基于小波系数图和卷积神经网络的太赫兹光谱物质识别

A Method of Terahertz Spectrum Material Identification Based on Wavelet Coefficient Graph

许多太赫兹光谱物质识别方法依靠寻找该物质在太赫兹波段范围内不同光谱表现出的不同特征来识别特定物质。吸收峰提取法是常用的光谱特征提取算法,但当光谱无明显特征吸收峰或峰位、峰值相近或难以识别时,难以利用吸收峰特征辨别物质。将机器学习和统计学习技术用于太赫兹光谱的识别中虽减少了吸收峰的干扰,但常常需要人为定义特征而导致分类误差。深度学习法能自动提取特征,但在识别前往往需要进行复杂的预处理操作,并且在特征提取的过程中容易丢失部分特征从而导致分类误差。针对以上问题,提出了一种基于小波系数图和卷积神经网络的太赫兹光谱识别方法。利用太赫兹光谱信号进行小波变换时,由于小波系数矩阵的每一行系数与原始光谱信号存在着对应关系,因此将太赫兹光谱的吸收系数通过小波变换在频率域上展开,能得到不同的二维的频率-尺度分布图,又称小波系数图。然后构造一个卷积神经网络(CNN)对小波系数图进行分类,可得到太赫兹光谱物质的分类结果。为了验证所提出算法的有效性,将三组小波系数图数据与原始光谱数据分别输入CNN、Support Vector Machin(SVM)、Multilayer Perceptron(MLP)三种不同的分类器作对比,从实验结果可以发现本文算法在三组数据中的识别率均达到了100%,说明相比于传统方法,本文方法能准确分类没有明显特征吸收峰的光谱,证明了使用卷积神经网络识别小波系数图的有效性。为了体现本文算法的优势,与小波脊线寻峰识别算法作对比,实验结果表明本文算法几乎不受峰频、峰位、峰值的影响,无论是识别不存在吸收峰的淀粉,还是识别相似度高的蔗糖和葡萄糖,都具有较高的识别率,分类准确率达97.62%,证明了所提算法的优越性。该算法为太赫兹光谱数据识别提供了一种新思路,同时也可以推广运用到其他谱图物质的识别中。

The terahertz spectrum material identification method mainly relies on finding the different characteristics of the different spectra of the substance in the terahertz band to identify a specific substance.The methods of absorption peak extraction are commonly used spectral feature extraction algorithm.However,when the spectrum has no obvious characteristic absorption peaks or peak positions,and peaks are similar or difficult to distinguish,it is difficult to use the absorption peak characteristics to distinguish substances.Although machine learning and statistical learning techniques to identify terahertz spectra reduces the interference of absorption peaks,it often requires an artificial definition of features to cause classification errors.The deep learning method can automatically extract features,but it often requires complex preprocessing operations before recognition,and it is easy to lose some features in the feature extraction process,leading to classification errors.A method of terahertz spectrum identification based on wavelet coefficient graph and convolutional neural network is proposed.When using the terahertz spectrum signal for wavelet transformation,each row of the wavelet coefficient matrix has a corresponding relationship with the original spectrum signal.The absorption coefficient of the terahertz spectrum is expanded in the frequency domain through wavelet transformation to obtain different two-dimensional frequency-scale distribution diagrams,which are also known as wavelet coefficient maps.Then a convolutional neural network(CNN)is constructed to classify the wavelet coefficient graph,and the classification result of the terahertz spectrum material can be obtained.To verify the effectiveness of the proposed algorithm,the three sets of wavelet coefficient maps and the original spectral data were input into three different classifiers of CNN,Support Vector Machin(SVM),Multilayer Perceptron(MLP)respectively for comparison.From the experimental results,we can find the recognition of the algorithm in the three sets of data.The rates reach 100%,indicating that compared with traditional methods,the method in this paper can still accurately classify spectra without obvious characteristic absorption peaks,which proves the effectiveness of using convolutional neural networks to identify wavelet coefficient maps.To show the advantages of the proposed algorithm in this paper,we compared it with the wavelet ridge peak-finding recognition algorithm.The experimental results show that the proposed algorithm is hardly affected by peak frequency,peak position,and peak value.Whether to identify the starch without an absorption peak or to identify high similarity sucrose and glucose,a high recognition rate is achieved by the proposed algorithm,and the classification accuracy rate is up to 97.62%,which proves the superiority of the proposed algorithm.The proposed algorithm provides a new idea for identifying terahertz spectrum data and can also be extended to the identification of other spectrum substances.