PCANet: A Common Solution for Laser-Induced Fluorescence Spectral Classification

Abstract

Laser-induced fluorescence (LIF) technology is an advanced optical detection method, which has the advantages of fast, high precision and nondestructive testing, and is widely used in many fields. The general pattern recognition method for fluorescence spectral classification is highly dependent on pretreatment and dimension reduction. Specific pretreatment and dimension reduction methods are required for specific substances. Deep learning, especially the convolutional neural network (CNN), has the advantage of low dependence on data preprocessing and dimensionality reduction process, which makes it perform well in spectral classification. However, learning a useful CNN model for classification depends crucially on the expertise of parameter tuning and some ad hoc tricks, which is not suitable for chemometrics researchers. This paper presents a novel chemometrics tool for fluorescence spectra, principal component analysis network (PCANet), and more specifically a PCANet model with the optimized hyper-parameters (only optimized once). A two-stage cascaded PCANet model is constructed based on the liquor dataset, and the hyper-parameters are optimized and determined, which can make PCANet recognition model with the highest accuracy. Comparing the CNN model with two convolutional layers, the PCANet model is less affected by the size of the input image and the number of samples in the training set. At the same time, the performance of the two models on the mine water dataset is analyzed, and PCANet has higher recognition accuracy. That is to say, the PCANet is more accurate than CNN in fluorescence spectral classification, and its ability to expand application is stronger than that of the CNN model. The successful application of PCANet model with the optimized hyper-parameters (only optimized once) in the liquor dataset and the mine water dataset has important reference significance for the classification of fluorescence spectra of other substances in the future. INDEX TERMS Laser-induced fluorescence, spectral classification, principal component analysis network, hyper-parameter. I. INTRODUCTION Laser-induced fluorescence (LIF) [1], [2] is a spectroscopic method in which an atom or molecule is excited to a higher energy level by the absorption of laser light followed by spontaneous emission of light [3]. It was first reported by Zare [4], [5] and coworkers in 1968. The technology has many advantages, such as fast, high precision, high sensitivity, and nondestructive testing, so it is widely used in many fields like medicine, environment, biology, chemical industry, as well as food science. The associate editor coordinating the review of this manuscript and approving it for publication was Sungroh Yoon.

Artificial Intelligence Image Recognition Method Based on Convolutional Neural Network Algorithm

“…In Table 5, the test methods are AlexNet, VGGNet, ResNet, and Random Forest. There are also algorithms for HSC proposed in reference [30], Scatt-Net proposed in reference [31] and PCAnet proposed in reference [32]. It can be obtained from Table 5 that the random forest is inferior to the traditional CNN in network performance, and the algorithm proposed in this paper is higher than the single structure algorithm in network test accuracy.…”

“…Carroll et al [31] proposed a scattering convolutional neural network (Scatter-Net) based on wavelet transform, which uses wavelet transform to extract image high-frequency information hierarchically instead of the parameter learning process, which shows good performance in image recognition and classification tasks. Hu et al [32] proposed a model PCAnet that initializes the CNN convolution layer parameters by extracting the features of the image principal component, and has achieved good results in image recognition tasks. Wang et al [34] pointed out that the structure of the convolutional neural network itself is the main factor for the network to extract multi-level and multiscale features.…”

High-Throughput Screening of Hydrogen Evolution Reaction Catalysts in MXene Materials

“…Most relationships between the descriptors and target values were nonlinear in heterogeneous catalysis. Three types of nonlinear ML algorithms were employed to predict the HER activity in this work, including neural network methods (Elman ANNs), tree ensemble methods (RFR), and kernel methods (KRR and SVR). − The selected algorithms have showed excellent performance in predicting material properties . They have respective advantages, such as fast global searching (Elman ANNs), strong resistance of overfitting (RFR), and good generalization (KRR and SVR) .…”

Multispectral Imaging: A New Solution for Identification of Coal and Gangue

“…SVM (Support Vector Machine) [26], [27] is a supervised learning model with associated learning algorithms that analyzes data used for classification and regression analysis. An SVM constructs a hyperplane or a set of hyperplanes in a high or infinite-dimensional space, which can be used for classification, regression, and so on.…”