Generation of realistic radiographic noisy images to train an image de-noising neural network - Université de Lorraine Accéder directement au contenu
Communication Dans Un Congrès Année : 2022

Generation of realistic radiographic noisy images to train an image de-noising neural network

Résumé

A new way to illustrate digital radiographies issued from flat panel detectors is suggested. It is based on studying the noise distribution occurred in radiography through multiple histograms for each experimental real radiography with low dose exposure then generated in non-noisy image to create an artificial noisy image, and finally validate the similarity of the pair real/artificial noisy image with the real underexposed radiography achieved in same exposure conditions. The quantum noise is based on the hypothesis that it consists the major contribution between all kinds of noise caused by quantum sink in photon counting process. Indeed, despite all research confirm the dominance of quantum noise in low- dose radiography[1], but other type of noise coming from the detectors after a gain factor and from screen electronics consist also a considerable noise known as "impulsive noise"[2]. The sum off all these noise type is considered in order to draw the real noise distribution, and in a forward step, have the capacity to train correctly a convenient neural network giving as output a maximum de-noised radiography.

Dates et versions

hal-03889394 , version 1 (08-12-2022)

Identifiants

Citer

Toni Khalil, Fabrice Monteiro, Chafic Salame. Generation of realistic radiographic noisy images to train an image de-noising neural network. International Conference on Technologies and Materials for Renewable Energy, Environment and Sustainability 2021 - TMREES21-Gr International Conference, May 2021, Athenes, Grece, Greece. pp.020189, ⟨10.1063/5.0092300⟩. ⟨hal-03889394⟩
35 Consultations
0 Téléchargements

Altmetric

Partager

Gmail Mastodon Facebook X LinkedIn More