基于DCGAN的拉曼光谱样本扩充及应用研究

拉曼光谱检测方法依赖于化学计量学算法，深度学习是当下最炙手可热的方向，可应用于拉曼光谱进行建模。但是深度学习需要大样本进行训练，而拉曼光谱采集受制于器材和人力成本，获取大批量的样本需要更大成本，且易受荧光等因素干扰，这些问题都制约了将深度学习应用于拉曼光谱。针对以上问题，通过引入深度卷积生成对抗网络（DCGAN）提取拉曼光谱内部特征，对抗生成新的拉曼光谱，从而达到扩充数据集目的。同时和另一个扩充数据集的方法--偏移法进行对比，证明DCGAN的可靠性。设计生成光谱选取标准，选取高相似性的光谱填充数据集，为深度学习在拉曼光谱中的应用奠定基础。为了验证生成的光谱比原始光谱有更好的适用性，设计四组实验：（1）使用原始拉曼光谱输入到SVM进行分类，得到51.92%的分类准确率；（2）使用原始拉曼光谱输入到CNN进行分类，得到75.00%的分类准确率；（3）采用偏移法生成光谱，输入到CNN里进行分类，得到91.85%的分类准确率；（4）使用DCGAN生成光谱，输入到CNN里进行分类，得到98.52%分类准确率。实验结果表明，DCGAN能在只有少量拉曼光谱的情况下，通过对抗学习得到较好的生成光谱，且生成的光谱相比原光谱更加清晰，减少了可能的干扰因素，具有光谱预处理效果。通过DCGAN对抗生成大量高质量的数据填充到原有拉曼光谱数据集，扩充数据集的样本量，使得深度学习模型能够得到更好的训练，从而提高模型的准确率。该研究为深度学习方法应用于拉曼光谱分析技术提出了一个可行的方案。

Data Augmentation of Raman Spectral and Its Application Research Based on DCGAN

The detection method of Raman spectroscopy relies on the chemometrics algorithms,and deep learning is the most popular are at present,which can be applied to the modeling of Raman spectroscopy.However,deep learning requires large samples for training,while Raman spectral collection is limited by equipment and labor cost.Obtaining large quantities of samples requires a higher cost,and also is suffered by fluorescence and other factors,which all restrict the application of deep learning to Raman spectral.In view of the above problems,the paper introduces the deep convolution generation counter network(DCGAN)to extract the characteristics of Raman peaks in the Raman spectrum,and generates a new Raman spectrum to expand the data set.At the same time,the reliability of DCGAN was proved by comparing with the slope-bias adjusting method,another method to expand the data set.In this paper,spectral selection criteria are designed and generated to fill the dataset with highly similar spectra,which is the first step for the application of deep learning in Raman spectra.In order to demonstrate that the generated spectrum has good comformality with the original spectrum,the paper sets up four groups of experiments for comparison:(1)the original Raman spectrum is input to SVM for classification,and the classification accuracy is 51.92%,(2)the original Raman spectrum was input to CNN for classification,and 75.00%classification accuracy was obtained,(3)the slope-bias adjusting method was used to generate the spectrum,which was input into CNN for classification,and the classification accuracy of 91.85%was obtained,(4)DCGAN was used to generate the spectrum,which was input into CNN for classification,and the classification accuracy was 98.52%.The comparison of the four groups of results proves the superiority of the Raman spectrum generated by DCGAN.The experimental results show that DCGAN can generated much alike spectrum through antagonism learning with only a small amount of Raman spectrum,and the generated spectrum is clearer than the original spectrum,reducing some interference factors,and has a preprocessing effect on the spectrum.Taking the advantage of DCGAN,a large number of high-quality data can be generated and filled into the original Raman spectral data set,and the sample size of the data set can be expanded,so that the deep learning model could be better trained,thus improving the accuracy of the classification or other model.This paper proposes a feasible scheme for applying deep learning method to Raman spectroscopy.