Research On The Technology Of Image Restoration Based On Prior Learning

Image restoration is the process of restoring original images from their degraded or damaged observations.Image restoration algorithms are widely used in medical imaging,satellite imaging,surveillance systems,and remote sensing imaging,etc.Image restoration methods are usually based on filter theory,spectrum analysis,wavelet,partial differential equations or stochastic modeling.In this thesis,we focus on stochastic modeling.The image is modeled as a random variable that follows to certain prior distributions.The statistical features of the natural image are captured,and then the restored image is reconstructed using the priors based on the maximum a posteriori estimate.Image priors are the key to solving the ill-posed image restoration problem.The early design of image priors is mainly based on the physical characteristics or local characteristics of the image.In recent years,researchers pay more attention to prior learning which is based on the statistical characteristics of the images.The performance of image restoration has been improved by the prior learning.This thesis studies image restoration techniques based on prior learning,which is supported by National Natural Science Foundation of China“low-dose CT reconstruction based on statistical distribution of sparse variables”.In this thesis,our research focuses on three aspects:patch priors model based on Gaussian mixture model,global image priors model based on deep generative model,and deep discriminative learning model for underwater image restoration.The major research contributions are listed as follow:1.In order to apply multi-scales and global statistical constraints to the Gaussian mixture patch prior model and further improve the quality of restored image,a multi-scale patch prior model based on gradient histogram constraint with Wasserstein distance is proposed.The global statistical priors of the gradient histogram are combined with the multi-scale patch prior model.The Wasserstein distance is used to measure the statistical distance between the gradient histogram of the restored image and the reference gradient histogram,and the difference constraint is integrated into the multi-scale patch prior framework,which further enhance visual quality of restored image.Different scales images are obtained after applying filtering and down-sampling to the target images.The same-sized scale patches are extracted from the different scales images and are enforced with local low-dimensional priors.Scale-invariance is achieved by applying Gaussian filtering before down-sampling to the original scale image and adjusting the parameters of the filter,then all local models of all scales can remain unchanged.The half-quadratic splitting and optimal transfer theory are used to optimize the objective function.The algorithm has an effective analytical solution and good convergence.Applying the proposed model to the image denoising and deblurring tasks,the visual quality of the restored image is better than the traditional method.2.The global Gaussian mixture patch prior model fails to fully utilize the coherence of neighboring patches in the image and local patch model is not stable in solving image inpainting problems.A hierarchical Bayesian-based local Gaussian mixture patch prior model is proposed.Using the priors knowledge of the model parameters,the probability distribution of the mean and covariance matrices are modelled by the Normal-Wishart distribution,which makes the patch estimation process more stable.Based on the coherence of neighboring patches,the set of similar patches in the window can be derived by the multivariate Gaussian probability distribution of specific mean and covariance.The similarity is measured by the ₂ norm metric,which is accelerated by using the summed square image and fast Fourier transform.Using the aggregation weights of the Gaussian distribution similarity based on Mahalanobis distance,combined with the Gaussian similarity of the spatial domain on the images,the statistical characteristics of the natural image are better fitted.Applying the proposed model to image denoising and inpainting,the denoising average results of the proposed method is better than those of the sparse-based methods.The latent texture can be well recovered in image inpainting whether with random sampling or uniform sampling.3.Aiming at the shortcomings of fix components and mainly relying on external learning of the traditional Gaussian mixture patch prior model,a new image prior model based on Dirichlet process mixture model is proposed.The model learns external generic priors from a set of external clean images and internal priors from a given degraded image.Due to the accumulative property of the statistics in the model,the integration of internal and external priors is naturally achieved in image restoration.Through the addition and merging of model components,the model complexity can be adaptively changed with data,more high-quality and compact models can be learned.In order to solve the variational posterior of all hidden variables,a scalable variational algorithm based on batch-updating method with adding-merging mechanisms is proposed.The new algorithm improves the traditional coordinate ascent algorithm which is relatively inefficient under large datasets and often falls into the local optima.Our model has advantages both on objective quality assessments and visual perception in image denoising and inpainting experiments comparing to traditional methods.4.Well-trained deep generative networks can learn the low-dimensional manifolds of images.Aiming at the incompleteness of theoretical research on the deep generative prior model,the inverse problems of the deep generative model is studied.It is proved that for the shallow deconvolution generative network,the gradient descent can effectively inverse the network and estimates the latent code.It is proved that the projection gradient algorithm is convergent under the condition that the objective function satisfies the condition of restricted strong convex and restricted strong smoothing.Aiming at the situation of this inability of the current deep generative network completely learning the distribution of rich and complex natural images,a new image restoration algorithm beyond the range of the generative network is proposed.Both the loss terms of the images in the range of generator and those of out-of-range images are considered and tied by the additional range error between in-range images and out-of-range images.The amount of loss slack can be controlled by adjusting the weights attached with each loss term in the final objective,which achieve the ability to exceed the representation of the generative networks.The proposed algorithm is applied to non-blind image restoration,such as compressed sensing and image restoration,and blind image deblurring application.Compared with the traditional method,both the vividness and fidelity of the restored images are better.The proposed algorithm can be further extended to solve other inverse problems in signal processing and computer visions.5.The single prior in the traditional model-based underwater image method often produces inaccurate estimation of the medium transmission in some areas of the image.This thesis proposes a new underwater image restoration with saliency-guided multi-scale priors’fusion.To accurately predict the medium transmission,the underwater dark channel prior and the intensity attenuation difference prior are fused.The contrast and brightness of restored underwater images can be _{iauosdnu utvdpseeeurrstdw}^{irddoimevbcpeuorrsdostiev} _a ^nt _taorotfy^{reey d sn} _e ^{o s}ee _{l rt ce}^{fat ossw}_{hioae rmm}^fe _{n a}^{oawtr hwkee lal i}_{eptgnehetcei o}^{ts e e}_{tbii daomesnra}^{axrpsi r ts} _e^o,,^{ytt ih}_{g d} ^{pp neoeg}_{ewao n fhn}^sss _e,_di^ce _cat^{mod al nooa drun}_{tthhuh efresie}thhuh efresie ^e_{e n s e}^{ndld-drebeevarais}_{afnulelcnaytor uedclreoaertsm-e}^{wlaeitazidteo esnu}._d^{r ne icdnmaeda}_{pdTe lhtceoeto} ^nr-_{etdd sh}^{gtw eoba-}_{ie setr} ^e _ch^tr _s^{e ee}_{rae ei}^nrr _n^s _{atmdrwi a}^{tdei omur maonatvdigoeee}_{niadnatiitesnorc} ^rd_rrgt ⁿ _{i yc}^{rw emeasft}_{omlpfaie sn tsahai}^{otfeoedrr ceiatlit mivbo}_{ntefo ideer t sn} ^aeanlgs_{h cttiea hr}^{ye e.}_need ^{dm r}_{e lnbcaeetot}^{Ieeontsn th} _{e dlwneetor a}^{ovoariddaes tv}_rcr,_rk^wi _one^eo._{dse ted}^{rc nsoo}._Bo^af,^{n r} _{yr t}^Tsit _wehcshr^{iahhlde ee}_{oito c}^e _{hmuhegb iihcsl}^{relninancccogok} _n ^d _e×_ittnahhngee^{td ehoeref-r}₁loss,multi-scale structural similarity measure loss and adversary loss,more details of the restored underwater image can be preserved.Experiments were carried out on underwater image datasets with various scenes.Compared with the traditional methods,the proposed model is more advantageous on both visual perception and image quality measurement.