JP2009111667A - Prediction of printing result - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To correctly predict printing results with less printing and measurements. <P>SOLUTION: In an initial stage, a gradation patch of primary colors is printed and a spectral reflectivity R(λ) is measured and a color patch is printed and a spectral reflectivity (λ) is measured only for the few first representative points. Thereafter, a spectral reflectivity (λ) is measured for second representative points using a neural network NNR wherein some of the first representative points, which are stable, are used as learned data CD1. By this method, a spectral reflectivity (λ) for a sufficient number of nodes (first representative points + second representative points) for predicting a spectral reflectivity (λ) using the Cellular Yule-Nielsen Spectral Neugebauer Model can be obtained. Since a color patch is printed and a spectral reflectivity (λ) is measured only for a few first representative points, a printing or measurement workability can be improved and thereby printing results can be predicted and a LUT can be created efficiently. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to prediction of printing results, and more particularly to prediction of printing results printed by depositing ink on a recording medium.

Conventionally, printing and colorimetry are performed only for some of the colors that can be printed by the printing apparatus, and the measurement results for all colors that can be printed by the printing apparatus are predicted based on the measurement results. For example, based on the cell division Yule-Nielsen spectroscopic Neugebauer model, a method for predicting the spectroscopic reflectance at an arbitrary ink amount has been proposed. In such a configuration, the ink amount space is divided into a plurality of cells surrounded by nodes having known spectral reflectances (lattice points that are evenly distributed in the ink amount (coverage) space), and each node constituting the cell is divided. Based on the spectral reflectance, the spectral reflectance of an arbitrary ink amount in the cell is predicted. Since the spectral reflectance of an arbitrary ink amount can be predicted based on the spectral reflectance of a node that is located at a relatively close position in the ink amount space, good prediction accuracy can be realized.
Pamphlet of International Publication WO2005 / 043884

In recent years, the types of ink that can be used are increasing with the increase in image quality of home printers, and the ink amount space in the above-described cell division Yule-Nielsen spectroscopic Neugebauer model is also becoming multidimensional. At the same time, the number of nodes described above increases exponentially with the increase in ink type. In addition, since it is impossible to reduce the density of nodes in the ink amount space in order to ensure good prediction accuracy, color patches must be printed / colorimetric (spectral reflectance measurement) for a large number of nodes. It was. Since a large amount of time is required for printing / colorimetry of a large number of color patches, there has been a problem that a period required for work such as creation of a conversion profile is prolonged. Furthermore, since the ink state fluctuates depending on the elapsed time after printing, there is a problem that measurement noise due to the difference in elapsed time is unavoidable in color measurement over a long period of time.
In order to solve the above problems, an object of the present invention is to accurately predict a printing result with a small number of printing / measurements.

  In order to achieve the above object, a printing result prediction method of the present invention is a printing result prediction method for predicting a printing result formed by adhering a plurality of types of ink on a recording medium, and includes a first measurement step, The first setting step, the second measurement step, the second setting step, the first prediction step, and the second prediction step are provided. In the first measurement step, printing is performed with a plurality of ink amount sets that form gradations with respect to the ink amount for each ink, and the printing result is measured. In the first setting step, a plurality of representative coordinates in the ink amount space are set as first representative points. In the second measurement step, printing is performed with an ink amount set corresponding to the first representative point, and the printing result is measured. In the second setting step, another representative coordinate different from the first representative point in the ink amount space is set as the second representative point. In the first prediction step, a printing result printed with the ink amount set corresponding to the second representative point is predicted based on the color measurement results in the first measurement step and the second measurement step. In the second prediction step, a print result printed with an arbitrary ink amount set is based on the color measurement result in the first measurement step, the color measurement result in the second measurement step, and the prediction result in the first prediction step. Predict.

The gradation in the first measurement step means that the ink amount changes stepwise, but the change in the ink amount for each step change may be equal or different at each step. As an example of the case where the change in the ink amount is different, it is conceivable that the color difference is substantially equal between the results of printing with different ink amount sets in one step.
Since the measurement result when printing is performed with the ink amount set corresponding to the second representative point is predicted in the first prediction step, printing is actually performed with the ink amount set corresponding to the second representative point. There is no need to measure the print result. Therefore, even when the type of ink increases, an increase in the number of the first representative points that actually require printing / measurement can be suppressed, and work for predicting the printing result can be made efficient.

  It is desirable that the prediction methods used in the first prediction step and the second prediction step are different from each other. This is because a suitable prediction method can be applied according to each prediction stage. For example, in the first prediction step of predicting the printing result when printing is performed with the ink amount set corresponding to the second representative point, a prediction method with higher prediction accuracy than that of the second prediction step is applied. Also good. This is because the prediction result of the first prediction step is used for prediction performed by the second prediction step, and thus higher accuracy is required. Of course, on the contrary, in the second prediction step, it is also possible to apply a prediction method with higher prediction accuracy than the first prediction step.

  As an example of a specific prediction method, a method using a neural network in which learning is performed based on the measurement result of the measurement step in the first prediction step may be used. According to the neural network, by inputting an arbitrary ink amount set including the second representative point, a print result corresponding to the ink amount set can be output as a prediction result.

  As a specific method for setting the first representative point, first, grid points distributed in the ink amount space may be generated, and the first representative point may be selected from the grid points. By doing so, it is possible to prevent the first representative points from being extremely close to each other in the ink amount space, and it is possible to prevent the prediction accuracy from being biased. Further, an area where the ink amount set used in the printing apparatus is distributed in the ink amount space is an effective area, and the density is higher in the vicinity of the outer edge of the effective area than in the vicinity of the center of the effective area. One representative point may be selected. In this case, the measurement result obtained by actually measuring the prediction result by the second prediction step for the arbitrary ink amount set existing in the vicinity of the outer edge of the effective area is more than the prediction result by the first prediction step. Can greatly contribute. In the vicinity of the outer edge of the effective area, for example, a phenomenon that is difficult to predict, such as ink bleeding, may occur. Therefore, the prediction is based on the actual measurement result in the second measurement step, not the prediction result in the first prediction step. It is desirable to do. On the other hand, since the behavior of ink coloring is stabilized near the center of the effective area, high prediction accuracy can be realized even if prediction is performed based on the prediction result of the first prediction step.

