The color accuracy of an LED-based multispectral imaging strategy has been evaluated with respect to the number of spectral bands used to build a color profile and render the final image. Images were captured under select illumination conditions provided by 10-channel LED light sources. First, the imaging system was characterized in its full 10-band capacity, in which an image was captured under illumination by each of the 10 LEDs in turn, and the full set used to derive a system profile. Then, the system was characterized in increasingly reduced capacities, obtained by reducing the number of bands in two ways. In one approach, image bands were systematically removed from the full 10-band set. In the other, images were captured under illumination by groups of several of the LEDs at once. For both approaches, the system was characterized using different combinations of image bands until the optimal set, giving the highest color accuracy, was determined when a total of only 9, 8, 7, or 6 bands was used to derive the profile. The results indicate that color accuracy is nearly equivalent when rendering images based on the optimal combination of anywhere from 6 to 10 spectral bands, and is maintained at a higher level than that of conventional RGB imaging. This information is a first step toward informing the development of practical LED-based multispectral imaging strategies that make spectral image capture simpler and more efficient for heritage digitization workflows.
In this paper, we evaluate the quality of reconstruction i.e. relighting from images obtained by a newly developed multispectral reflectance transformation imaging (MS-RTI) system. The captured MS-RTI images are of objects with different translucency and color. We use the most common methods for relighting the objects: polynomial texture mapping (PTM) and hemispherical harmonics (HSH), as well as the recent discrete model decomposition (DMD). The results show that all three models can reconstruct the images of translucent materials, with the reconstruction error varying with translucency but still in the range of what has been reported for other non-translucent materials. DMD relighted images are marginally better for the most transparent objects, while HSH- and PTM- relighted images appear to be better for the opaquer objects. The estimation of the surface normals of highly translucent objects using photometric stereo is not very accurate. Utilizing the peak of the fitted angular reflectance field, the relighting models, especially PTM, can provide more accurate estimation of the surface normals.