Method of color correction using multi-level halftoning

[0002] The present invention is directed to the digital imaging arts. It finds particular application to a method of color correction using multi-level halftoning and will be described with particular reference thereto. It is to be appreciated that the present invention can also be applied to any type of color transformation process including color space transformation, device calibration and characterization, color correction, and the like.

[0003] Computers and other electronic equipment typically depict color in 3-D coordinates such as RGB. Many printers, on the other hand, print in either three-dimensional colorant space, cyan, magenta, yellow (CMY) or four-dimensional colorant space, cyan, magenta, yellow, and black (CMYK) which correspond to the input values, e.g. RGB. Frequently, a device independent color space is used as an intermediate representation of the image. A common choice for such a device independent space is a luminance-chrominance space, denoted generically as LC₁C₂. The L component represents luminance or lightness, and C₂and C₂are the two chrominance coordinates representing red-green and yellow-blue variations, respectively. An example of such a space is CIELAB. Translations are then derived from input RGB to LC₁C₂, and from LC₁C₂to printer colorant space.

[0004] While this invention is applicable to a wide variety of transformations, the discussion will focus on transforms from a luminance chrominance space LC₁C₂to printer colorant space CMYK. Such transformations are typically implemented by a 3-D look-up table (LUT), which converts each digital LC₁C₂input to the corresponding output CMYK value before being received by the printer.

[0005] A printer which has an ideal colorant behavior has a one-to-one correspondence of cyan-to-red, magenta-to-green, and yellow-to-blue. This means that when printed, the cyan colorant will only absorb red light, the magenta colorant will only absorb green light, and the yellow colorant will only absorb blue light. However, typical printer colorants deviate from this ideal behavior, and in fact absorb light in bands of the electromagnetic spectrum other than the intended absorption band. These so called unwanted absorptions lead to interactions between the colorants that result in a complex nonlinear relationship between digital values that drive the printer, and the resulting colorimetric response. A response, or other value, labeled as “colorimetric” refers to a measurement of the printed color, as seen by an average human observer, and represented in a device independent color coordinate system such as CIELAB. Modeling the calorimetric response across the entire range of CMYK values therefore cannot usually be achieved by a simple function, and in fact requires many parameters and measurements. The number of measurements required to characterize the printer adequately, can easily number 1,000 or more. Usually, in order to represent such a complex function at reasonable computational cost, a color correction LUT is built which approximates the mapping between colorimetric space and CMYK values. More specifically, the color correction LUT corrects for non-linearities and unwanted absorptions of colorants. For every input color specified in some luminance-chrominance space LC₁C₂, the LUT retrieves the corresponding CMYK which, when printed and measured, will yield the requested LC₁C₂color, provided that this color is within the reproducible gamut of the device.

[0006] To build the LUT, a predefined set of CMYK digital values are sent to the printer. The printer prints a corresponding set of color patches. The calibration color patches are measured and a colorimetric LC₁C₂coordinate is determined for each patch, i.e. for each of the predefined CMYK values. Each of the measured LC₁C₂coordinates then, identifies a three-dimensional vector location within the three-dimensional space. Each LC₁C₂coordinate is typically represented by 8-bit values for each of the L, C₁, and C₂components. Although such an 24-bit LC₁C₂coordinate is capable of addressing 256³locations, the look-up table is typically partitioned into a smaller size, such as 16×16×16 (4096) table locations, each node of which stores a CMYK value. CMYK values at intermediate LC₁C₂points are determined by some form of interpolation among the LUT nodes. The size of the look-up table is a compromise between the desired accuracy of the look-up table (i.e. the fidelity of the output) and the expense of storing a large number of values. Thus after the calibration patches are produced, each measured LC₁C₂coordinate has a corresponding known CMYK value. Unfortunately, the LC₁C₂coordinates do not, in general, perfectly coincide with the node locations (i.e. the three dimensional intersection points) of the look-up table. Hence, the CMYK values placed at the nodes are estimated by some multidimensional data fitting technique such as Shepard's algorithm.