  By increasing the number ratio of the second representative point to the first representative point, it is possible to suppress the work of actual printing / measurement. On the contrary, by reducing the number ratio of the second representative point to the first representative point, the influence of the prediction error in the first prediction step is suppressed, and the prediction result in the second prediction step is made more suitable for actual measurement. High accuracy can be achieved. Thus, although workability and accuracy are in a trade-off relationship, the number of the grid points inside the effective area that are selected as the first representative point and the first representative point are selected. If the ratio with respect to the number of non-performed ones (that is, the second representative point) is about 1: 3, an appropriate balance between workability and accuracy can be ensured.

  The print result measured in the first and second measurement steps and predicted in the first and second prediction steps may be the spectral reflectance on the recording medium. By measuring / predicting the spectral reflectance as the printing result, it is possible to further specify the color when the color specifying step irradiates an arbitrary light source. Of course, the color on the recording medium may be measured / predicted as the printing result.

  Furthermore, it is also possible to create a lookup table using the printing result prediction method described above. That is, a color under a desired light source can be specified using the prediction method of the present invention, and a lookup table that defines the correspondence between the color and the corresponding ink amount set can be created. Also in this case, the work required for creating the lookup table can be reduced. Further, another conversion profile may be created using a lookup table that defines the correspondence between the ink amount set and the color. For example, a device link profile may be created by combining with the source ICC profile of the input device. Hereinafter, the ink amount set for which the correspondence relationship is defined in the lookup table is represented as an LUT ink amount set.

  Since the color corresponding to the LUT ink amount set is used when performing the interpolation process using the lookup table, it is desirable that the color is distributed with a high degree of smoothness in the color space. This is because the higher the degree of smoothness in the color space, the more uniform and better interpolation accuracy can be realized in the entire color space. Therefore, the LUT ink amount set is optimized so that the degree of smoothness of the distribution in the color space of the colors acquired by the spectral reflection prediction step and the color specifying step is increased.

  Furthermore, optimization may be performed by adding other viewpoints. For example, the LUT ink amount set can be optimized so that the image quality when printing with the LUT ink amount set is good. More specifically, good image quality may be realized by reducing the difference in color when two or more types of light sources are irradiated. According to the spectral reflectance prediction step, since the spectral reflectance when printing with the LUT ink amount set can be acquired, two or more types of light sources are applied to the spectral reflectance. More than one type of color under the light source can be specified. Further, as one of the image quality, the LUT ink amount set may be optimized so as to improve (reduce) the graininess.

  On the other hand, not only the image quality and smoothness, but also the LUT ink amount set may be optimized so that the total amount of each ink constituting the LUT ink amount set satisfies a predetermined condition. Thereby, for example, the total amount of ink deposited on the recording medium can be controlled.

  The technical idea of the present invention is specifically realized not only as a method but also as hardware such as a computer that executes the method or a program that causes a hardware such as a computer to execute a processing procedure according to the above method. It goes without saying that it is possible. In addition, the method of the present invention is not limited to a single unit, and may be incorporated as part of a certain method. For example, it goes without saying that the present invention can also be realized in a printing method or an image processing method that includes a part of a method for predicting a printing result by the method of the present invention. Moreover, each process which comprises this invention may be implement | achieved by the process performed in a distributed manner in several substantive apparatus. For example, some processes of the present invention may be realized by a personal computer, and other processes may be realized by a printing apparatus. Of course, each process of the present invention may be distributed via a network.

Next, embodiments of the present invention will be described in the following order.
A. Various converters and their preparation:
A-1. Spectral printing model converter:
A-2. Color converter:
A-3. Granularity converter:
A-4. Neural network learning:
B. Create a lookup table:
C. Summary and variations:

A. Various converters and their preparation A-1. Spectral Printing Model Converter FIG. 1 is a block diagram showing a configuration of a printing result prediction apparatus as a first embodiment of the present invention. In the figure, the printing result prediction apparatus is realized by hardware and software of a computer 10. Specifically, the CPU 12 included in the computer 10 reads the program data 11a stored in the hard disk drive (HDD) 11 or the like, and executes the calculation according to the program data 11a while expanding the program data 11a on the RAM 13. . And various processes which comprise the printing result prediction method of this invention by controlling external apparatuses, such as the printer 20, the spectral reflectometer 30, and the scanner 40 as a printing apparatus of this invention via the said interface by the said calculation. Is realized.

  The printer 20 includes an ink mounting unit 21 that mounts ink cartridges 22a, 22b,... The ink cartridges 22a, 22b,... Correspond to the color material container of the present invention. The printer 20 forms a print image on the printing paper by discharging a plurality of types of ink supplied from the ink cartridges 22a, 22b... Onto the printing paper as a recording medium. Specifically, it is possible to eject ink from nozzles formed on the carriage while main-scanning the carriage provided with the ink mounting portion 21 and performing sub-scanning by the paper feed roller, and combining each ink color Thus, a large number of colors are formed, thereby forming a printed image on the printing paper. The printer in this embodiment is an inkjet printer, but the present invention can be applied to various printers other than the inkjet printer.

  The printing result prediction apparatus is mainly composed of a spectral printing model converter RC, a color converter CC, and a graininess converter GC, and the spectral printing model converter RC is further composed of a preparation unit RC1 and a conversion unit RC2. The preparation unit RC1 uses a spectral printing model so that the spectral reflectance when the printer 20 prints with an arbitrary ink amount set can be predicted using the spectral printing model. Processing for preparing the spectral reflectance of the ink amount set corresponding to the representative point is executed. In order to execute this process, the preparation unit RC1 spectroscopically analyzes the first setting unit RC1a that sets the first representative point in the ink amount space and the patch that has been printed with the ink amount set corresponding to the first representative point. A measurement unit RC1b that causes the spectral reflectometer 30 to measure the reflectance; a second setting unit RC1c that sets another representative coordinate different from the first representative point in the ink amount space as a second representative point; A first prediction unit RCd for predicting a spectral reflectance when printing is performed with an ink amount set corresponding to a representative point based on a measurement result by the measurement unit RC1b; and a primary color gradation patch is printed in an ink amount space. And a next color gradation printing part RC1e. Hereinafter, the details of the units RCa1 to RC1e constituting the preparation unit RC1 will be described together with the flow of processing executed by these units.

