Research On Deconvolution Algorithm Of Infrared Spectrum Based On Generative Adversarial Network

With the development of optoelectronic technology,infrared spectroscopy has been widely used in industry and scientific research in recent years,especially in compound determination and spectral processing.For the collection of infrared spectra,it will be affected by many factors,among which random noise and frequency band overlap are the most common.The traditional method of simulating the degradation process is to use partial differential equations,with the goal of denoising,improving edges,and preserving the original image information.However,this method requires manual design of the model and selection of parameters,which will reduce performance and cause problems in real-world applications.For the deconvolution of the infrared spectrum,the traditional method is to obtain the potentially clean infrared spectrum through non-blind deconvolution and semi-blind deconvolution methods such as DSPNet and SBD,but both of these deconvolutions require fuzzy kernel priors.This thesis attempts to apply deep learning to the blind deconvolution of the most advanced infrared spectrum for the first time,and proposes a generation of anti-blind self-deconvolution neural network,which can reconstruct the original infrared spectrum from the fuzzy infrared spectrum without any prior.This thesis uses two generation networks and,which are used to model the clear infrared spectrum and the fuzzy kernel respectively.Join the generative confrontation network GAN,draw on the idea of the derivative algorithm of VAE/GAN,and use the output of the infrared spectrum generation network as the generator model G of GAN to improve the quality of the reconstructed infrared spectrum.For the optimization of the loss function,this thesis introduces Total Variation(TV)regularity to the latent infrared spectrum and the fuzzy kernel respectively,which makes the boundary characteristics of the infrared spectrum maintain better results.The whole process is a kind of self-supervised learning.The degraded infrared spectrum is directly input into the generation network without training data set.This method can be called Self Deblur-GAN.Finally,experiments are carried out on manual data sets and real infrared spectroscopy,and Self Deblur-GAN is compared with traditional infrared spectroscopy deconvolution methods.The experimental results show that the Self Deblur-GAN method can achieve infrared spectroscopy deblurring more significantly,and has more good superiority.