[0007] An illustration of the operation of a look-up table is instructive. Referring to FIG. 1, an input LC₁C₂value 10 is sent into the table 12 for conversion into a printer specific CMYK value. The conversion is accomplished by interpolating the known CMYK values corresponding to the nodes 14 nearest the input LC₁C₂coordinate location 10. Because the color is defined in three dimensions, the interpolation is similarly done in three dimensions. Common examples of 3-D interpolation techniques include trilinear, tetrahedral, and prism interpolation. Of these, tetrahedral interpolation is the fastest method, requiring interpolation only among 4 nodes. (The trilinear and prism schemes utilize 8 and 6 nodes, respectively.) All these techniques require several multiplication, addition, shift, and comparison operations for each output signal at each pixel; and are often implemented with special purpose hardware. This interpolated CMYK value is then output by a printer. Unfortunately, three-dimensional interpolation presents a significant computational burden for many applications.

[0008] The present invention provides a new and improved method of color correction which overcomes the above-referenced problems and others.

[0009] In accordance with the present invention an improved method is proposed for lookup table based color transformations that combines multilevel halftoning with interpolation Specifically, if the input color space to the lookup table is in luminance-chrominance coordinates, multilevel halftoning is applied to each of the chrominance coordinates to achieve coincidence with the nodes of the lookup table along these two coordinates. The multilevel halftoning step introduces errors at high spatial frequencies in the chrominance channels. However, these errors are not easily perceptible by the human visual system. Finally, 1-dimensional interpolation is performed on the luminance component to obtain the output color values.

[0010] In accordance with an aspect of the present invention, a method of color transforming an input color value to an output color value includes a multi-dimensional color transformation table having an arrangement of nodes. Each node defines an output color value. The method includes receiving the input color value and applying multilevel halftoning to a predetermined subset of the input color components to select a plurality of nodes constrained to a selected number of dimensions. Remaining others of the input color components are then used to interpolate among the selected plurality of nodes to obtain the output color value.

[0011] In accordance with another aspect of the present invention, the halftoning step includes independently halftoning certain ones of the predetermined subset of input color components to select the plurality of nodes constrained to the selected number of dimensions.

[0012] In accordance with another aspect of the present invention, the halftoning includes applying a halftone screen function to ones of the predetermined subset of input colors to select the plurality of nodes.

[0013] In accordance with another aspect of the present invention, input color components are defined by L*a*b* components. The halftoning includes halftoning the a* and the b* components to select a plurality of nodes constrained to the L* dimension.

[0014] In accordance with another aspect of the present invention, the arrangement of nodes within the color transformation table is nonrectangular.

[0015] In accordance with the present invention, an electronic imaging system including a receiver and a conversion processor are provided. The conversion processor includes a multi-dimensional table having a plurality of nodes each containing predefined output color data. A first quantizer receives a predetermined subset of the input color components and identifies nodes constrained to a selected number of dimensions. An interpolator calculates an output color from the predefined color data contained at the nodes identified based on remaining ones of the input color components.

[0016] In accordance with another aspect of the present invention, the quantizer comprises a multilevel halftoner applying a halftone screen function to ones of the predetermined subset of input color components, and identifies the nodes constrained to the selected number of dimensions.

[0017] In accordance with another aspect of the present invention, the electronic imaging system further includes a second quantizer depending on the nodes identified by the first quantizer to identify the nodes constrained to the selected number of dimensions.

[0018] One advantage of the present invention is that the expensive 3-D interpolation is reduced to a much simpler process of halftoning and 1-D interpolation, which brings substantially savings in computation, with minimal loss in visual quality. Another advantage of the present invention is that the halftoning and 1-D interpolation steps require less storage and memory for precomputed quantities in the lookup table. This allows alternative grid structures to be employed for the lookup table.

[0019] Still further advantages of the present invention will become apparent to those of ordinary skill in the art upon reading and understanding the following detailed description of the preferred embodiments.

[0020] The invention may take physical form in certain parts and arrangements of parts, and in various steps and arrangements of steps. The drawings, then, must be presented for illustrative purposes only and should not be construed as limiting the scope of the present invention, wherein:

[0021] FIG. 1 is a graphic depiction of the prior art three-dimensional interpolation look up table;

[0022] FIG. 2 an illustration of one C₁C₂plane and an off-node arrival of an input color;

[0023] FIG. 3 is a graphical depiction of a logical flow diagram in accordance with an embodiment of the present invention; and

[0024] FIG. 4 is an illustration of interpolation along a remaining dimension following multilevel halftoning on the other dimensions;

[0025] FIG. 5 is graphical depiction of a sequential grid according to the present invention; and

[0026] FIG. 6 is a graphical depiction of a logical flow diagram in accordance with an alternate embodiment of the present invention.