  FIG. 2 shows a flow of spectral printing model converter preparation processing executed by the preparation unit RC1. In step S100, the first setting unit RC1a sets an ink set for preparing the spectral printing model converter RC. Here, as an example, the following description will be made assuming that an ink set of C (cyan) M (magenta) Y (yellow) K (black) lc (light cyan) lm (light magenta) is set. The ink set set here is set based on a combination of ink cartridges 22a, 22b,... That can be mounted on the ink mounting portion 21 of the printer 20, for example. When the ink set is set, the primary color gradation setting unit RC1e sets a plurality of ink amount sets G constituting the gradation from the primary colors.

  FIG. 3 is a diagram schematically showing the primary color. In the figure, for convenience of illustration, a CMY ink amount space is shown. The primary color is a color expressed by a single color ink without mixing two or more inks, and is a color expressed by an ink amount set on each ink amount axis in the ink amount space. The ink amount set for the primary color constituting the primary color gradation changes in the ink amount step by step, but the change in the ink amount for each step change may be equal or different at each step. Also good. For example, as an example of a case where the change in the ink amount is different, it is conceivable that the color difference is substantially equal between the results of printing with different ink amount sets in one step. Here, in order to make the color difference substantially uniform between the results of printing with different ink amount sets in one step, for example, it is possible to predict the spectral reflectance of the printing result based on the coverage according to the amount of ink applied It is. If the color differences are made substantially equal, normally, as shown in FIG. 3, in the ink amount space, the set density of the ink amount set increases as the ink amount increases.

  When the setting of the plurality of ink amount sets G is completed, the measurement unit RC1b acquires the ink amount set of the primary color gradation, and the measurement unit RC1b prints the color patch of the ink amount set. In the present specification, the color patch has a broad meaning including not only a chromatic patch but also an achromatic patch. When printing a color patch, image data in which a predetermined area is filled with pixels of the ink amount set of the first representative point is generated, and halftone processing and microweave processing are sequentially performed on the image data. Data is generated, and each part of the printer 20 is driven based on the print data to print a color patch. The color patches are printed by the number of first representative points, for example, a color chart in which each color patch is arranged so as to form a gradation. Of course, it is assumed that ink cartridges 22a, 22b,... That hold CMYKlclm ink are mounted on the ink mounting portion 21.

Further, the ink amount set corresponding to each of the first representative point in the printing of each color patch _{(d c, d m, d} y, d k, d lc, d lm) of ink approximately proportional to the discharge amount discharged It can be considered that it is attached to the printing paper. Accordingly, in this specification, the ink amount d _c prior to the halftone process or microweave _{process, d m, d y, d} k, d lc, ink discharge amount to be deposited on the actual printing paper of d _lm It will be handled without distinction. The printing paper on which the color patch is printed is a printing paper on which a lookup table to be described later is created. It is desirable to change the set number and the set ratio of the first representative point and the second representative point depending on the printing paper. For example, in the photo matte paper for pigment ink, the change in color with respect to the amount of ink is more abrupt than in normal photographic paper, so the number of first representative points and second representative points to be set is large (with high density). It is desirable.

  When the measurement unit RC1b confirms that printing of all the color patches has been completed and the specified drying time has been completed, the spectral reflectance meter 30 measures the spectral reflectance of each color patch. Each color patch is irradiated with light having a plurality of visible wavelengths, and the reflectance at each visible wavelength is measured. The measurement unit RC1b causes the HDD 11 to store the data storing the spectral reflectance obtained by the measurement as the primary color spectral reflectance data D0.

  When step S105 ends, the first setting unit RC1a sets the first representative point from the ink amount space in step S110. Here, since CMYKlclm is set as the ink set, the first representative point is set from a six-dimensional ink amount space having CMYKlclm as axes. In subsequent step S120, the second setting unit RC1c sets the second representative point from the ink amount space. Again, the second representative point is set from the CMYKlclm ink amount space.

  FIG. 4 schematically shows the processing of steps S100 to S120. In the figure, for the sake of simplification, the CM plane in the ink volume space (the YKlclm ink volume has a certain value) is shown, the vertical axis represents the ink volume axis of C ink, and the horizontal axis represents the M ink volume. The ink amount axis is used. In the figure, lattice points (representative coordinates) are formed with respect to the ink amount axis of CM, and these lattice points are set as candidates for the first representative point in step S110. The lattice points may be formed at equal intervals, but are preferably optimized for each ink in accordance with the color development characteristics of each ink color. For example, it is preferable to optimize so that the interval becomes narrower as the non-linearity of the ink coloring property becomes stronger. More specifically, for example, the rate of change of the color development characteristic with respect to the ink ejection amount is higher in linearity as the ink ejection amount is smaller, and the linearity tends to decrease as the ink ejection amount is larger. It can be considered that the interval between lattice points is set to be narrower. The accuracy of prediction of spectral reflectance R (λ) in a neural network NNR, which will be described later, increases as the linearity of the spectral reflectance increases, and the prediction accuracy also decreases as the linearity decreases. Therefore, the inter-lattice distance is narrowed at the portion where the linearity deteriorates in the measurement result of the primary color gradation described above. That is, by increasing the learning data at a site where the linearity is lowered, it is possible to improve the prediction accuracy of the spectral reflectance R (λ) of the unmeasured lattice point using a neural network described later.

  In selecting this grid point as a candidate for the first representative point, it is preferable that the grid point and the grid point selected in the gradation of the primary color do not overlap. That is, by avoiding measuring the same color twice, it is possible to avoid useless actual measurement, to prevent a substantial decrease in actual measurement grid points due to overlapping color measurement, and to prevent a decrease in the prediction system of the neural network NNR.

Although not shown, an equivalent grid is similarly formed for the ink amount axes of the other inks (YKlclm), and a lattice point in the entire ink amount space is a candidate for the first representative point. Accordingly, the number of grid points that are candidates for the first representative point is the number of grids raised to the power of the ink type. In this specification, the ink amount set means a combination of the ink amounts of the respective inks, and one ink amount set can be specified by specifying one coordinate in the ink amount space. Ink amount of each ink CMYKlclm is intended to represent _{d c, d m, d y} , d k, d lc, and d _lm. Ink amount set can be represented by a vector of _{(d c, d m, d} y, d k, d lc, d lm).

