Multi-components Low-dimensional Manifold Regularization Based Construction Of Photon-limited Poisson Noisy Image

Photon counting imaging is widely used in astronomical imaging,night vision imaging,medical imaging and other important fields.However,due to the influence of the imaging environment and hardware equipment,the number of photons collected by the equipment is seriously insufficient.At this time,shot noise that obeys the Poisson distribution will seriously reduce the imaging quality and generate low-quality photon-limited Poisson images.In the photon-limited Poisson image,the self-similarity of the local pixel gray,the gray consistency and continuity of the local geometry of the image are severely damaged,as a result,there are bias in the detection of the local geometric structure of the image and the geometric similarity measurement of non-local image blocks.It also makes many methods are no longer effective for photon-limited Poisson images,which are feasible for sufficient photon Poisson images.In recent years,the demand for high-quality imaging in the fields of security surveillance,military,astronomy and medicine has become higher and higher.The high-quality reconstruction of photon-limited Poisson images has become a research hotspot in the field of image reconstruction theory and applications and one of the difficulties.For photon-limited Poisson image reconstruction,a low dimensional manifold regularization method suitable for Gaussian noise is introduced into Poisson image processing.Considering the differences in image components,a multi-component low dimensional manifold regularization Poisson image reconstruction method is proposed.In addition,according to the situation of photon-limited Poisson image information defect,under the framework of 'repair+reconstruction',the low dimensional mani-fold regularization method is coupled with the matrix compensation method based on compressed sensing to further improve the quality of Poisson image reconstruction.The main research contents and innovations of this thesis are:1.A photon-limited Poisson image reconstruction method based on multi-component low dimensional manifold regularization is proposed.This thesis introduces the low dimensional manifold regular term for Gaussian noise into the Poisson image reconstruction model firstly.Then,based on the different dimensions of the block manifolds of different components(smooth region and texture region)in a natural image,a multi-component low dimensional manifold regularization Poisson image reconstruction model.In the multi-component low dimensional manifold Poisson reconstruction model,the block manifold of the image is divided into different image component sub-manifolds firstly,the dimension of each manifold is used as regu-lar term,and the corresponding dimensions of different sub-manifolds are regular terms are given different weights to better reconstruct different morphological components in the image.Numerical experiments show that,in terms of visual effects and objective indicators,the multi-component low-dimensional manifold regularized Poisson image reconstruction model further improves the image reconstruction effect based on the low-dimensional manifold regularized Poisson image reconstruction model.2.A multi-component low dimensional Poisson image reconstruction method based on matrix completion is proposed.Due to photon-limited Poisson image information defect,this thesis utilizes the"repair+reconstruction" framework to reconstruct the image.In this thesis,the matrix completion method based on compressed sensing and the multi-component low dimensional manifold regularization method are coupled.Based on Poisson im-age features,a special sampling operator is designed,and the model is solved using the ADMM algorithm.Numerical experiments show that the matrix completion based multi-component low dimensional manifold regularization Poisson reconstruction mod-el proposed in this thesis effectively improves the visual effect and objective index of the reconstructed image.On this basis,based on the ADMM algorithm,the above method is further ex-tended to the'plug-and-play' framework,and different Gaussian denoising methods can be used to substitute the algorithm,which reflects the flexibility of the method.