Spectral Color Reproduction of Paintings

A spectral-based imaging system was constructed consisting of a two-sequential-absorption-filter-CFA digital camera and a seven-color inkjet printer and used to image and print a post-impressionist style painting such that matches were generated for CIE Illuminants D65 and A. Camera calibration was learning based using the Matrix R method. Rendered images for both illuminants were inputted to a color separation algorithm. First, conventional colorimetric gamut mapping was performed for the designated primary illuminant. Second, each pixel was transformed into a metameric printer gamut, that is, all possible ink combinations matching the primary illuminant. The ink combination with the smallest color difference for the secondary illuminant was selected as the color separation. Prints were prepared for each combination of primary and secondary illuminant and repeated for both CIE standard observers. The method was successful within limits of camera and printer spectral accuracy, ink design, and illuminant and observer metamerism. Introduction Reproducing the appearance of paintings in print is a common occurrence. In most museums, visual editing is an integral part of the workflow [1]. The amount of visual adjustment and the number of iterations required to achieve acceptability depends principally on the spectral properties of the artwork, the spectral sensitivities and color management of the camera, differences between the actual and assumed viewing illuminants, differences in size between the original and reproduction, the spectral properties of the printing materials and color management, and the matching objective (colorimetric, preferred, etc.). If the painting and print have the same size and matching is desired for multiple conditions, spectral reproduction becomes the matching objective. This objective has been achieved in the past using a multi-spectral camera and spectral printing models [2-5]. The limiting factor was the extreme computational load in generating color separations because, in essence, instrumental-based color matching using non-linear constrained optimization was performed at each pixel. During the last years, there have been significant advances in spectral imaging of artwork [6], spectral processing [7-9], and spectral printing [10]. (These references are exemplars and not a definitive list.) Although these advances overcome the processing limitations of past research, spectral reproduction of artwork is still limited by metamerism, a result of printing inks’ inability to span the spectral properties of artist materials. Thus, it was of interest to evaluate spectral color reproduction of a painting with a focus on metamerism. Experimental The general methodology is shown in Figure 1. A painting is imaged using a multi-spectral camera. Following spectral and spatial processing (denoising and sharpening), two CIELAB images are rendered for a standard observer and for primary and secondary illuminants. The coordinates for the primary illuminant are mapped within the printer’s color gamut. For each pixel, all ink combinations and their amounts are calculated yielding a metameric ensemble of spectra (metameric mismatch gamut for the secondary illuminant). This is possible when the number of inks exceed three. For each metamer, a color difference is calculated for the secondary illuminant. The ink combination leading to the smallest difference is selected. Figure 1. Experimental workflow. Yellow indicates imaging system and blue indicates printing system. Painting A set of artist acrylics were defined that reasonably spanned the spectral gamut of artist pigments [11]. These were used to make a painting in the style of Vincent van Gogh’s Church at Auvers, shown in Figure 2. Imaging system A 22-megapixel Sinar digital camera system was modified as described by Berns [12], resulting in a six-channel multispectral camera. Lighting consisted of two tungsten-halogen Elinchrom Scanlite Digital 1000 sources affixed with Chimera diffusers. Lee #201 bluish gelatin filters were placed between the lamps and diffusers to achieve more spectrally uniform lighting (CCT = ~5400 K). Each light illuminated the object plane at 45° from the normal. Images were collected of a gray surface for flat fielding, a GretagMacbeth ColorChecker DC (CCDC) and a custom artist material target for calibration, and the painting. 