In FIG. 4, a boundary line LL between the effective area and the ineffective area is shown at a position where the amount of ink of CM has increased to some extent. The boundary line LL, the ink amount of _{YKlclm d y, d k, d} lc, at the time when a certain value d _lm, the printer 20 is capable CM ink amount d _c attached to the printing paper, a combination of d _m Means the upper limit of. For example, an area where ink bleeding does not occur on the printing paper is defined as an effective area, and an outer edge of the effective area is defined by a boundary line LL. If the printer 20 uses the ink amount set only inside the effective area, it is possible to accurately control the print result without being affected by ink bleeding. The boundary line LL can be determined according to the hardware, ink physical properties, surface physical properties of the printing paper, and printing environment of the printer 20 that is the target of printing result prediction. The effective area in the ink amount space corresponds to a color reproduction gamut that can be actually reproduced by the printer 20.

  In step S110, the first setting unit RC1a excludes, from the above-described lattice points, lattice points that are greatly out of the boundary line LL from the candidates for the first representative point. For example, grid points that are in the invalid area and that do not pass the boundary line LL within the unit grid surrounded by the grid points are excluded from the candidates for the first representative point. This is because it is not necessary to predict the printing result for an ink amount set that is not actually used by the printer 20. Furthermore, the first setting unit RC1a excludes some of the lattice points in the effective area from the candidates for the first representative point. Then, the lattice points that remain without being excluded are set as the first representative points. Here, it is desirable that the first representative points are uniformly distributed in the effective region, and a part of the lattice points is excluded from the candidates for the first representative points so as not to be biased. However, the first representative points do not have to be completely uniform, and for example, the distribution density of the first representative points in a part of the region may be increased. For example, a large number of first representative points may be set in a region of an ink amount set such as an achromatic color or a memory color that requires high color reproducibility when printing is performed by the printer 20. On the other hand, since the lattice points near the outer edge (boundary line LL) of the effective region that are outside the effective region are not excluded, they are all set as the first representative points.

In the above processing, the following four types of attributes are generated at the lattice points in the ink amount space. (In FIG. 4, the attribute numbers are shown in circles indicating the positions of the respective grid points.)
Attribute 1: Lattice point excluded outside the effective area Attribute 2: First representative point within the effective area Attribute 3: First representative point outside the effective area Attribute 4: Lattice point excluded within the effective area (Second representative point)
Here, the first setting unit RC1a excludes a predetermined number of grid points from the effective area so that the number ratio of the grid points of attribute 2 (first representative points) and the grid points of attribute 4 is approximately 1: 3. . In step S120, the second setting unit RC1c sets the grid point of attribute 4 that is excluded from the first representative point in step S110 as the second representative point. Eventually, the effective area is distributed in a state where the first representative points are thinned out. Therefore, the distribution density of the first representative points is higher near the outer edge of the effective area than inside the effective area.

  When each representative point is set as described above, in step S130, the measurement unit RC1b acquires the ink amount set of the first representative point, and prints the color patch of the ink amount set. In the present specification, the color patch has a broad meaning including not only a chromatic patch but also an achromatic patch. When printing a color patch, image data in which a predetermined area is filled with pixels of the ink amount set of the first representative point is generated, and halftone processing and microweave processing are sequentially performed on the image data. Data is generated, and each part of the printer 20 is driven based on the print data to print a color patch. The color patches are printed by the number of first representative points, for example, a color chart in which each color patch is arranged so as to form a gradation. Of course, it is assumed that ink cartridges 22a, 22b,... That hold CMYKlclm ink are mounted on the ink mounting portion 21.

Further, the ink amount set corresponding to each of the first representative point in the printing of each color patch _{(d c, d m, d} y, d k, d lc, d lm) of ink approximately proportional to the discharge amount discharged It can be considered that it is attached to the printing paper. Accordingly, in the present specification, the ink quantity d _c prior to the halftone process or microweave _{treatment, d m, d y, d} k, d lc, ink discharge amount to be deposited on the actual printing paper of d _lm It will be handled without distinction. The printing paper on which the color patch is printed is a printing paper on which a lookup table to be described later is created. It is desirable to change the set number and the set ratio of the first representative point and the second representative point depending on the printing paper. For example, in the photo matte paper for pigment ink, the change in color with respect to the amount of ink is more abrupt than in normal photographic paper, so the number of first representative points and second representative points to be set is large (with high density). It is desirable.

  When the measurement unit RC1b confirms that printing of all the color patches is completed and the specified drying time is completed, the spectral reflectance meter 30 measures the spectral reflectance of each color patch in step S140. Each color patch is irradiated with light having a plurality of visible wavelengths, and the reflectance at each visible wavelength is measured. The measurement unit RC1b causes the HDD 11 to store data storing the spectral reflectance obtained by the measurement as the spectral reflectance data D.

  FIG. 5A shows the contents of the spectral reflectance data D. In the figure, the reflectance for each wavelength is stored for each first representative point represented by an ink amount set. Each first representative point has the attributes 2 and 3 described above, and the attributes 2 and 3 are also stored so as to be identifiable. In step S150, the first prediction unit RCd extracts spectral reflectance data corresponding to the attribute 2 and primary color spectral reflectance data D0 from the spectral reflectance data D, and the extracted data is learned by a neural network described later. It is assumed that data CD1 is used. In the spectral reflectance data D, the correspondence between the ink amount set and the spectral reflectance is specified, and this correspondence can be used as the learning data CD1. In step S160, the first prediction unit RC1d learns the neural network using the learning data CD1. The neural network learning will be described in detail in section A-4.

FIG. 6 shows the neural network NNR learned in step S160. In the figure, the ink amount d _c of the _{ink, d m, d y, d} k, d lc, d lm has become possible input to the input layer of the neural network NNR, the output layer reflectance at each wavelength, That is, the spectral reflectance R (λ) can be output. In step S170, the first prediction unit RC1d inputs the ink amount set set as the second representative point in step S120 to the input layer of the neural network NNR, and the spectral reflectance R ( λ) is calculated. When the spectral reflectance R (λ) for all the second representative points can be calculated, the first predicting unit RC1d adds the second representative point and its spectral reflectance R (λ) to the spectral reflectance data D in step S180. Add the correspondence of. FIG. 5B shows the spectral reflectance data D after additional recording. As shown in the figure, the correspondence relationship between the spectral reflectance R (λ) for both the first representative point and the second representative point is described.

