Research On Blind Deconvolution Of Spectrum Based On Unpaired Learning

With the development of spectral detection and photoelectric imaging,multiband spectrum is always degraded by the random noise and band overlap during the acquisition of spectrum devices.Owing to the fixed spectrum degradation model,the existing spectrum deconvolution technologies are sensitive to the handcrafted model designed and manually selected parameters.The fundamental cause of these limitations during spectral analysis is that spectral processing is limited by 1-D signal without structural information available and insufficient training samples.It is feasible to collect unpaired noisy and clean spectrals in most real-world applications.This training method of unpaired learning is practical and valuable.The noise can be signal dependent but is spatially uncorrelated.In order to facilitate the unpaired learning of deconvolution networks,this thesis presents a two-stage scheme by incorporating self-supervised learning and knowledge distillation.The details are as follows:(1)For self-supervised learning,we suggest a dilated blind-spot network to learn deconvoluting solely from real noisy spectrals.Due to the spatial independence of noise,we adopt a network by stacking 1 × 1 convolution layers to estimate the noise level map for each spectral.Both the dilated blind-spot network and spectral-specific noise model can be jointly trained via maximizing the constrained log-likelihood.Given the output of the dilated blind-spot network and estimated noise level map,improved deconvoluting performance can be further obtained based on the Bayes’ rule.(2)As for knowledge distillation,we first apply the learned noise models to clean spectrals to synthesize a paired set of training spectrals,and use the real noisy spectrals and the corresponding deconvoluting results in the first stage to form another paired set.Then,the deconvolution network based on U-Net adapted to one-dimensional spectral data is trained by using the two pairing sets.Finally,identity loss is introduced to learn richer and more accurate structure and feature expression,and the final deconvolution model can be extracted.(3)Spectral analysis of medical tissue is an effective information extraction method.The model saved by machine learning training is used to classify the medical spectrum.In order to further verify the performance of infrared spectrum blind deconvolution algorithm based on unpaired learning,the prediction results of past convolution algorithm on two medical spectrum data sets of breast and ovary were compared.Through the classification accuracy of medical spectrum,the performance of each deconvolution algorithm in real medical scene is evaluated.(4)Experiments show that in terms of quantitative and qualitative evaluation,the unpaired learning method performs well in both synthetic noise spectrum and real noise infrared spectrum,and the blind deconvolution based on unpaired learning has good robustness and adaptability.