[0027] A method is provided for reducing the cost of interpolation operations required in 3-D lookup table (LUT) based color correction. The input color value will be described in luminance-chrominance (LC₁C₂) coordinates, however, the present invention is equally applicable to other input formats with no loss of functionality. Multilevel halftoning is used to restrict C₁and C₂to a limited set of levels; namely those levels that coincide with the LUT node locations along C₁and C₂. The LUT calculation then reduces from 3-D interpolation in LC₁C₂to 1-D interpolation only along L. Testing shows that acceptable images can be obtained, with significant gains in computational cost as compared to the prior art 3D interpolation.

[0028] The present invention recognizes that reducing color correction from a 3-D interpolation problem to a one-dimensional problem will significantly reduce the cost and speed of color correction. Typically, device independent image values are stored as luminance-chrominance coordinates, for example the CIELAB color space. Generally, the human visual system is less sensitive to chrominance errors than it is to distortions in luminance at high spatial frequencies. In other words, an observer is more likely to detect small errors in luminance than in chrominance at high spatial frequencies. It is this recognition that is the basis of the present invention.

[0029] For clarity, a single C₁-C₂plane 20 is shown in FIG. 2. Input color 22 comprises input color components L_in, C_1inand C_2in. To reduce the computational cost of the ensuing color correction operation, some distortion in the C₁and C₂channels is accepted. Accordingly, the C₁component is limited to a finite set of dimensions or levels 24, 25, 26, 27, 28 each coinciding with a plurality of nodes (i.e. along the C₂and L axes). Similarly, the C₂component is constrained to match levels 31, 32, 33, 34, 35. One skilled in the art can appreciate that if an input color value is forced to always coincide with a level or a predefined dimension in the LUT, then there is no need to perform interpolation along that dimension. In other words, of the three input components, two (C₁and C₂) have been determined through multilevel halftoning. Hence, the color correction reduces to 1-D interpolation along the remaining axis (e.g. luminance or L) as will be shown below.

[0030] Reducing the precision of a signal to a small set of levels or dimensions is a quantization problem. A typical color correction LUT would assign 16 node locations along each of the L, C₁and C₂dimensions (from FIG. 1) resulting in a total of 16³=4096 node entries. Thus, for a typical 8-bit input, the present invention requires for each chrominance channel, a quantization mapping from 256 possible input levels to the 16 output node locations. Straightforward quantization involves mapping each input value to the closest node. However, experiments show that this approach can result in objectionable contouring artifacts.

[0031] A method of minimizing such artifacts in the quantized image is through dithering, or more generally, multilevel halftoning. This effectively pushes quantization errors into high spatial frequencies, exploiting another insensitivity in the human visual system. Preferably then, halftoning is applied to the chrominance channels C₁and C₂at high spatial frequencies thus, exploiting the greatly reduced sensitivity of the human visual system to high frequency chrominance errors. There are two known methods of multilevel halftoning: screening and error diffusion. The present invention preferably employs a blue noise stochastic screen, as it enjoys the computational ease of any screening technique, and borrows some of the more desirable qualitative behavior from error diffusion. An equally viable alternative is to use a dispersed dot screen. Still referring to FIG. 2, the halftoning operation can be envisioned as an input color 22 entering the LUT. As is typical, the input color 22 lies between the predefined levels 24-28 and 31-35 on both the C₁and C₂axes. Starting with C₁, a determination is made whether to place the input color 22 on either a slightly lower level 25 or a slightly higher level 26. No interpolation is performed, the input color is “forced” onto one of the levels, thus inducing an error. A similar determination is made for C₂forcing the input color 22 onto either the lower level 33 or the higher level 34. Assume for illustration purposes, the thresholding placed the input color on node 40.