  According to the spectral reflectance data D described above, the ink amount set corresponding to all the grid points except the attribute 4 shown in FIG. 4 and the spectral reflectance R (λ) corresponding thereto are prepared. The reason why the attribute 3 that is the first representative point outside the effective area is not used as the learning data CD1 of the neural network NNR is that the spectral reflectance R (λ) may change irregularly due to, for example, ink bleeding. This is to prevent the learning data CD1 from causing noise. On the contrary, all the grid points are set as the first representative points around the outer edge of the effective region, and the color patch is actually measured, and the predicted value by the neural network NNR is not used. This is because, for a region that is difficult to predict, a decrease in accuracy can be suppressed by relying on an actual measurement value in prediction performed later. On the other hand, since it can be considered that the behavior of the spectral reflectance R (λ) with respect to the variation in the ink amount set is stable inside the effective area, the first representative point inside the effective area and the actually measured spectral reflection are reflected. The correspondence relationship with the rate R (λ) is used as the learning data CD1, and the predicted value by the neural network NNR learned thereby is adopted as the spectral reflectance R (λ) of the second representative point.

  Further, by using the measurement result of the primary color gradation patch as the learning data CD1 of the neural network NNR, the fluctuation of the spectral reflectance R (λ) corresponding to the fluctuation of each ink amount alone is surely included in the learning data CD1. I can do it. Since the spectral reflectance R (λ) corresponding to the variation of each ink amount alone is the minimum value when mixing the ink colors, the prediction accuracy of the neural network NNR is improved. In particular, when the number of primary color patches set in the vicinity of the maximum grid of each ink amount axis in the effective region of the ink amount space is equal to or larger than a predetermined number, the prediction accuracy of lattice points near each vertex in the ink amount space is determined by the neural network NNR. Is preferable.

  When the spectral reflectance data D can be prepared, the spectral reflectance R (λ) can be predicted by the conversion unit RC2. Note that the hardware that executes the conversion unit RC2 and the conversion unit RC2 corresponds to the second prediction unit of the present invention. In the present embodiment, the spectral reflectance when the conversion unit RC2 performs printing with an arbitrary ink amount set by using a Cellular Yule-Nielsen Spectral Neugebauer Model as a spectral printing model. R (λ) is predicted. In the following, the Cellular Yule-Nielsen Spectral Neugebauer Model will be described. This model is based on the well-known spectroscopic Neugebauer model and the Yule-Nielsen model. In the following description, for simplification of description, a model in the case of using three types of CMY inks will be described. However, a similar model was used using an arbitrary ink set including CMYKlclm of this embodiment. It is easy to extend to the model. For cell splitting Yule-Nielsen spectral Neugebauer model, Color Res Appl 25, 4-19, 2000 and R Balasubramanian, Optimization of the spectral Neugebauer model for printer characterization, J. Electronic Imaging 8 (2), 156- See 166 (1999).

ここで、ａ_iはｉ番目の領域の面積率であり、Ｒ_i（λ）はｉ番目の領域の分光反射率である。添え字ｉは、インクの無い領域（ｗ）と、シアンインクのみの領域（ｃ）と、マゼンタインクのみの領域（ｍ）と、イエローインクのみの領域（ｙ）と、マゼンタインクとイエローインクが吐出される領域（ｒ）と、イエローインクとシアンインクが吐出される領域（ｇ）と、シアンインクとマゼンタインクが吐出される領域（ｂ）と、ＣＭＹの３つのインクが吐出される領域（ｋ）をそれぞれ意味している。また、ｆ_c，ｆ_m，ｆ_yは、ＣＭＹ各インクを１種類のみ吐出したときにそのインクで覆われる面積の割合（「インク被覆率(Ink area coverage)」と呼ぶ）である。 FIG. 7 is a diagram showing a spectral Neugebauer model. The spectral Neugebauer model, optional ink amount sets _{(d c, d m, d} y) spectral reflectance of the printed matter when printed with R (lambda) is given by the following equation (1).

Here, a _i is the area ratio of the i-th region, and R _i (λ) is the spectral reflectance of the i-th region. The subscript i includes an area without ink (w), an area only with cyan ink (c), an area only with magenta ink (m), an area only with yellow ink (y), magenta ink and yellow ink. A region (r) where yellow ink and cyan ink are ejected, a region (b) where cyan ink and magenta ink are ejected, and a region where three inks CMY are ejected (region) (r) k) respectively. Further, f _c , f _m , and _fy are the proportions of the area covered with only one CMY ink when it is ejected (referred to as “Ink area coverage”).

The ink coverages f _c , f _m , and _fy are given by the Murray-Davis model shown in FIG. In the Murray-Davies model, for example, the ink area coverage f _c of the cyan ink is a nonlinear function of the ink amount d _c of the cyan, be converted to the ink amount d _c in the ink coverage f _c, for example by one-dimensional lookup table Can do. Ink coverage f _c, f _m, f _y is the ink amount d _c, d _m, reason for the non-linear function of d _y is spread enough ink in the case where a small amount of ink ejected to the unit area, This is because, when a large amount of ink is ejected, the ink is overlapped and the area covered with the ink does not increase so much. The same applies to other types of MY inks.

ここで、ｎは１以上の所定の係数であり、例えばｎ＝１０に設定することができる。（２ａ）式および（２ｂ）式は、ユール・ニールセン分光ノイゲバウアモデル(Yule-Nielsen Spectral Neugebauer Model)を表す式である。 When the Yule-Nielsen model for spectral reflectance is applied, the above equation (1) can be rewritten as the following equation (2a) or (2b).

Here, n is a predetermined coefficient of 1 or more, and can be set to n = 10, for example. Expressions (2a) and (2b) are expressions representing the Yule-Nielsen Spectral Neugebauer Model.

  The Cellular Yule-Nielsen Spectral Neugebauer Model adopted in the present embodiment is obtained by dividing the ink amount space of the above-described Yule-Nielsen Spectral Neugebauer Model into a plurality of cells. is there.

FIG. 8A shows an example of cell division in the cell division Yule-Nielsen spectroscopic Neugebauer model. Here, for simplification of description depicts the cell division in a two-dimensional ink amount space including two axes of the ink amount d _c, d _m of the CM inks. It notes that it for ink coverage f _c, is f _m with at Murray-Davis model described above the ink amount d _c, a unique relationship with d _m, the ink coverage f _c, also be considered as an axis indicating the f _m . White circles are cell division grid points (called “nodes”), and a two-dimensional ink amount (coverage) space is divided into nine cells C1 to C9. The ink amount set (d _c , d _m ) corresponding to each node is an ink amount set corresponding to the first representative point and the second representative point. That is, the spectral reflectance R (λ) of each node can be obtained by referring to the spectral reflectance data D described above. Therefore, the spectral reflectances R (λ) ₀₀ , R (λ) ₁₀ , R (λ) ₂₀ ... R (λ) ₃₃ of each node can be obtained from the spectral reflectance data D.

