1. Introduction

Towards the end of the 1980s digital imaging was developing at a rapid pace. New devices, particularly CCDs, with higher resolution and quality were making it possible to think about replacing conventional photography with direct digital imaging. This paper summarises roughly a decade of work on developing high quality imaging systems especially relating to the imaging of fine art. The systems were specifically built for imaging paintings in museums and galleries, however the techniques used are widely applicable. Experience was drawn from a wide range of fields: optics, mechanical engineering, image processing, colour science, computer science and painting conservation.

Conventional photography of art captures fine detail; the resolution of large-format 10x8" film is very high (about one Gigabyte equivalent) but the colour accuracy and permanence is poor. Digital imaging promised higher quality and permanence but it was not clear how the high resolution would be obtained. Imaging devices were limited to around 1000x1000 pels for arrays or 4096x1 for linear sensors whereas resolutions of around 20,000x 20,000 were required.

 

1. The VASARI Project

In 1989 a group interested in exploring the limits of the emerging digital imaging started a European project called VASARI: Visual Arts System for Archiving and Retrieval of Images. The main aim was to produce an imaging system of sufficient quality to replace conventional photography of fine art. It was also hoped that the system could acquire spectral information directly from paintings, over the whole surface. Previously researchers had only been able to make spot measurements of colour using a spectrometer. The final results included two scanners and calibrated CIE Lab output rather than spectra. Several improvements have been made since then as well as an additional scanner in Florence.

It was determined early on at the design stage that the scanner should operate in a vertical plane rather than horizontally as in the inspirational system designed to scan the declaration of independence in Washington [ref!?]. This would help to prevent accidental damage to the paintings as well as maintain their normal position. Also rather than using precise robotics and a line imager, rectangular image areas would be acquired and a stitching technique developed to join them. This software stitching was also expected to compensate for other mechanical and optical distortions.

The resolution required by such a system was not clear at the start of VASARI. To replace 10x8" transparencies a calculation based on their resolving power (40 line pairs per mm gives around 20000 samples per 10") could be used to set a maximum resolution of around 20k x 20k. Studies of the sizes of cracks in paintings using X-radiographs seemed to confirm that the closer the study the more that could be seen: paintings are fractal in their nature. At one large meeting in Luxembourg a vote was even taken to decide on a resolution, with around 300dpi coming out a winner for no apparent reason. There was certainly a scepticism about the large file sizes envisaged in VASARI. 20k x 20k x 3 bytes was already over 1GB and the team were talking about multispectral 12 bit data. Fortunately we were consoled by the fact that storage size and processing power progresses at a fast rate. It was decided that the system should be capable of acquiring at least 12 pels/mm as well as 2-3 times this.

A high resolution camera was chosen: the Progres 2000/3000 which provides around 3k x 2k resolution. The current versions use 12 bits A/D conversion with a non-cooled CCD. In order to move the camera a known distance in pels the resolution in pels per millimeter is measured by moving the camera known distances and finding an edge on a target. The scanner can then move the camera to provide an array of patches overlapping by around 200 pels. The overlaps are correlated later to provide an exact join of the patches. [REF SPIE paper]

Many other calibration steps remove non-uniformities due to variations in dark-current and sensitivity of the CCD sensor. In the VASARI scanners, the lights move with the camera, providing a consistent illumination pattern. The illumination non-uniformities are removed using an image of a smooth white PTFE target. This also helps compensate for variations in sensor gain.

The colour accuracy was obtained by using more than three filters and it was hoped that colour errors could be as low as 1 CIELab D E. The conventional film imaging the system was expected to replace has average errors much higher than this [TOO WOOLY?] but the aim in VASARI was to achieve as low an error as possible under reasonable conditions. These included a low illumination (< 1000 Lux [David?]) and fast capture as paintings can not be left in a scanner for longer than an afternoon for example.

The filters were placed in the lighting system in order to avoid focussing problems if used in the camera. A system of fibre-optic light-guides was used to provide six adjustable light sources quite close to the painting. Another advantage of this approach was that painting was not comlpetely exposed to light all the time. ENST [spelling/accents?] carried out some simulations using spectra of known pigments and produced sets of idealised filters tuned to paintings. Studies were also made of the number of bits required per sample, concluding that 8 bits was insufficient, 10 bits was reasonable but 12 bits was ideal. Fewer bits/pel lead to quantisation steps larger than one just noticeable difference (or 1 D E).

Unfortunately the ideal filters calculated theoretically were not practical or affordable so narrow band gaussian filters were evaluated and chosen instead. Very narrow filters with low transmission were not used because of the necessity for more light. Medium widths of 70nm were finally selected at 50nm intervals so seven filters covered the 400-700nm wavelength range. Using many narrow band filters also made the system more capable of separating metameric colours.

[picture of vasari scanner]

[detail of light system? - David's drawing?]

[equations for calibration? Or refer to SPIE etc?]

[explanation of newer calibration - Ruven et al]

[picture of results - hi/low res]

2. Colour calibration

At the start of every capture session an image is taken of various calibration targets set up permanently on the easil. These include a smooth white used to correct the non-uniform illumination. A Macbeth ColorChecher Chart with precisely measured spectra is imaged to generate the calibration matrix. A small line is also imaged twice in different horizontal positions in order to measure the resolution in pels/mm.

The colour calibration principle is simple. With linearised camera data from the colour chart is used with the known colours to solve for the unknown factors in an equation:

cameravalues = unknownfactors . knowncolours

This involves inverting a matrix in such a way as to obtain the lowest least-mean-square errors and has been described in more detail elsewhere [SPIE1901]. The result is that colour chart colours are calibrated very well, while others are less precise. A more recent improvement minimises the errors in CIE Lab space afterwards by small adjustments, which spreads the errors more evenly. Once the calibration factors are known they are applied to all the camera images to obtain CIE XYZ colour values.

