光谱的预处理在光谱分析中占有非常重要的地位。 针对现有光谱去噪算法对弱峰保存能力差、 基线校正算法对光谱能量过扣除、 光谱特征峰定位不准确以及各种预处理算法串行处理造成的误差累计等问题, 设计了一个端到端的卷积神经网络。 该网络由两个模块组成: 基线校正和去噪模块和特征峰定位模块。 这两个模块相互连接又独立输出。 理想条件下, 可以依据光谱的线型函数和特征峰的位置拟合出无噪声无基线的光谱, 所以在基线校正和光谱去噪模块中连接特征峰定位模块的输出可以有效的提高去噪和基线校正的精度; 而高质量的光谱有助于更加精确的估计光谱峰的位置, 因此这两个模块相互连接可以有效提高重建光谱的质量。 光谱基线校正和去噪模块是一个前馈网络, 该模块由多个卷积层、 激活函数和批归一化层构成, 每一层均连接了特征峰定位模块的输出。 特征峰定位模块是一个多尺度特征融合网络, 该模块使用不同尺寸的卷积核将光谱分为不同的尺度, 融合大小不同尺度的特征估计光谱特征峰的具体位置。 在网络训练时, 使用不同温度、 湿度和不同预热时间的光谱仪获得光谱作为输入样本, 使用中国计量院的标准仪器获得光谱数据作为输出样本。 在实验中, 首先对合成的光谱分别添加不同信噪比的噪声和不同峰值的高斯基线, 分别评价该网络在噪声抑制、 基线校正、 光谱特征峰校正的能力; 然后将添加噪声和基线后的玉米的近红外光谱作为样本, 用最先进的算法对它们进行预处理, 然后用偏最小二乘法估计玉米中的水和油的浓度。 估计的浓度与用标准仪器测量的真实浓度进行比较, 以证明所提出的CNN的优势。 实验证明, 所设计的网络在单任务和多任务处理中均能取得良好的结果。 而且经过该网络处理的光谱在定量分析中可以得到更准确的结果, 具有较强的实用价值。

The techniques of Spectral Pre-Processingis important in qualitative spectral analysis. They were aiming at the problems of poor preservation of weak peaks in denoising algorithms, over-deduction of real spectral in baseline correction algorithms, inaccurate location of spectral peaks method, and error accumulation caused by serial use of various preprocessing algorithms. In this paper, we design a convolutional neural network (CNN). This CNN consists of two modules: baseline correction and denoising module and peak detection module, which are connected and output independently. In the ideal conditions, according to the spectrumlinear function, the ideal spectrum can be obtained by fitting the spectral peak location. Therefore, connecting the output of the spectral peak detection module to the baseline correction and denoising module can improve the accuracy of baseline correction and denoising effectively; and high-quality spectrum is the basis ofspectral peak detection. Hence the interconnection of two modules can improve the quality of the reconstructed spectrum effectively.The spectral baseline correctionand denoising module is a feed forward network consisting of several convolution layers, activation functions and batch normalization layers.The characteristic peak detection module is a multi-scale feature fusion network, that uses different-sized convolution kernels to divide the feature spectrum into different scales. The module fuses feature of different sizes to estimate the location of the characteristic spectral peak. During the training of the proposed CNN, the spectrum obtained by spectrophotometers with different temperatures, humidity and preheating times were used as input samples. The spectrum obtained from standard instruments of the National Institute of Metrology was used as output samples.In the experiment, noise with different SNR and Gaussian baselines with different peak values were added to the synthesized spectrum applied to evaluate the superiority of proposed CNN in denoising, baseline correction and spectral characteristic peak detection. Then the near-infrared spectrum of the corn, which was added noise and baseline, were taken as samples. Moreover, they, which were preprocessed using preprocessing algorithms, were applied to estimate the concentration of water and oilin corn by partial least squares. The estimated concentration compares with the real concentration, which is measured with a standard instrument, to prove the advantage of the proposed CNN. Experiments show that theproposed CNN can achieve good results in both single-task and multi-task preprocessing. Moreover, the spectrum preprocessed by the proposed CNN can get the most accurate results in quantitative analysis, which demonstrates that the proposed CNN has strong practical value.