Munsell Color Science Laboratory Technical Report. Direct Digital Imaging of Vincent van Gogh s Self-Portrait A Personal View

Transcription

1 Munsell Color Science Laboratory Technical Report Direct Digital Imaging of Vincent van Gogh s Self-Portrait A Personal View Roy S. Berns berns@cis.rit.edu May, 2000 A Note About This Document in Terms of Color Management The various digital images of van Gogh are shown at the end of this document to make visual comparisons easier. The images will look best using a CRT display with its white point set to between 5000 and 6500 K. The choice of gamma is not critical though the images have a gamma of 1.8. Abstract The ability to accurately record the color of a painting within an American museum is explored. Limitations in imaging techniques and hardware are current barriers to high quality. Interestingly, simple colormanagement techniques provide large improvement. Perhaps the greatest advantage is a decrease in visual adjustment. A review is presented of the author s experiences in trying to digitally record the color of Vincent van Gogh s Self-Portrait. 1

2 Introduction Beginning September 1, 1999 and through May, 2000, I was in residence at the National Gallery of Art, Washington, DC as a Visiting Senior Fellow in Art Conservation Science while on sabbatical. I have been involved in colorant selection for restorative inpainting using instrumental color-matching techniques, the selection criterion that of minimizing metamerism; characterizing the optical properties of varnishes using image analysis; and a proof-of-concept for spectral-based color reproduction including capture and printed output. All of these projects apply to paintings. During my first weeks, Vincent van Gogh s Roses was under study in the painting conservation laboratory (see van Gogh was known to have used a red-lake pigment with extremely poor lightfastness. In its current condition, the roses are white. Following a microscopic evaluation and the removal of a very small amount of green paint covering the edge of a rose, a vibrant red color was revealed. By placing chips of the Munsell Book of Color under the microscope adjacent to the red area, its color was quantified: 1.25R 5/14. This is a very high-chroma color. Similar fading has been reported in other van Gogh paintings. 1 Another van Gogh in the Gallery s collection, Self-Portrait, shown in Fig. 1 [1889, oil on canvas, x m (22 1/2 x 17 1/4 in.), collection of Mr. and Mrs. John Hay Whitney, see had also undergone a large color change. One of the intriguing aspects of this painting is that its edges have been protected behind the frame s rabbet; a close-up of an edge is shown in Fig. 2. The blue background was probably a purplish blue; the red-lake pigment used by van Gogh has faded. Figure 2. Detail of Vincent van Gogh, Self-Portrait. Looking from top to bottom: canvas, brown paper tape, purple area, blue area. Using a GretagMacbeth Spectroeye reflection spectrophotometer, measurements were made at an edge and an adjacent interior position, plotted in Fig. 3. The red lake absorbs the short-wavelength peak of cobalt blue (identified using X-ray fluorescence and visible reflectance spectroscopy). If the painting could be imaged in a way to facilitate spectral reconstruction, it would be interesting to add back red to various parts of the painting using Kubelka-Munk theory. Having a spectral image would also enable a printed reproduction to be produced with minimal metamerism. 2

3 0.4 Reflectance factor Blue background Purple behind rabbit Wavelength (nm) Figure 3. Vincent van Gogh, Self-Portrait spectral reflectance factor of edge covered by frame rabbit and adjacent location not covered. Thus I embarked on a quest to digitize the van Gogh Self-Portrait. In the process of imaging this painting, the ubiquitous GretagMacbeth Color Checker Color Rendition chart, and a Kodak gray scale, as well as observing imaging practices, it became clear that at the Gallery, digital imaging, and in particular, direct digital imaging of artwork, is still in its childhood. Visits to other museums, libraries, and archives indicated a similar level of development. Of course, my expertise in imaging artwork was completely theoretical, limited to the GretagMacbeth ColorChecker Color Rendition Chart, gray scales, and test samples and paintings without texture, impasto, or gloss. The purpose of this paper is to share my experiences and implore the American color and imaging science community to become more active in the imaging of our cultural heritage. Available Imaging Systems At the Gallery, there are two digital imaging systems. The first is a Dicomed scanning back coupled with a Mamiya 4 x 5 conventional camera. It uses tungsten-halogen illumination and polarizing optics to reduce specular reflection. This scanner is capable of capturing 6000 x 8000 pixels. The spectral sensitivities of the system are plotted in Fig. 4 along with the human visual system s L, M, and S cone fundamentals. The CCD sensor is designed for flat-bed scanners. As a consequence, it s spectral sensitivities are densitometric rather than colorimetric. Clearly, this will lead to large color errors when imaging artwork. 3

