Acquisition and processing of multispectral data for texturing 3D models

This thesis deals with three problems concerning the use of a multispectral imaging spectrograph in applications of cultural heritage. The multispectral camera is part of an instrument developed within the project "Shape&Color" (CARIPARO, 2003-2005), coupling the spectrograph with a 3D laser scanner. Although the issues we have addressed arose from the characteristics of this specic instrument, they can be regarded as general problems concerning multispectral imaging, and are therefore of broader interest. The first part relates on the characterization of the spectrograph performance in measuring spectral reflectance under different illumination conditions. Four different illumination setups have been used to acquire a set of colored calibrated tiles. The system performance has been evaluated through a metrologically-inspired procedure, using as descriptors the average error (AE) and the average error standard deviation (AESTD), calculated by means of error propagation formula. The best results have been obtained with a metallic iodide lamp and an incandescence lamp used in a sequence, juxtaposing the spectral reflectance measured with the metallic iodide lamp in the 400-600 nm interval and that obtained with the halogen lamp in the 600-900 nm interval. The second presented issue concerns the problem of separating spectral illumination and spectral reflectance from the acquired color signal (the global radiation signal reflected by a target object). Since the latter can be considered as the product of illumination and spectral reflectance, this is an ill-posed problem. Methods in the literature estimate the two functions apart from a scale factor. The proposed solution attempts at the recovery of this scale factor using a statistical-based approach. The core of the algorithm consists of the estimation of the illumination intensity through a modification of the RANSAC algorithm, using relations derived from the physical constraints of the illumination and the spectral reflectance. The spectral reflectance is subsequently computed from the measured color signal and the estimated illumination function. The algorithm has been tested on four case studies, representing artworks of different pictorial techniques, color characteristics and dimensions. The results are good in terms of mean relative error, while the infinity norm of the relative error sometimes assumes high values. The last problem we have dealt with is that of using the multispectral images acquired with the Shape&Color scanner to texturize uncalibrated 3D data. What makes the problem worth addressing is that the spectral camera is not pinhole, but can be classified as a cylindrical panoramic camera. In this thesis, the general problem of estimating the extrinsic parameters of the camera from a known set of 3D-2D correspondences has been considered. The chosen approach is the classical reprojection error minimization procedure. As the projection operator is nonlinear, the objective function has a very complicated structure. Due to this and to the high dimensionality of the problem, the minimization results are strongly sensitive to the choice of the initial parameter values. This work proposes a way of finding a reliable initial point for the minimization function, so as to lower the risk of being trapped into local minima.