In this thesis, we will focus on the physical implementation of visual cryptography, which consists in sharing a secret message between several unmeaning images, so-called shadow images, at least one of them being printed. By the principle of the method, no information leaks about the message until the images are properly stacked together. As the alignment of the shadow images hampers the deployment of the visual cryptography in practice, we develop a dedicated image registration method. In contrast with existing methods, ours is not intrusive. We make use of the particular shape of the elementary constituents of the shadow images, the shares, to extract in the Fourier domain the main parameters of the geometrical transformations occurring between the superposed images. We prove that this method allows subpixel accuracy in shadow images registration. We benefit from such ability by implementing visual cryptography in an image projection configuration : the digital shadow image is projected onto the printed one. In this way, the registration is performed automatically by using a digital camera (the resulting superposition being observable by the eye). For the purpose of authentication, one has to deal with specific attacks: the shadow image attached to a given document could be tampered with or copied. In order to prevent such attacks, we have increased the di_culty to reproduce the shadow image by considering color. This approach requires a complete management of colors. Thanks to recent advances in color reproduction, we are able to predict the reflectance and transmittance spectra of supports printed in color. In this thesis, we develop new spectral prediction models namely for piles of printed transparencies as well as for transparencies stacked onto papers, all printed in color. Thus, we are able to predict the color of each share in a shadow image to be printed and to achieve color matching i.e. we are able to reach a color by various combinations of superposed colors. Such a prediction allowed us to introduce a new approach in visual cryptography: color matching when revealing the secret message to be shared into two (or more) shadow images, in order to authenticate the shadow images provider. As the prediction models are sensitive to the calibration of the printing system (printer, inks, supports, halftoning and geometry measurement conditions), the use of special materials will increase the di_culty to generate visually acceptable fake pairs of shadow images