4 Relative sensitivity Wavelength, nm Figure 4. Spectral sensitivities of a typical scanback (solid lines), normalized to equal area. These sensitivities include the CCD spectral sensitivity, filter transmittances, and infrared radiation blocking filter. The human visual system s spectral sensitivities are also plotted (dashed lines). The Dicomed system was built in order to digitize the Gallery s Stieglitz photographic collection with the ultimate goal of creating a catalog and website (see and A Kodak gray scale and color-separation guide are imaged along with each work of art. The camera software is used for gray balance and tone reproduction based on visual adjustment and several aim values for the gray scale. Watching the system being used, it became clear that the importance of fully utilizing the camera s 12-bit dynamic range was not recognized. As long as the scan time did not lead to blooming, the minimum and maximum 8 bit digital counts were set using the curves adjustment, essentially defining the transfer function from 12 to 8 bits per channel. A computational analysis similar to Burns 2 was performed in which L* errors caused by a change in one digital count were determined for a Kodak gray scale imaged at various bit levels. The results are shown in Table I. Errors increase with decreasing lightness. It is quite common to see image archives with high image noise and blocking in shadow areas caused by incorrect integration times. The general belief in the American museum community is that all deficiencies in image capture can be corrected in Photoshop. Once the image has been captured, tone and color correction are adjusted visually in Adobe Photoshop. In this system the CRT is set up to 5000 K using software resident with the operating system. (There seems to be more confidence in a visual approach than instrumentation.) Since this adjustment affects the video look-up tables, the peak luminance of the monitor is reduced. The operator will look at the original art and adjust the tone and color balance of the CRT image to match, in general, the original object. A light booth with fluorescent 5000 K daylight is used, although at considerably higher luminance than the CRT display. However, differences in viewing geometry, difficulties in properly setting a CRT s black point, and optical flare during image capture result in unmatched black levels between illuminated objects and their CRT representations. The operator may boost contrast to compensate for these differences. The second system is an IBM Pro It consists of a monochrome scanning back and filter wheel. It uses tungsten-halogen illumination and can digitize both reflective and transmissive materials. It can capture 3000 x 4000 pixels. Its spectral sensitivities are plotted in Fig. 5. This system was optimized to have nearcolorimetric properties. We have an identical system at RIT and have reported its performance previously

5 Table I. The effects of quantization levels on lightness errors, L*, for a Kodak gray scale. Kodak gray scale L* L* for 16 bits L* for 12 bits L* for 10 bits L* for 8 bits B M A average maximum

6 Relative sensitivity Wavelength, nm Figure 5. Spectral sensitivities of a monochrome scanback with three optimized color filters (solid lines), normalized to equal area. These sensitivities include the CCD spectral sensitivity, filter transmittances, and infrared radiation blocking filter. The human visual system s spectral sensitivities are also plotted (dashed lines). Most recently, a Nikon D-1 digital camera was purchased by the Gallery. It is a typical professionalquality digital camera employing a two-dimensional color CCD array. Its spectral sensitivities are unknown but are presumed to lie between densitometric and colorimetric. The fourth system consists of a 8 x10 view camera, conventional positive transparency film processed by E-6, digitization using either a flat-bed or drum scanner, and image editing to match the photograph. This fourth system represents the most common digital imaging at the Gallery. One advantage of this system is that the positive, in most cases, will reasonably match the original art. The Gallery photographers will use various lighting, exposure, and color filtration, iterating until a match is achieved between the transparency and the object, both illuminated by 5000K fluorescent daylight. Ideally, the painting would be imaged using all of these systems and perhaps this may yet occur. However, each system is owned by a separate department and coupled with the popularity of this painting, coordination has been difficult. During October, 1999, the painting was imaged using the Dicomed system and measured spectrophotometrically in a number of positions using a GretagMacbeth SpectroEye. During March, 2000, the painting was imaged using the Nikon D-1 using available Gallery lighting. More analyses using all four systems are in progress. Image Comparison The van Gogh Self-Portrait serves as a useful case study. In creating the truth, the 8 x10 positive transparency, the photographer was unable to find a combination of film type, lighting, and colorcompensating filtration to achieve correct color for the jacket and background, simultaneously. The transparency was balanced to achieve reasonable color accuracy for the foreground. The transparency was digitized using a flat-bed scanner, shown in Fig. 6. The background was purplish rather than a dark blue. Spectral measurements of blue areas of the jacket and the background show the cause of this discrepancy. The jacket was painted with a combination of emerald green and cobalt blue while the background was painted with a combination of cobalt blue and an earth pigment, possibly yellow ochre. The jacket s spectrum has low reflectance towards the long-wavelength end of the visible spectrum while the background has a large reflectance tail, typical of cobalt blue. In order to color balance the foreground, there was excessive red exposure in the background areas. Thus, it was impossible to achieve truth through conventional photography. 6

