Color Conversion for Desktop Scanners

Transcription

1 Conversion for Desktop Scanners Jon Y. Hardeberg Conexant Systems Inc., Redmond, Washington, USA 1 Introduction Why do we need color? Digital color imaging systems process electronic information from various sources; images may come from the Internet, a remote sensing device, a digital camera, a desktop scanner, etc. Employing the ever increasing available computing power, a vast variety of image processing algorithms might be applied to the images. Examples of such algorithms include image compression, object recognition, segmentation, sharpening, noise reduction, and red eye removal [1]. After processing, an image may be transmitted to one or several destinations via a computer network for further processing and rendering on a computer monitor or a printer (Figure 1). The different image I/O devices generally process color information differently, for example if you scan the same photo with two different desktop scanners, you will get different digital images, or if you print a digital image on two different printer, the colors of the resulting prints will be different. Using more scientific terms: every imaging device quantifies color information using its own device-dependent color space. Figure 1: Today s color imaging systems are typically very complex. How can we achieve consistent communication of color? To achieve color consistency throughout such a widely distributed system, it is necessary to understand and control the way in which the different devices in the entire color imaging chain treat colors. This can be achieved by performing colorimetric characterizations of the color image acquisition and reproduction devices so that the device-dependent color spaces of the scanner, the monitor, the printer, and other color imaging devices, can be linked to a device-independent color space ( Connection Space), see Figure 2. In an imaging system with N color image input devices and M color image output devices the number of required color paths is reduced from NM to N + M. This is the basic principle of color management. For more information avout color management, refer for example to [2-4] or the books by Giorgianni and Madden [5] and Kang [6]. Up to this point, everyone in the color imaging science and ering community would probably 1

2 agree. The controversy comes about, however, when it comes to choosing a device-independent color space, and how, when, and where the color s shall be done. We will in this paper contribute to this discussion by presenting some of the considerations, recommendations, and design choices we have made concerning color s. Some of this work has been done in the context of the design of Conexant s controller solutions for multifunction peripherals (MFPs). First, in Section 2, we present three different color imaging architectures. The choice of a color architecture determines which device-independent color space to use, and when to perform the color s. Then, in Section 3, we discuss how the color should be done, that is, which computational method is advantageous for our applications. In Section 4, we address the question of where to perform the color, in particular, how the color from the scanner s RGB space into srgb can be integrated in the imaging pipeline. Finally, in Section 5 we discuss the definition of the color for a given scanner, by means of a method of colorimetric scanner characterization. Image Device-Independent Space Scanner Camera Monitor Printer Figure 2: The principle of Management. Each imaging device is characterized by a profile. information is communicated using a Device-Independent Space. 2 imaging architecture In this section we discuss the device-independent color space that When to convert? should be used for image exchange, and when the color s should be done. We present and discuss advantages and To which colorspace? disadvantages of three different color architectures, based on ICC color management, color facsimile standards, and the srgb color space. Our analysis may apply to several types of digital imaging systems, but we wish to emphasize that this study was done in the context of designing PC software solutions for multifunction peripherals (MFPs) for the consumer market. Notice also that even if in our presentation we are most concerned with the scanner and the printer, the color s needed for other imaging devices such as digital cameras and displays should also be considered. 2.1 management architecture ICC The International Consortium (ICC) was established in 1993 by eight industry vendors for the purpose of creating, promoting and encouraging the standardization and evolution of an open, vendor-neutral, cross-platform Management System architecture and components. Its standard for color management [7] has today wide acceptance, even if some issues still are unsolved.

3 In an ICC color management architecture, a scanner is characterized by its ICC profile. The ICC profile defines the relationship between the device-dependent scanner RGB color space and the Connection Space (PCS), which is typically either the CIEXYZ or CIELAB color space [8]. Similarily, the relationship between the PCS and the printer s device-dependent color space (typically CMYK) is defined by an ICC profile. The scanned image is communicated in the scanner s device-dependent RGB space. The color occurs when the image is rendered (typically on a monitor or a printer) by coupling the ICC profile of the scanner and that of the output device (see Figure 3). This implies that the scanner profile must be associated with each image, typically by embedding it in the image file. This requires using an image file format that Scanner RGB image Scanner ICC profile defining the color from Scanner RGB to XYZ Printer ICC profile defining the color from XYZ to Printer CMY(K) Figure 3: ICC color management architecture. allows embedded profiles, such as for example TIFF. Depending on the size and resolution of the image, and the type of profile, the size overhead added by this may or may not be significant. One advantage with this solution is that no intermediate color space is used directly. The color is done directly between the color spaces of the input and the output devices, thus minimizing the loss of information involved in any color with fixed-point arithmetics. An important drawback with this solution is its complexity. Even if the underlying principles are reasonably well understood, it is clear that many potential users of ICC color management systems have been discouraged by a rather complicated user experience. The success of this architecture in ensuring color consistency is dependent on whether the user has properly installed and enabled a color management system, such as ICM [3] for Microsoft Windows or Sync [4] for Macintosh. L*a*b* image Scanner profile defining the color from Scanner RGB to L*a*b* Printer profile defining the color from L*a*b* to Printer CMY(K) Figure 4: facsimile architecture using CIELAB color space for image interchange.

