基于反向传播神经网络的拉曼光谱去噪方法

拉曼光谱技术作为一种典型的光学检测方法，因其独特的非侵入性、快速、原位和极高的特异性，在生物分析、疾病诊断及分子识别等众多领域得到广泛应用。拉曼光谱的指纹特性使其成为生物医学分析领域的重要工具，但拉曼散射信号微弱，数据处理分析大量依赖分析人员、自动化处理能力低等因素都会极大影响该技术在实际中的应用。实验设备、环境产生的噪声、待测生物样本的自发荧光等各种干扰因素使得高质量拉曼光谱数据的获得变得较为困难，各种随机噪声会干扰拉曼光谱图中指纹谱峰信息的识别，增大拉曼特征提取的难度，因此，噪声抑制在拉曼光谱预处理中显得十分重要。采用反向传播算法，从数据本身的特征出发，基于理想光谱片段和含噪光谱片段的线性差异，分别构建噪声判定神经网络和光谱去噪神经网络模型。以随机生成的一系列洛伦兹峰数学模型叠加生成拉曼光谱，对生成的拉曼光谱分别加入不同强度的噪声，以此为实验数据，对比新方法和经典滑动窗口均值法、Savitzky-Golay滤波法、傅里叶变换法、小波阈值变换方法去噪的结果。对均方根误差和信号噪声比两个指标进行分析，结果显示，在低噪声干扰下所有去噪方法都能较好地完成任务，但滑动窗口均值方法在光谱的边缘去噪效果会出现下降。随着噪声信号的增大，滑动窗口均值方法、S-G滤波方法、傅里叶变换方法去噪性能都出现明显下降。而基于反向传播神经网络的去噪方法要优于傅里叶变换法、滑动窗口均值法、S-G滤波器法，同时该方法在避免复杂参数寻优设置的同时，获得了和最优阈值小波变换方法近乎一致的去噪效果，大大简化了参数设置，更适合拉曼光谱去噪的自动化实现。

A Denoising Method Based on Back Propagation Neural Network for Raman Spectrum

As a typical optical detective method, Raman spectroscopy has been widely used in many fields such as biological analysis, disease diagnosis and molecular recognition due to its unique non-invasive, fast, in-situ and extremely high specificity. The fingerprint characteristics of Raman Spectroscopy make it an important tool in the biomedical field. However, many problems greatly influence the application of this technology, including weak Raman scattering signal, the large dependence on analysts for data processing and analysis, and the low capabilities of automated processing. It is usually difficult to obtain an effective and stable Raman spectrum of various interference factors, including the noise from experimental equipment or environment and spontaneous fluorescence of the biological sample. All kinds of random noises will interfere with identifying fingerprint peak information in Raman spectroscopy and increase the difficulty of Raman feature extraction, so denoising is an important task in the preprocessing of Raman Spectroscopy analysis. In this paper, using the backpropagation algorithm and considering the characteristics of the data, we built the noise judgment and denoising neural network model respectively, based on the linear difference between the ideal spectral segment and the noisy spectral segment. For the simulation experiments, the data of Raman spectra are composed of a series of randomly generated mathematical models of the Lorentz peak. Different intensities of noises were added to the randomly generated single Raman spectrum and 100 groups of Raman spectra, respectively. In addition, the new method has been compared with the classic sliding window average method, Savitzky-Golay filter method, Fourier transforms, and wavelet threshold transform method. Using the Root mean square error and Signal to noise ratio indicators for analysis, the results show that all methods can complete the denoising task under low noise, but the denoising effect of the sliding window average method is obviously reduced at the edge of the spectrum. With the increase of the noise signal, the denoising performance of the sliding window average method, S-G filtering method, and Fourier transform method have all decreased significantly. Overall, the denoising method based on a backpropagation neural network is better than the Fourier transform method, sliding average window method, and S-G filter method. This method avoids complex parameter optimization settings and at the same time, obtains the denoising effect that is almost the same as the optimal wavelet transform method. It greatly simplifies the parameter setting and is more suitable for the automated denoising of the Raman spectrum.