7 The scanned transparency was color corrected using Photoshop masks to correct mainly the background, and secondarily the face and jacket. This is shown in Fig. 1. The image was output to 8 x10 positive film resulting in the official photographic image of Self-Portrait, used to produce various printed publications The digital manipulations were performed having the painting in the same room as the computer system. The official photograph was digitized for the Gallery s web site using the IBM Pro 3000 system. The 24-bit TIFF files are approximately linearized with respect to CIE L* but not color corrected. Color correction is defined in the header and only used when displaying the images on the IBM system. When transferred to other computer platforms, considerable color correction is required. This is accomplished visually using a set of predefined Photoshop actions and additional editing. The Self-Portrait web image is shown in Fig. 7. There is a large boost in chroma and contrast compared with Fig. 1. The Dicomed image is shown in Fig. 8. When viewed on the system s CRT display, the color seemed reasonable, particularly following gray balance and visually adjusting the tone-reproduction curves. However, in comparison to the official digital image, shown in Fig. 1, there is quite a large difference in color. Similar to the photographic process, the background has become purplish. The Nikon D-1 image is shown in Fig. 9. It also resulted in large color errors. Color Management All of the images shown are a result of visual adjustment. An obvious question is whether simple color management can provide significant improvement over visual techniques and overcome limitations in spectral sensitivity. A ColorChecker and Kodak gray scale were imaged using both cameras. The gray scale was used to characterize the photometric response of the system and the ColorChecker was used to derive a linear transformation to convert from linearized R, G, and B to CIE tristimulus values. Nonlinear optimization was used to derive the matrix in which the objective function was minimizing CIE94 color differences, described in more detail in references 8 and 9. The results are shown in Table II. For comparison, a similar analysis was performed for the IBM Pro 3000 and Marc II (manufactured by the Lenz brothers and based on a Sony ICX085AK sensor and Schott BG39 infrared cut-off filter). The errors correlate with each camera s spectral sensitivities: The closer a camera s spectral sensitivities are to being colorimetric, the better the color accuracy. Table II. Optimization results. System Gray scale ColorChecker E* 94 E* 94 E* ab Average Maximum Average Maximum Average Maximum IBM Pro Marc II Nikon D Dicomed Obviously, these results are not independent validation, rather they indicate modeling performance. Because it is difficult to determine the accuracy of estimating the colorimetry of the van Gogh except by visual means, a target was produced with spectral properties similar to the painting. Essentially, direct spectrophotometric measurements were made in different colored areas of the painting such as the background, greens and blues of the jacket, reds and browns of the beard, yellows and greens of the hair, flesh tones, and the palette colors. From a database of retouching paints created for other research at the Gallery, a set of modern colorants were selected that well reproduced the spectral properties of the painting: cobalt blue, prussian blue, yellow ochre, cadmium red, naples yellow light, ivory black, and titanium white. (It would be unwise to use the identical pigments used by van Gogh; emerald green is arsenic based, for example!) Specific combinations of colorants were used to match specific image areas. The target is shown in Fig

8 Figure 10. Painted target with similar spectral properties to Vincent van Gogh Self-Portrait. (Samples B13 B17 are the chromatic colorants mixed with titanium white.) The test target was imaged using the Nikon D-1. The average digital counts of each patch was used to estimate colorimetric coordinates using the transforms derived from the Kodak gray scale and ColorChecker. The average and maximum CIE94 values were 3.7 and 11.2, respectively. A color-difference histogram is shown in Fig Frequency <0 [0,1) [1,2) [2,3) [3,4) [4,5) [5,6) [6,7) [7,8) [8,9) [9,10) [10,11) [11,12) 12 E*94 Figure 11. CIE94 histogram comparing measured and estimated colorimetric coordinates for test target shown in Fig. 11 for illuminant D65 and the 1931 standard observer. The empirical transformation and a CRT profile for a typical Macintosh system were concatenated. The concatenated profile was applied to the image shown in Fig. 9, resulting in Fig. 12. Although there are still differences, the improvement is striking. In particular, the purple background is largely eliminated and the face reproduction is much improved. 8