4 2.2 facsimile architecture CIELAB In the international standard for color facsimile ITU-T.42 it is specified that images shall be transmitted in the CIELAB color space, using the JPEG compression scheme. CIELAB is a widely used device-independent color space with many interesting features such as pseudo-uniformity with regards to human perception [8]. At the time of scanning/transmitting, the document colors must be converted from Scanner-RGB to CIELAB, and, at the receiving/printing end, the colors must be converted from CIELAB to CMYK (see Figure 4). It is expected that color facsimile functionalities will be increasingly present in imaging devices, in particular in multifunction peripherals (MFPs). It is therefore natural to consider using this architecture also for the communication between the device and the computer, and between computers, since the required algorithms might already be implemented for color facsimile. The major disadvantage with this architecture for consumer imaging is that very few PC imaging software applications are currently accepting images in this format. Another problem is that several encoding standards for CIELAB images exist: the one used for ITU color facsimile is different from the one used in TIFF. Also, CIELAB images cannot be viewed directly and rendering requires compute-intensive color. 2.3 Internet color architecture srgb This color architecture is very much similar to the previous. The only difference is that it uses the srgb color space instead of CIELAB for image interchange (see Figure 5). The srgb color space was introduced in 1996 as a standard color space for image interchange, especially over the Internet [9]. Its definition is based on the average performance of typical PC monitors under reference viewing conditions. Despite some controversy, it has now been widely accepted in the consumer imaging industry. It is used by key industry players such as Microsoft, Hewlett-Packard, and Adobe, and has been adopted as an international standard (IEC ). srgb image Scanner profile defining the color from Scanner RGB to srgb Printer profile defining the color from srgb to Printer CMY(K) Figure 5: Internet color imaging architecture using srgb color space for image interchange.

5 2.4 Discussion For the products we have considered in this project, that is, low-cost consumer imaging devices, not high-end graphic arts devices, we have chosen to use the srgb color space for image interchange. There are two main reasons for this choice. The first the increasingly widespread use of srgb in the consumer imaging industry, and the second and most important is the simplicity of the resulting imaging system. It is not necessary to manage device profiles associated with the images. And as more and more devices are using srgb as their way of describing colors, the users of digital imaging devices will be able to easily achieve unambigous consistent colors in their images. One of the most important uses of digital images for a typical user is to display it on a computer monitor. Since the srgb color space is based on the performance of an average PC monitor, images can be displayed directly on the monitor without further processing. It is not necessary to use ICC profiles when rendering the image, and thus the image can be processed by standard image editing applications. The ICC and CIELAB approaches both may require significant computation to display an image on a monitor. Furthermore, the use of srgb does not limit the possibility of processing the image in an ICC color management system framework. This can be done easily by means of the srgb ICC profile [9]. For example in Microsoft ICM 2.0 [3], if no source profile is associated with an image, it is assumed to be in srgb. There are also disadvantages with this approach. The srgb color space has a reduced color gamut compared to CIELAB. Some of the colors of a typical color printer fall outside of the srgb color gamut. It may be necessary to apply gamut mapping techniques [10]. However, we have not found this to be a major problem for the class of devices we consider. As a concluding remark concerning the choice of color imaging architecture we quote from http: // The goal of srgb is to develop an 80% solution that puts a single stake in the ground on a single recommendation that solves most of the color communication problems for office, home and web users. Using srgb is not an attempt to solve all color imaging problems. For example for high-end graphic arts systems the ICC solution may be preferrable. 3 Choice of color algorithm Now that we have decided to use the srgb color space for image interchange, we must define efficient algorithms to convert between the different color spaces. Our assessment of this problem How to convert? is described in this section. A general color between two color spaces can be described by the equation O = f (P ), wherep and O denote the input and output color signal, respectively. For our application, P will typically be the raw scanner RGB values, P =[R scan ;G scan ;B scan ],ando =[R srgb ;G srgb ;B srgb ], the calibrated srgb values. One possibility for performing this transformation is to implement directly the analytical function f ( ) using floating point arithmetic. As we see in Section 5 no exact analytical function exist for the from scanner RGB to srgb. Then another possibility would be to use the nth order polynomials described in Section 5. However, for images over a certain size, such floating-point calculations tend to be prohibitively slow.

