Skip to Main content Skip to Navigation
Theses

Multispectral Analysis and spectral Reflectance Reconstruction of Art Paintings

Abstract : This thesis is devoted to the analysis of high definition multispectral images of art painting masterpieces and to the reconstruction of the spectral reflectance in each pixel of these images. To this end, we were mainly interested in: 1. the methods for spectral reflectance reconstruction of the surface of coloured materials in each pixel of an N-channel multispectral image (N >3), 2. the radiometrically controlled acquisition of multispectral images and the automatic calibration of the acquisition system. The problem of the spectral reconstruction being an ill posed inverse problem, multiple methods of reconstruction were developed in the scientific community. We propose a classification which enables us to better compare these techniques and to improve some of them. The improved methods have been implemented and tested. The methods usually used for the reconstruction of spectral reflectance from an N-channel image are mainly linear. To improve their precision and noise tolerance we introduced nonlinear techniques based on neural networks. Initially, multi-layer networks mixed with Principal Component Analysis (PCA) obtained good results. From this base we then worked on reconstruction using "Mixture Density Networks". This technique uses neural networks to estimate a probability distribution which is treated a posteriori to obtain a solution of the problem to be solved. The method was adapted to spectral reconstruction and very good results were obtained. This method was then enriched by developing an automatic system of architecture selection. Two strategies, genetic algorithms and random searches, were developed to this end. The radiometrically controlled acquisition of multispectral images and the automatic calibration of the acquisition system relate to the European project CRISATEL. The goal of this project is the spectral analysis and the virtual removal of varnish of art painting masterpieces. Within the framework of this project, we evaluated the performance of a new high definition multispectral camera. In particular, we studied the influence of the noise on the acquisition parameters and the spatial inhomogeneity of the light sources (two rotating elliptical projectors creating a luminous band synchronized with the movement of the camera CCD). We also characterized spectrally the filters, the CCD and the lamps. We evaluated the elements of the acquisition system. From the results of this evaluation an automatic calibration procedure was conceived and implemented. This automatic system determines the acquisition parameters (exposure time, CCD amplifiers gains and offsets). Its goal is to obtain the best dynamic range possible of the signal, and to gather the necessary data for the correction of the images. The elements to be corrected are dark noise, pixel sensitivity gain and the illuminant inhomogeneity. We also worked on other aspects related to multispectral imaging, i.e. the selection of the optical filters most adapted to the spectral reconstruction of a specific material, in our case oil pigments. For that we developed an optimisation technique which determines the parameters of a family of Gaussian filters which maximizes a quality criterion on the reconstruction obtained by using these filters. Finally, the work completed during this thesis was applied to art works. In the results chapter we present two examples: "Saint-Jacques le mineur" painted by George de la Tour and "Le départ pour Jersey" painted by Guillaume Fouace. The multispectral camera of the CRISATEL project is exploited by the Centre de Recherche et de Restauration des Musées de France (C2RMF). Currently, the camera is located in the Museum of Louvre where it indeed digitises art painting masterpieces.
Document type :
Theses
Domain :
Complete list of metadatas

https://pastel.archives-ouvertes.fr/pastel-00000761
Contributor : Ecole Télécom Paristech <>
Submitted on : Tuesday, February 15, 2005 - 8:00:00 AM
Last modification on : Friday, July 31, 2020 - 10:44:07 AM
Long-term archiving on: : Saturday, November 26, 2016 - 2:57:24 PM

Identifiers

  • HAL Id : pastel-00000761, version 1

Citation

Alejandro Ribes Cortes. Multispectral Analysis and spectral Reflectance Reconstruction of Art Paintings. domain_other. Télécom ParisTech, 2003. English. ⟨pastel-00000761⟩

Share

Metrics

Record views

2358

Files downloads

1370