9 Conclusions The principles of color and imaging science have yet to permeate the American museum and library communities. Many imaging professionals, although highly skilled in conventional photography, graphic design, or computer science, are less knowledgeable in color digital imaging. They are being asked to make important decisions regarding imaging hardware and software for creating digital archives of our cultural heritage. It is clear that by using color standards in addition to gray scales, it is possible to significantly improve color accuracy while simultaneously decreasing the reliance on visual adjustment. For truly high accuracy, it will be necessary to design imaging hardware that can produce high spatial image quality, high colorimetric accuracy, and high spatial resolution. The software and operating systems must be compatible with current museum professionals. Ultimately, we will be best served by an influx of imaging scientists and engineers to the art world. Acknowledgements This publication was written while the author was a Senior Fellow in Conservation Science at the National Gallery of Art, Washington; the Gallery s financial support is sincerely appreciated. References 1. A. Distel and S. A. Stein, Cézanne to Van Gogh: The Collection of Doctor Gachet, The Metropolitan Museum of Art, New York, P. D. Burns and R. S. Berns, Quantization in multispectral color image acquisition, Proc. 7 th IS&T/SID Color Imaging Conference, (1999). 3. R. S. Berns, F. H. Imai, P. D. Burns and Di-Y. Tzeng, Multi-spectral-based color reproduction research at the Munsell Color Science Laboratory, Proc. of SPIE 3409, (1998). 4. F. H. Imai and R. S. Berns, High-resolution multi-spectral image archives: A hybrid approach, Proc. IS&T/SID Sixth Color Imaging Conference, (1998). 5. F. H. Imai and R. S. Berns, A comparative analysis of spectral reflectance reconstruction in various spaces using a trichromatic camera system, Proc. IS&T/SID Seventh Color Imaging Conference, (1999). 6. F. H. Imai and R. S. Berns, Spectral estimation using trichromatic digital cameras, Proc. Intl. Symp. Multispectral Imaging and Color Reproduction for Digital Archives, Chiba University, Japan, (1999). 7. F. H. Imai, R. S. Berns, and Di_Y. Tzeng, A comparative analysis of spectral reflectance estimated in various spaces using a trichromatic camera system, J. Imaging Sci. Tech. 44, (2000). 8. R. S. Berns, Billmeyer and Saltzman s Principles of Color Technology, 3 rd ed., John Wiley & Sons, New York, R. S. Berns, The science of digitizing two-dimensional works of art for color-accurate image archives concepts through practice, Munsell Color Science Laboratory Technical Report, (2000). See 9

10 Figure 1. Vincent van Gogh, Self-Portrait, [1889, oil on canvas, x m (22 1/2 x 17 1/4 in.), collection of Mr. and Mrs. John Hay Whitney]. This image was digitized using a flat-bed scanner from a 8 x10 positive photographic transparency. Adobe Photoshop was used to correct the color of the background, jacket, and portions of the face to improve the color match in comparison to the painting illuminated by fluorescent daylight. Figure 6. Vincent van Gogh, Self-Portrait. This image was digitized using a flat-bed scanner from a 8 x10 positive photographic transparency. Figure 7. Vincent van Gogh, Self-Portrait. This image was digitized using an IBM Pro 3000 from a 8 x10 positive photographic transparency. The digital image was adjusted visually to match the photographic transparency. 10

11 Figure 8. Vincent van Gogh, Self-Portrait. This image was created using a Dicomed scan back, a Mamiya 4 x5 camera and lens, and tungstenhalogen illumination. Dicomed software was used to gray balance the image and adjust tone reproduction. Figure 9. Vincent van Gogh, Self-Portrait. This image was created using a Nikon D-1 digital camera, and tungsten illumination. Figure 12. Vincent van Gogh, Self-Portrait. This image was created using a Nikon D-1 digital camera, and tungsten illumination, followed by color correction as described in the text. 11