3. Vasari scanner Improvements

A much simpler GUI has recently been made for the VASARI scanner which makes it useable by a non-specialist. This was inspired by the simple interface made for the later MARC camera described below. Other developments increased the scanning speed over the years. Initially a PC with an AT bus card was used for the camera, with an Ethernet connection to the controlling SUN. The current system uses an SBUS interface card in the SUN, which is also considerably faster (16MHz 68020 to 50 MHz SPARC to 167MHz UltraSPARC). The disk system improvements also make a noticeable difference (12ms, 650MB, 3MB/s disks to 8ms, 28GB, 10MB/s disks) [JOHN?] due to the mass of data handled. Mass storage was a shelf of DATs read at 140 kB/s and is now an NSM CD-R jukebox.

4. VASARI Results

The image stitching was highly successful, with no visible joins and the illumination correction provided a seamless brightness across the images. The Progres camera is deliberately band-limited so its highest resolution looks like it requires sharpening to compensate. Recently the makers of the camera have provided a better filter and the results are better. The average colour error using a MacBeth chart was consistently around 1 CIELab D E. Suddenly there was the capability to carry out image-wide comparisons and colorimetric experiments which are reported later. An unexpected result was the popularity of the project with the funders in the European Community. [cheeky!] The scanner in the National Gallery has imaged around 100 paintings so far and these probably remain the highest accuracy collection of art images.[ DAVID] The VASARI image processing system VIPS [VIPSREF] was produced specifically to handle large images and is still in use today. It was subsequently used in all the follow-on projects.

5. MARC

In 1992 the MARC project (Methodology for Art Reproduction in Colour) concentrated on perfecting the path from accurate imaging to accurate printing: from the painting on the wall to the printed page. It also designed a new scan-back camera which was capable of high resolution while being more portable and faster. It was expected that the use of a more conventional RGB filter set would compromise colour accuracy slightly but be appropriate for most printing purposes as the errors introduced there are greater.

[DO WE SAY MUCH ON PRINTING? - DOES THE JOURNAL NEED IT?]

The new camera was designed and built by Udo and Reimar Lenz in the Universities in Munich while the camera software was developed in London by the authors. It used a similar masked CCD as used in their Progres cameras but also had a macro-scan in order to cover the image plane of a large format camera (Bronica). The CCD has 500x582 elements with small (6m m), widely separated (11x17m m) sensor sites. Piezo-transducers used for microscanning in the Progres cameras were dispensed with and the microscope translating stage stepper motors used for both macro and micro-scanning.

Each of the 7x9 patches has an overall gain characteristic due to the variations in angle of the incident light, especially noticeable at the edges of the image. To compensate for this a full-field calibration target is used. [NOT TRUE? - ISN't IT FUDGED OUT?]

A quantised CIE Lab standard was made for the final calibrated images in order to avoid using 16 bits for each component. Tests showed that the 8 bits per channel used commonly used can lead to contouring when printed. This led to the MARC project standard (LabQ) of 10 bits for L and 11 bits for each of a and b, producing an efficient 32 bit colour value for use in display and printing. The ten bits for L may seem low given the initial twelve bits per channel but L is more visually linear than RGB. Calibrated images are stored in a LabQ VIPS image file uncompressed, while the raw source data is stored losslessly compressed for reference.

6. MARC results

The camera has a resolution of 20k x 20k and can grab this image in around 20 minutes depending on the integration time used. The average D E is around 3 and the images look almost indistinguishable from those obtained by the VASARI scanners except that the lighting is more conventional.

[image of camera]

[do we need macro-micro scan pic???]

[mention Lenz bros new cameras? - new refs?]

 

7. Viseum: putting the images on the Web

In 1996 the Viseum project worked on a client-server system [WEBREF] to allow the large images produced in previous work to accessible over the Web. This used the Internet Imaging Protocol with extensions to allow a Java client in a normal Web browser to fetch image tiles (64x64 pels) in JPEG sRGB [REF?] format for display. A multiresolution TIFF file was used with full, half, quarter etc resolutions stored as tiled JPEG images. By only fetching the tiles necessary for the specific view required at any one time by the viewer the system was useable over the internet at low bandwidths. It was also possible for the C server to convert LabQ uncompressed images on the fly. A system was also tested with a monitor colour profile (ICC) server so that images could be transformed specifically for the user's monitor rather than to sRGB.

 

8. Conclusions

 

9. Acknowledgements

The MARC and VASARI projects were funded by the European Commission’s ESPRIT programme. VASARI involved: The Doerner Institute, Munich; Brameur Ltd. (UK); Birkbeck College London, Telecom Paris/ENST; Thomson-CSF LER, Rennes (F); TÜV-Bayern, Munich (D). The ESPRIT III project MARC involved: Thomson Broadband Systems, The National Gallery, Birkbeck College, The Doerner Institute Munich, Crosfield Ltd (UK), Hirmer Verlag (D) and Schwitter (Switzerland).

The new VASARI scanner GUI was written by Ruven P while at the National Gallery funded by HP-Labs.

Thanks to Barco for advice and monitor models for colour display as well as support for their Calibrator.

 

10. References

K. Martinez, J. Cupitt, and S. Perry. High resolution Colorimetric Image Browsing on the Web, Computer Networks and ISDN Systems, Vol. 30, pp 399-405, 1998.