(CARS)and random forest algorithm(RF)were used to extract the feature spectral data of the samples,and a one-dimensional convolutional neural network(1D-CNN)model was established.The original classification labels were analyzed and verified by combining the feature wavelengths extracted by the two algorithms,and the unknown samples were identified by prediction.The model recognition accuracy of CARS-1D-CNN and RF-1D-CNN were 84.21%and 94.73%,respectively.By comparison,it was found that RF-1D-CNN could realize more effective classification of paper express document bags.The proposed method is simple and rapid,with a small sample size and no damage to sample,which can provide scientific basis for the identification of physical evidence in express document bags.