Table In fact, performs also the cell division in a six-dimensional ink amount space of CMYKlclm, the ink amount set coordinates also six dimensions of each node _{(d c, d m, d} y, d k, d lc, d lm) by Is done. Then, the ink amount set for each node _{(d c, d m, d} y, d k, d lc, d lm) first representative point or spectral reflectance of the second representative point R corresponding to (lambda) is the spectral reflectance It is acquired from the rate data D. The measured spectral reflectance R (λ) is acquired for the node corresponding to the first representative point, but the predicted value by the neural network NNR is obtained from the spectral reflectance R (λ) of the node corresponding to the second representative point. Get as. Further, since the first representative point and the second representative point are originally set from the coordinates existing on the lattice points in the ink amount space, the orthogonal cells C1 to C6 can be formed in the ink amount space.

FIG. 8 (B) shows the relationship between the ink area coverage f _c and the ink amount d _c which are used in the cell division model. Here, one kind of the ink amount in the range 0 to D _cmax of ink is also divided into three sections, the virtual used in the cell division model by non-linear curve which increases monotonically from 0 for each section to the 1 A typical ink coverage _fc is determined. For other inks, the ink coverages f _m and f _y are obtained in the same manner.

ここで、（３）式におけるインク被覆率ｆ_c，ｆ_mは図８（Ｂ）のグラフで与えられる値である。また、セルＣ５を囲む４つのノードに対応する分光反射率Ｒ（λ）₁₁，（λ）₁₂，（λ）₂₁，（λ）₂₂は分光反射率データＤを参照することにより取得することができる。これにより、（３）式の右辺を構成するすべての値を確定することができ、その計算結果として任意のインク量セット（ｄ_c，ｄ_m）にて印刷を行った場合の分光反射率Ｒ（λ）を算出することができる。波長λを可視光域にて順次シフトさせていくことにより、可視光領域における分光反射率Ｒ（λ）を得ることができる。インク量空間を複数のセルに分割すれば、分割しない場合に比べてサンプルの分光反射率Ｒ（λ）をより精度良く算出することができる。このように、本実施形態においては分光反射率データＤを参照しつつ変換部ＲＣ２が印刷結果としての分光反射率Ｒ（λ）を予測する。 FIG. 8C shows a method of calculating the spectral reflectance R (λ) when printing is performed with an arbitrary ink amount set (d _c , d _m ) in the center cell C5 of FIG. 8A. Is shown. The spectral reflectance R (λ) when printing is performed with the ink amount set (d _c , d _m ) is given by the following equation (3).

Here, the ink coverages f _c and f _m in the equation (3) are values given in the graph of FIG. The spectral reflectances R (λ) ₁₁ , (λ) ₁₂ , (λ) ₂₁ , and (λ) ₂₂ corresponding to the four nodes surrounding the cell C5 can be obtained by referring to the spectral reflectance data D. it can. Thereby, all the values constituting the right side of the expression (3) can be determined, and the spectral reflectance R when printing is performed with an arbitrary ink amount set (d _c , d _m ) as the calculation result. (Λ) can be calculated. The spectral reflectance R (λ) in the visible light region can be obtained by sequentially shifting the wavelength λ in the visible light region. If the ink amount space is divided into a plurality of cells, the spectral reflectance R (λ) of the sample can be calculated with higher accuracy than when the ink amount space is not divided. Thus, in the present embodiment, the conversion unit RC2 predicts the spectral reflectance R (λ) as a printing result while referring to the spectral reflectance data D.

  Although the spectral reflectance R (λ) is predicted in each divided cell, the node constituting the vertex of each cell may be the first representative point or the second representative point. However, since the distribution density of the first representative points is higher in the vicinity of the outer edge of the effective area than in the inner area of the effective area, the vicinity of the outer edge of the gamut is accurately measured by the spectral reflectance R (λ) actually measured. Can be predicted. On the other hand, since the behavior of the ink is stabilized inside the effective region where the distribution density of the first representative point is low, the spectral reflectance R (λ) is accurately predicted by the neural network NNR. Therefore, even if the spectral reflectance R (λ) predicted by the neural network NNR is used and the spectral reflectance R (λ) is predicted by the cell division Yule-Nielsen spectral Neugebauer model, the accuracy is greatly reduced. Never do. The spectral printing converter RC as a printing result prediction device can accurately predict the spectral reflectance R (λ) when printing with an arbitrary ink amount set by using it alone. A look-up table with good workability can be created by using the above converter together or using the predicted spectral reflectance R (λ). Hereinafter, other converters will be described.

A-2. Color Converter FIG. 9 schematically shows a process in which the color converter CC that executes the color specifying step of the present invention specifies a color based on the spectral reflectance R (λ). In the figure, the spectrum of the reflected light from the printed material is predicted by multiplying the spectrum of the desired light source at each wavelength λ of the spectral reflectance R (λ) predicted by the spectral printing converter RC. Further, the tristimulus values XYZ are calculated by convolving and normalizing the sensitivity functions x (λ), y (λ), and z (λ) under the desired observation conditions with respect to the spectrum of the reflected light. In this embodiment, unless otherwise indicated, the tristimulus values XYZ are calculated under the observation conditions of the CIE 1931 2 ° observer. As the light source, CIE standard D50 light, D65 light, F-system light, A-system light, or the like can be input. Further, the color converter CC calculates the L ^* a ^* b ^* value of the CIELAB color system by applying the conversion formula of the CIE standard to the tristimulus values XYZ. In this way, by using the spectral printing converter RC and the color converter CC sequentially, the L ^* a ^* b ^* value when printing is performed with an arbitrary ink amount set can be obtained.