[0032] As seen in FIG. 3, a conversion processor 50 receives the input color 22 comprising three components i.e. L 22_L, C₁22_Aand C₂22_B. In the illustrated embodiment, C₁22_Aand C₂22_Bare quantized independent of each other. The first step in multilevel halftoning is to normalize the coordinate C₁(x, y) 22_Aat pixel (x, y) within the interval formed by the two neighboring nodes (e.g. levels 22 and 23 from FIG. 2) in the LUT along the C₁. The normalized coordinate C₁′ is obtained by

[0033] C₁′(x,y)=[C₁(x,y)−C_1L(x,y)]/[C_1G(x,y)−C_1L(x,y)]  Equation (1)

[0034] where C₁(x, y) is the original input coordinate; C_1L(x, y) is the nearest LUT node less than C₁(x, y); and C_1G(X, y) is the nearest LUT node greater than C₁(x, y). The normalized coordinate C₁′(x, y) therefore always lies between 0 and 1. Note that the normalization operation can alternately be carried out ahead of time and stored in a 1-D lookup table array. The normalized value C₁′(x, y) is then compared with the halftone screen threshold value, otherwise termed the dither signal D(x, y) 52, which is also normalized between 0 and 1. If C₁′(x, y) is less than or equal to D(x, y), then C_1L(x, y) is picked as the output halftoned level, otherwise C_1G(x, y) is chosen. This output level is denoted CC_1q. An identical operation is performed for C₂(x, y) 22_B.

[0035] At this point, and with cross-reference to FIG. 4, the quantized values or dimensions 25 and 33 enter a 2-D lookup able 54 (FIG. 3) to determine a luminance line 60 defined by the intersection of the quantized C₁chrominance level or plane, e.g. 25 and the C₂chrominance level or plane 33. The result passes to the 1-D lookup and interpolation device 56 so that the L component of the input color 22_Lcan be interpolated (in 1-D) based on the remaining dimension (i.e. L) between the two nodes 64, 66 surrounding the input color value 22, thus determining the output color.

[0036] Referring now to FIG. 5, an alternative embodiment employs a sequential grid 70 to be used as an alternative to that shown in FIG. 2 for the 2-D chrominance plane in the 3-D LUT. The sequential grid 70 allows for a more flexible placement of nodes in the C₁-C₂plane, than the strictly rectangular grid placement as seen in FIG. 2. It allows nodes to be placed where the LUT transformation has greatest variation or visual importance, and in general, will result in a better trade-off between LUT accuracy and LUT size. In this embodiment, multilevel halftoning is applied to the first chrominance coordinate of the input color 72. This maps the C₁value or component to one of the levels 74 or 75. In the given example, the halftoning process constrains the possible nodes to the two-dimensional space defined by level 74. In the next step, multilevel halftoning is applied in the C₂dimension along the level 74 between nodes 76 and 77. In the given example, this results in the input color being mapped to node 77. Finally, as with the original embodiment, a 2-D lookup, followed by 1-D interpolation along the L axis is applied to obtain the output color.

[0037] A diagram of a conversion processor 80 according to this alternate embodiment process is given in FIG. 6. The flow processor 80 is identical in all essential respects to that of the original embodiment processor 50 (illustrated in FIG. 3) except that the halftoned output C_1qalong the first dimension is used to determine the set of LUT nodes along the second axis C₂, which may be different for different node values along C₁. Hence, the halftoning is carried out sequentially, rather than independently along the two chrominance axes.

[0038] In Table 1, three methods, trilinear interpolation, tetrahedral interpolation, and the original embodiment of the present invention, are compared in terms of the computations required to perform the color correction operation at each pixel, for N output colors. This operation includes 1) retrieving the nodes of the sub-cell enclosing the input color; and 2) performing the actual interpolation. Note that for the present invention, the comparison step in the multilevel halftoning operation has been included in the cost analysis. In all three cases, any quantities that do not depend on the input color are assumed to be precomputed and stored. These quantities include differences between output values at adjacent nodes, and the normalization of input values given by Equation (1). The last row of Table 1 shows the savings achieved by applying the present invention over tetrahedral interpolation. The savings would be even larger in comparison to trilinear interpolation.

[0039] The invention has been described with reference to the preferred embodiment. Obviously, modifications and alterations will occur to others upon a reading and understanding of this specification. It is intended to include all such modifications and alterations insofar as they come within the scope of the appended claims or the equivalents thereof.