484 ©2008 Society for Imaging Science and Technology A learning-based technique, known as the Matrix R method, was used to estimate spectral reflectance [13]. This method optimizes colorimetric and spectral accuracy simultaneously. Accordingly, transformations were derived for illuminants D65 and A and for both the 1931 and 1964 standard observers. Thus, four 16-bit CIELAB images were rendered. Sharpening was performed on the L* plane and noise was reduced on the a* and b* planes. Figure 2. Auvers, Bernard Lehmann (16” x 20”) 2007: rendered for D65 and the 1931 standard observer. Measurement locations notated. Printing system An HP Z3100 Photo inkjet printer was controlled by an Onyx Production House RIP 7.0. Of its 12 inks, only cyan, magenta, yellow, black, red, green, and blue were used. To reduce fluorescence in the final prints, a paper without optical whitener was used, Felix Schoeller (H74261) 270g/m2. A calibration target of 7725 patches spanning the printer’s spectral gamut and constrained to the paper’s ink limits was printed and measured using an X-Rite i1iSis. The ink gamut was divided into four-ink sub-gamuts [2] and the cellular extension of the Yule-Nielsen Spectral Neugebauer equations was used to characterize each spectral sub-gamut in similar fashion to Chen [14]. The separation method combined spectral gamut mapping as well as model inversion in one single step. The basis of the separation is the spectral gamut-mapping framework described in detail in another CGIV 2008 paper [15]. The separation method compensated for both color discrimination and printing quantization artifacts. Using a traditional gamut mapping (chroma compression while preserving hue and lightness) within a hue-linearized [16] CIELAB color space for the primary illuminant, the CIELAB image was transformed into a metameric printer gamut. A 3D histogram was created for this image and for each sub-model the colorant space was sampled in 1% steps resulting in ~100 million different ink combinations. For the 20 sub-models a total of 2 billion colors were transformed by the forward model for the primary illuminant and tested using the 3D histogram for matching pixel-CIELAB values of the already gamut-mapped image. For each ink combination matching a CIELAB pixel value for the primary illuminant, the corresponding CIELAB value for the second illuminant was calculated using the forward printer model and compared with the corresponding pixel CIELAB value for the second illuminant using ∆E00. The ink combination with the smallest difference was used for the separation. The whole separation process required ~5 min for a 22-megapixel image on an Intel Q6600 quad-core processor using a performance optimized C++ implementation. Results and Discussion The spectral reflectance factors of fifteen positions, selected as representative colors, were measured using an X-Rite i1 on the painting and each print. The metameric mismatch gamuts for illuminant A (secondary illuminant) are plotted in Figure 3 for seven of the colors that did not overlap in the a*-b* projection. For some colors (e.g., blue and greenish yellow), the color of the painting was not within the mismatch gamut; thus it was not possible to produce a reproduction matching the painting except under a single reference condition. The spectral properties of the printer did not span the acrylic paints, in particular, ultramarine blue, the dominant paint used in the sky. This is shown in Figure 4 (position 8); the prints had different spectral characteristics than the painting. Although the printer’s blue ink has a long wavelength reflectance tail, it did not coincide with ultramarine and the spectral differences beyond 600 nm are striking. Another example is the greenish yellow used between the split in the road (position 14). Hansa yellow medium, with a transition wavelength above 500 nm, was redder than the printer yellow, having a transition wavelength near 480 nm. Consequently, the print was metameric. An example of good performance is shown in Figure 6 for a grayish green color (position 6). For each measurement position, the print spectra exhibited similar shape with appreciable variation. The color separation algorithm resulted in the same set of inks for each color, but a range of ink amounts. This was caused by changes in the primary and secondary illumininants, changes in the observer, and measurement uncertainty caused by positioning the spectrophotometer. Figure 3. Metameric mismatch gamuts for seven positions on the painting (marked by arrow tips): illuminant A and the 1931 standard observer. The printer’s gamut boundary is shown and the inner colored regions indicate where a match of CIEDE2000<2.0 is possible and the dark regions in their centers <1.0 for the colors constrained to match under illuminant D65. CGIV 2008 and MCS’08 Final Program and Proceedings 485 Figure 4. Spectral reflectance factor at measurement position 8 for painting (original) containing appreciable ultramarine blue and four prints. Figure 5. Spectral reflectance factor at measurement position 14 for painting (original) containing appreciable hansa yellow medium and four