Furthermore, the color converter CC can perform chromatic adaptation conversion on the tristimulus values XYZ. For example, by applying a chromatic adaptation conversion formula based on CIECAT02 to the tristimulus values XYZ calculated with D50 light, for example, L ^* a ^* expressing the appearance of color under D50 light with the corresponding color of D65 light b ^* value can be converted. Regarding CIECAT02, for example, “The CIECAM02 Color Appearance Model”, Nathan Moroney et al., IS & T / SID Tenth Color Imaging Conference, pp.23-27, and “The performance of CIECAM02”, Changjun Li et al., IS & T / SID Tenth Color Imaging Conference, pp.28-31. However, as the chromatic adaptation conversion, any other chromatic adaptation conversion such as von Kries's chromatic adaptation prediction formula may be used. The L ^* a ^* b ^* value obtained by this chromatic adaptation conversion is expressed as CV _{L1 → Ls} . This subscript “L1 → Ls” means that the color appearance under the light source L1 is an L ^* a ^* b ^* value expressed by the corresponding color of the standard light source Ls. The color converter CC obtains color values CV _{L1 → Ls} and CV _{L2 → Ls} representing the appearance under at least two comparative light sources L1 and L2 with colors corresponding to the standard light source Ls, and based on these values. A color constancy index CII is calculated. The color constancy index CII can be calculated by, for example, the following formula (4).

For color constancy index CII, see Billmeyer and Saltzman's Principles of Color Technology, 3rd edition, John Wiley & Sons, Inc, 2000, p. 129, p. 213-215. The right side of the equation (4) is a color difference ΔE ^* 94 (2 in the CIE 1994 color difference equation in which the values of the lightness and saturation coefficients kL and kC are set to 2 and the value of the hue coefficient kH is set to 1. : Corresponds to 2). In the CIE 1994 color difference equation, the denominator coefficients SL, Sc, SH on the right side of the equation (4) are given by the following equation (5).

In addition, as a color difference formula used for calculation of the color constancy index CII, other formulas can be used. The color constancy index CII is defined as the difference in color appearance when a color patch is observed under first and second different observation conditions. Therefore, an ink amount set that reduces the color constancy index CII when printed is preferable in that the difference in color appearance under different viewing conditions is small. In addition, since the color values CV _{L1 → Ls} and CV _{L2 → Ls} are the colorimetric values of the corresponding colors under the same standard observation conditions, the color constancy index CII, which is the color difference between them, makes the difference in color appearance fairly accurate It becomes the value expressed in. Next, the granularity converter GC and its preparation will be described.

A-3. Granularity Converter FIG. 10 shows the flow of the granularity converter GC preparation process. In step S200, a large number (N) of ink sets for graininess evaluation are prepared. The granularity evaluation ink amount set may be prepared on grid points that are evenly present in the ink amount space, or a colorimetric color space when printing is actually performed with the granularity evaluation ink amount set. You may prepare what exists equally in. Furthermore, it is also possible to focus on a specific color gamut such as an achromatic color or a memory color so as to accurately predict the graininess of a color gamut that requires a relatively large reduction in graininess. When the granularity evaluation ink amount set is prepared, a color patch is printed by the printer 20 based on the granularity evaluation ink amount set in step S210. The printing of the color patch here can be performed in the same procedure as in step S130 of the spectral printing model converter preparation process described above. However, since the first representative point where the color patch is printed in the spectral printing model converter preparation process and the ink amount set for graininess evaluation are different ink amount sets, it is necessary to print the color patches individually. Of course, it is also possible to share a color patch for the portion where the first representative point and the granularity evaluation ink amount set are common.

ＧＩについては、例えば、Makoto Fujino,Image Quality Evaluation of Inkjet Prints, Japan Hardcopy '99, p.291-294を参照。なお、（６）式のａ_Lは明度補正項、ＷＳ（ｕ）は画像のウイナースペクトラム、ＶＴＦは視覚の空間周波数特性、ｕは空間周波数である。 In step S215, the color patch is scanned by the scanner 40. Here, scanning is performed at a higher resolution than the resolution at which the printer printed the color patches. By doing so, it is possible to obtain image data capable of grasping in detail the distribution state of ink dots in each color patch. In step S220, the scanned image data is converted into image data L (x, y) having a lightness L ^* distribution on the print medium (x and y mean horizontal and vertical coordinates on the print medium, and x, The pixel specified by y is expressed as a sub-pixel). Processing for calculating the graininess index GI is started from the next step S225. The graininess index GI is the graininess (or the degree of noise) felt by an observer when an observer visually recognizes a certain printed matter. The smaller the graininess index GI, the smaller the graininess felt by the observer. . In this embodiment, the graininess index GI is defined by the following equation (6).

Regarding GI, see, for example, Makoto Fujino, Image Quality Evaluation of Inkjet Prints, Japan Hardcopy '99, p.291-294. In Equation (6), a _L is a brightness correction term, WS (u) is a winner spectrum of an image, VTF is a visual spatial frequency characteristic, and u is a spatial frequency.

  FIG. 11 illustrates how the graininess index GI is actually calculated. In this embodiment, the graininess index GI evaluates the graininess of a printed image by the spatial frequency (cycle / mm) characteristic of the brightness of the image. For this purpose, first, FFT (Fast Fourier Transformation) is performed on the spatial distribution L (x, y) in the brightness sub-pixel plane shown at the left end of FIG. 11 (step S230). In FIG. 11, the obtained spectrum of the spatial frequency is shown as S (u, v). The spectrum S (u, v) is composed of a real part Re (u, v) and an imaginary part Im (u, v), and S (u, v) = Re (u, v) + jIm (u, v). It is. This spectrum S (u, v) corresponds to the above-described winner spectrum.

Here, (u, v) has a dimension of the inverse space of (x, y). In this embodiment, (x, y) is defined as a coordinate, and in order to correspond to an actual length dimension, a scanner is used. For example, 40 scan resolutions must be considered. Therefore, even when evaluating S (u, v) in the spatial frequency dimension, dimension conversion is required. Therefore, first, the spatial frequency magnitude f (u, v) corresponding to the coordinates (u, v) is calculated. That is, the lowest frequency e _u in the main scanning direction X resolution per 25.4, the lowest frequency e _v in the sub-scanning direction is defined as Y resolution per 25.4. The X resolution and Y resolution are resolutions when the scanner 40 scans. Here, 1 inch is 25.4 mm. Minimum frequency e _u in each scanning direction, if e _v is calculated, the spatial frequency in arbitrary coordinates (u, v) size f (u, v) is ^{_{((e u · u) 2}} + (e v・ V) ² )) It can be calculated as ^1/2 .

