分组全连接的近红外光谱定量分析网络

全连接网络作为深度学习中的一种典型结构，几乎在所有神经网络模型中均有出现。在近红外光谱定量分析中，光谱数据样本数量较少，但每个样本的维度高。导致了两个问题：将光谱直接输入网络，网络的参数量会十分庞大，训练模型需要更多的样本，否则模型容易进入过拟合状态；在输入网络前对光谱进行降维，虽解决了网络参数量过大的问题，但会丢失一部分信息，无法充分发挥网络的学习能力。针对近红外光谱的特性，提出了一种分组全连接的近红外光谱定量分析网络GFCN。该网络在传统的两层全连接网络的基础上，用若干个小的全连接层替代第一个全连接层，克服了直接输入光谱导致网络参数量过大的缺点。采用Tecator和IDRC2018数据集对该方法进行测试，同时与全连接网络FCN和偏最小二乘PLS两种方法进行对比。结果显示：在两个数据集上，GFCN预测效果均优于FCN和PLS。在只有少量样本参与建模的情况下，GFCN依然能够保持较高的预测效果。表明，GFCN可以用于近红外光谱的定量分析，并且适应样本较少的场景，具有重要的研究价值和广泛的应用场景。

Near-Infrared Spectral Quantitative Analysis Network Based on Grouped Fully Connection

As a typical structure in deep learning, a fully connected network appears in almost all neural network models. In the quantitative analysis of near-infrared spectroscopy, the number of spectral samples is small, but the dimension of each sample is high. It leads to two problems: if the spectrum is directly input into the network, the number of parameters of the network will be very large, which requires more samples to train the model. Otherwise, the model is prone to over fitting; reducing the dimension of the spectrum before inputting it into the network solves the problem that the number of parameters of the network is too large, but it will lose some information and cannot give full play to the learning ability of the network. According to the characteristics of near-infrared spectrum, a group fully connected near-infrared spectrum quantitative analysis network(GFCN) is proposed. Based on the traditional two-layer fully connected network, the network uses several small fully connected layers to replace the first fully connected layer, which overcomes the disadvantage of too many network parameters caused by a direct input spectrum. The GFCN model was tested with Tecator and IDRC2018 datasets and compared with a fully connected network (FCN) and partial least squares (PLS). The results show that the prediction effect of GFCN is better than that of FCN and PLS on the two datasets. In the case of only a small number of samples participating in the modeling, GFCN can still maintain a high prediction effect. The experimental results show that the GFCN can be used for the quantitative analysis of near-infrared spectrum and adapt to the scene with few samples. It indicates that the proposed model has important research value and good application prospects.