Automatic Removal Algorithm Of Useless Scattering In Fluorescence Spectra Based On Deep Learning

Three-dimensional fluorescence spectra is widely used in fluorescence analysis technology and is also an important means for analyzing fluorescence data.It has the advantages of high sensitivity and rapid analysis speed.During the experiment,due to the limitation of the generation mechanism,the spectra has useless scattering such as Raman scattering and Rayleigh scattering.The useless scattering affects the accuracy of the effective information for analyzing the mixture and reduce the discriminating ability of the three-dimensional fluorescence spectra in the experiment.Therefore,it is meaningful to remove Raman scattering and Rayleigh scattering in the three-dimensional fluorescence spectra.In the existing traditional methods,The width of the useless scattering are often artificialy set by experience to remove useless scattering.This leads the is highly subjective and inefficient.The requirements knowledge of the spectra,make it impossible to remove useless scattering automatically.In order to solve problems,this dissertation proposes an automatic removing algorithm of useless scattering based on deep learning neural network.AttGAN,proposed by He and Zuo et al.in 2019,is a facial attribute editing algorithm.The algorithm marks various attributes of the face and generates a new face image with the required attributes while retaining other details.The algorithm proposed in this dissertation applies the AttGAN deep learning neural network to the process of removing useless scattering from three-dimensional fluorescence spectra.It locates the position of useless scattering in the three-dimensional fluorescence spectra automatically.And on the condition of accurately retaining the effective part of the information,the algorithm removes the useless scattering and fill the spectral intensity of the removed position.When training the AttGAN network,three-dimensional spectral data collected from the experiment is not enough to train the network.This dissertation combines the generated spectra with the experimentally collected spectra to form the AttGAN data set to impliment the training and testing of the neural network.In the process of generating spectra,both of the grid mode and random mode are applied.The results show that the AttGAN deep learning neural network has a good ability in removing useless scattering,which can effectively remove the useless scattering part and accurately retain the effective information.Finally,this dissertation compares the algorithm with the traditional algorithms.The comparing results show that the difference between the results of the algorithm in this dissertation and the target information in the original data is small,and the useless scattering can be removed accurately.Without artificial parameters setting,the dissertation proposed algorithm shows more efficient than the traditional methods.It is an ideal algorithm for removing useless scattering.