6 Another possibility would be to pre-calculate the transformation for all possible input values and store the results in a 3D look-up table (LUT). Then, for image, the transformation could be found almost immediately, simply by indexing into this table. Unfortunately this solution requires a very large amount of memory approximately 50 Mb for 24 bit per pixel input and output data, and it is thus not currently of practical interest. We wish to gain speed, storage space and flexibility by using an intermediate approach, using a reduced-size 3D LUT and interpolating between the missing values. The speed gain is achieved by a reduced number of operations, as compared to calculating the analytical function. The storage space gain is evident when compared to a conventional LUT. The flexibility stems from the fact that the algorithm is able to perform an arbitrary color from one three-dimensional space to another; the performed color depends uniquely on the choice of output values in the LUT. Several different schemes exist for this interpolation, such as: trilinear, tetrahedral, prism, and slant prism. One important design choice for this algorithm is the size of the LUT. Increasing the size gives a more dense interpolation grid, and thus more accurate results, at the expense of memory use. We have performed a comparative analysis of advantages and disadvantages of different algorithms for performing color [11]. For our applications we considered flexibility to be a very important requirement, and thus we recommended using interpolation in 3D LUTs for color. If a different color space such as CIELAB should be used for image interchange, it would suffice to use another LUT file, no architectural modification would be required. 4 Implementation Where to convert? Having decided to convert to and from srgb by means of interpolation in 3D LUTs we must now decide where in the imaging system these s shall be done. Several possibilities exist, as illustrated in Figure 6. Depending on the specific requirements of a particular device or system, one or the other of these options may be advantageous. One interesting solution which allows seamless integration of the color from Scanner RGB to srgb in any TWAIN-capable imaging application, the TWAIN Filter, was proposed in [11]. 5 Scanner Characterization It is often suggested that scanned color images can be considered How to define to be in srgb space if an appropriate gamma correction has been the? applied. We have seen that doing this introduces significant colorimetric errors [11]. The same image scanned on different scanners gives considerably different digital images, with color values far from being srgb values. In this section we describe our approach to fixing this problem by defining the relationship between the scanner s device-dependent RGB color space and the device-independent srgb color space. That is, we seek to define the transformation [R srgb ;G srgb ;B srgb ]=f (R scan ;G scan ;B scan ): Unless the scanner is colorimetric, that is, the spectral sensitivities of the three scanner channels

7 Storage, Distribution, etc. Imaging Application Perform colorspace in the application? Perform the in the application? Imaging Application TWAIN Filter Intercept the TWAIN data flow to convert to srgb? Let the system s CMS perform from srgb to Printer CMY(K)? ICM 2.0 PC TWAIN Data Source Perform to srgb in the scanner driver? Perform from srgb to Printer CMY(K) in the printer driver? Printer Driver Device Embedded Image Processing Convert from Scanner RGB to srgb in the device? Perform in the device? Embedded Image Processing Scan Sensor Develop new sensors that produce srgb directly? Printing Engine Figure 6: Different options concerning where to convert between the device-dependent color spaces and the srgb color space. equal the CIEXYZ color matching functions [8] or any nonsingular linear transformation of them, an exact analytical representation of this equation does not exist. Very few, if any, commercial desktop scanners can be classified as being colorimetric. We must thus try to approximate the function f ( ). The common approach to this problem is to scan a standard color target, extract the mean RGB values of the color patches, and apply an optimization scheme to these values and the measured CIELAB color values of the patches (see Figure 7). This procedure is known as scanner profiling or colorimetric scanner characterization. Scanner CMS Colour Chart R G B B Spectrophotometric Data Scanner profile RGB-to-L*a*b* 3D look-up tables R G L* a* b* Characterisation Process Polynomial Coefficients Figure 7: Scanner characterization. 3D Polynomial Transformations Now, the quality of the resulting color depends very much on the optimization scheme that is used. In the literature, the most common solution to this problem is to apply linear regression algorithms to convert from Scanner-RGB values to CIEXYZ values. Then the to any color space based on CIE colorimetry, in particular the srgb space, can be calculated analytically. The main drawback with such methods is that the error that is minimized by the regression algorithm, the RMS error in CIEXYZ space, is very poorly correlated to visual color differences. We remedy this by making sure that the output values of the regression algorithm are CIELAB values,

