Jon Yngve Hardeberg

TL;DR: A spectral characterization of the acquisition system taking into account the acquisition noise is performed and the spectral reflectance of each pixel of the imaged surface is estimated by inverting the model using a principal eigenvector approach.

Abstract: The goal of the work reported in this dissertation is to develop methods for the acquisition and reproduction of high quality digital color images. To reach this goal it is necessary to understand and control the way in which the different devices involved in the entire color imaging chain treat colors. Therefore we addressed the problem of colorimetric characterization of scanners and printers, providing efficient and colorimetrically accurate means of conversion between a device-independent color space such as the CIELAB space, and the device-dependent color spaces of a scanner and a printer. First, we propose a new method for the colorimetric characterization of color scanners. It consists of applying a non-linear correction to the scanner RGB values followed by a 3rd order 3D polynomial regression function directly to CIELAB space. This method gives very good results in terms of residual color differences. The method has been successfully applied to several color image acquisition devices, including digital cameras. Together with other proposed algorithms for image quality enhancements it has allowed us to obtain very high quality digital color images of fine art paintings. An original method for the colorimetric characterization of a printer is then proposed. The method is based on a computational geometry approach. It uses a 3D triangulation technique to build a tetrahedral partition of the printer color gamut volume and it generates a surrounding structure enclosing the definition domain. The characterization provides the inverse

TL;DR: This work proposes a new approach to denoising, inpainting, and super-resolution of hyperspectral image data using intrinsic properties of a CNN without any training, and shows the performance of the given algorithm to be comparable to theperformance of trained networks, while its application is not restricted by the availability of training data.

TL;DR: A novel method for reflectance recovery is introduced that finds the smoothest spectrum consistent with both the colorimetric data and a linear model of reflectance.

TL;DR: An automated melanoma recognition system based on deep learning method combined with so called hand-crafted RSurf features and Local Binary Patterns is proposed, which demonstrates high classification accuracy, sensitivity, and specificity of the proposed approach when it is compared with other classifiers on the same dataset.