Multi-Components Low Dimensional Manifold Model For Photonlimited Poisson Noisy Image Reconstruction

Abstract

The exploitation of prior of image is very important for the reconstruction of photon-limited Poisson image which is urgent demand and particularly challenging in many application fields. Recently, a low dimensional manifold model has attracted attention in image processing, in which all image patches are treated as samples of a same manifold and the dimension of patch manifold is utilized as a nonlocal regularization prior. But in fact, different image patches often belong to different manifolds, and existing analysis shows that the patch manifolds corresponding to different image components often have different dimensions. Considering this difference, we propose to cluster the image patches into several groups corresponding to certain image components such as cartoon component, texture and edges firstly, and then utilize different low dimensional manifold regularizations for different image patch groups and propose a multi-components low dimensional manifold model for Poisson noisy image reconstruction. Numerical experiments show that our method can improve the result both visually and in terms of the peak-signal-noise-ratio and the featuresimilarity-index-measurement efficiently, especially for the Poisson images with extremely small number of photons.