8 instead of CIEXYZ values, since the Euclidean distance in CIELAB space, corresponds well to perceptual color differences. E Λ ab = p L Λ2 + a Λ2 + b Λ2 There is clearly not a linear relationship between Scanner-RGB and CIELAB space, and we propose thus to model the transformation f ( ) given above by nth order polynomials whose coefficients may be optimized by standard regression techniques. We have evaluated the proposed method on several different desktop scanners [11, 12]. The results are very good. When evaluating the differences between the ideal and the obtained srgb values of an IT8.7/2 color target, we obtain mean CIELAB E Λ ab color differences as low as 1.4. For more information about this method and its applications, refer to [11-16], available from http: //color.hardeberg.com/. 6 Conclusion In this study we have investigated the problem of achieving high color image quality in a digital imaging system, with an emphasis on image acquisition by a desktop scanner. We have proposed methods for scanning to srgb, that is, converting from the scanner s device-dependent RGB color space to the device-independent color space srgb. Several aspects of this problem have been covered. We have defined the by a process of colorimetric scanner characterization. We have studied the problem of designing effective algorithms to perform the color, and we have proposed methods for integrating the color in the imaging workflow. We have also discussed advantages and disadvantages of a digital color imaging system using the srgb space for image exchange, as compared to using other color imaging architectures. By applying the principles and algorithms presented in this paper when designing consumer imaging devices and systems, one step towards unambiguous communication of color image information will be taken. References [1] Jon Yngve Hardeberg. Red-eye removal using digital color image processing. In IS&T s PICS 2001 conference, Montreal, Quebec, Canada, April [2] Jon Yngve Hardeberg. management: Principles and solutions. NORSIGnalet, the quarterly magazine of the Norwegian Signal Processing Society, (3), Available from [3] Microsoft Corporation. Introduction to Management in Microsoft Windows. Available at [4] Brian P. Lawler. Sync technology brief: The color management workflow standard. Apple White paper, available at colorsyncwhitepaper.pdf, 1999.

9 [5] Edward J. Giorgianni and Thomas E. Madden. Digital Management: Encoding Solutions. Addison-Wesley, [6] Henry R. Kang. Technology for Electronic Imaging Devices. SPIE Optical Engineering Press, [7] File format for color profiles. The International Consortium, Version ICC.1: , See [8] Günter Wyszecki and W. S. Stiles. Science: Concepts and Methods, Quantitative Data and Formulae. John Wiley & Sons, New York, 2 edition, [9] Gary Starkweather. space interchange using srgb. This White Paper and other information about the srgb color space is available at [10] Ján Morovic. To Develop a Universal Gamut Mapping Algorithm. PhD thesis, Colour & Imaging Institute, University of Derby, October [11] Jon Yngve Hardeberg. Desktop scanning to srgb. In IS&T and SPIE s Device Independent, Hardcopy and Graphic Arts V, volume 3963 of SPIE Proceedings, pages 47 57, San Jose, CA, January [12] Jon Yngve Hardeberg. Acquisition and reproduction of colour images: colorimetric and multispectral approaches. Ph.D dissertation, École Nationale Supérieure des Télécommunications, Paris, France, [13] Jon Yngve Hardeberg, Francis Schmitt, Ingeborg Tastl, Hans Brettel, and Jean-Pierre Crettez. management for color facsimile. In Proceedings of IS&T and SID s 4th Imaging Conference: Science, Systems and Applications, pages , Scottsdale, Arizona, November Also in R. Buckley, ed., Recent Progress in Management and Communications, IS&T, pages , [14] Jon Yngve Hardeberg. Transformations and colour consistency for the colour facsimile. Diploma thesis, The Norwegian Institute of Technology (NTH), Trondheim, Norway, April [15] Jon Yngve Hardeberg and Jean-Pierre Crettez. Computer aided colorimetric analysis of fine art paintings. In Oslo International Colour Conference, Colour between Art and Science, Oslo, Norway, October [16] Zhihong Pan, Ying X. Noyes, Jon Y. Hardeberg, Lawrence Lee, and Glenn Healey. scanner characterization with scan targets of different media types and printing mechanisms. In IS&T and SPIE s Device-Independent, Hardcopy, and Graphic Arts VI, San Jose, California, January Jon Y. Hardeberg received his sivilingeniør (M.Sc.) degree in 1995 from the Norwegian Institute of Technology, Trondheim, Norway. He finished his Ph.D. studies at the Ecole Nationale Superieure des Telecommunications, Paris, France in February His Ph.D. research concerned color image acquisition and reproduction, with applications in facsimile, fine-art paintings, and multi-spectral imaging. He is currently with Conexant Systems, Inc. where he architects and develops color image processing solutions for multifunction peripheral devices and systems. For further information about the research presented in this paper, the author can be contacted by at jon. hardeberg@conexant.com or jon@hardeberg.com. Further related information can be found at conexant.com/ and