  On the other hand, human eyes have different sensitivities to lightness depending on the magnitude f (u, v) of the spatial frequency, and the visual spatial frequency characteristic is, for example, VTF (f) shown in the lower center of FIG. It is a characteristic. VTF (f) in FIG. 11 is VTF (f) = 5.05 × exp (−0.138 · d · π · f / 180) × (1−exp (−0.1 · d · π · f / 180)). Here, d is the distance between the printed matter and the eyes, and f is the magnitude of the spatial frequency. Since f is expressed as a function of (u, v) described above, the visual spatial frequency characteristic VTF can be a function VTF (u, v) of (u, v).

  By multiplying the above-mentioned spectrum S (u, v) by this VTF (u, v), the spectrum S (u, v) can be evaluated in a state in which the visual spatial frequency characteristic is taken into consideration. If this evaluation is integrated, the spatial frequency can be evaluated for the entire sub-pixel plane. Therefore, in the present embodiment, the processing up to integration is performed in the processing of steps S235 to S255. First, both (u, v) are initialized to “0” (step S235), and a certain coordinate (u, The spatial frequency f (u, v) at v) is calculated (step S240). Further, the VTF at the spatial frequency f is calculated (step S245).

  When the VTF is obtained, the sum of the square of the VTF and the square of the spectrum S (u, v) and the variable Pow for substituting the integration result are calculated (step S250). That is, since the spectrum S (u, v) includes a real part Re (u, v) and an imaginary part Im (u, v), in order to evaluate the magnitude, first, the square of the VTF and the spectrum S ( Integration is performed by the square of u, v). Then, it is determined whether or not the above processing has been performed for all coordinates (u, v) (step S255). If it is not determined that the processing has been completed for all coordinates (u, v), unprocessed coordinates are determined. (U, v) is extracted, and the processing after step S240 is repeated. Note that, as shown in FIG. 11, VTF suddenly decreases and becomes almost “0” as the spatial frequency increases, so that the range of coordinates (u, v) is limited to a predetermined value or less in advance. Calculation can be performed within a necessary and sufficient range.

When the integration is completed, Pow ^1/2 / total number of pixels is calculated (step S260). That is, the dimension is returned to the dimension of the spectrum S (u, v) by the square root of the variable Pow, and normalized by dividing by the total number of pixels. By this normalization, an objective index (Int in FIG. 10) independent of the number of pixels of the input image is calculated. In the present embodiment, the graininess index GI is further corrected by taking into account the influence of the brightness of the entire printed matter. In other words, in this embodiment, even if the spatial frequency spectrum is the same, the printed matter gives different impressions to the human eyes when it is bright and dark, and the brighter one feels more grainy. Make corrections. For this reason, first, the lightness L (x, y) is added to all the pixels and divided by all the pixels to calculate the average lightness Ave of the entire image (step S265).

Then, the correction coefficient a (L) based on the brightness of the entire image is defined as a (L) = ((Ave + 16) / 116) ^0.8, and the correction coefficient a (L) is calculated (step S270) and the above Int The graininess index GI is multiplied (step S275). The correction coefficient a (L) corresponds to the brightness correction term a _L described above. With the above processing, the above equation (6) is specifically calculated. The correction coefficient may be a function that increases or decreases the coefficient value according to the average brightness, and various other functions may be employed. Of course, the component for evaluating the graininess index GI is not limited to the lightness component, and the spatial frequency may be evaluated in consideration of the hue and the saturation component, and the lightness component, red-green component, yellow- After calculating the blue component and Fourier transforming each, the graininess index GI may be calculated by multiplying the visual spatial frequency characteristic defined in advance for each color component.

  Through the processes in steps S205 to S275 described above, the granularity of the color patch actually printed in accordance with the granularity evaluation ink amount set can be quantified as the granularity index GI. In step S280, the correspondence relationship between each granularity evaluation ink amount set and the granularity index GI is stored in the HDD 11 as learning data CD2. In step S290, a process of creating a neural network NNG that predicts the graininess index GI when printing is performed with an arbitrary ink amount set by learning using the learning data CD2. The learning of the neural network NNG here will be described in detail in section A-4.

FIG. 12 shows the neural network NNG learned in step S290. In the figure, ink amounts d _c , d _m , d _y , d _k , d _lc , and d _{lm of} each ink can be input to the input layer of the neural network NNG, and the granularity generation number GI is output in the output layer. It is possible to do. Such, if the neural network NNG is possible preparations can convert any ink amount set _{(d c, d m, d} y, d k, d lc, d lm) of particulate generation number GI. Next, learning of the neural network NNR and the neural network NNG will be described. Although the neural network NNR and the neural network NNG have different output layers, almost the same learning method can be applied. Therefore, here, the neural network NNG will be described as an example.

A-4. Neural network learning:
FIG. 13 shows the flow of the granularity neural network (NNG) creation process (step S290 in FIG. 10). In step S300, each parameter for determining the structure of the neural network NNG is initialized. The neural network NNG has an ink amount vector d _j = (d _c , d _m , d _y , d _k , d _lc , d _lm ), (j = 1 to 6 = C) whose input layer corresponds to an arbitrary ink amount set. , M, Y, K, lc, lm). In general, since the input value and output value of a neural network are normalized to a predetermined range, the ink amount vector d _j = (d _c , d _m , d _y , d _k , d _lc , d _lm ) as an input value. It is desirable to normalize. The output layer has a graininess index GI, and the graininess index GI is also normalized and output in a predetermined value range.

The neural network NNG of the present embodiment has a three-layer structure, and one intermediate layer is set, and the number of intermediate units constituting the intermediate layer can be arbitrarily set. If the intermediate unit U _i (i = 1 to I) is described, the total number I (for example, I = 23, 18) of the intermediate units U _i is set in step S300. In general, if the number of intermediate units U _i is larger than the number of teacher signals (the number of correspondences between the ink amount set and the granularity index GI in the learning data CD2), the tendency of overfitting is strong, and the number of teacher signals If the number of intermediate units U _i is smaller than the number of intermediate units U _i , the tendency of underfitting becomes stronger. Therefore, considering the number N of granularity evaluation ink amount sets set in step S200 in FIG. It is desirable to set the total number I of perceptrons U _i .

上記の（７）式において各中間ユニットＵ_iは各インク量Ｉ_jに対して固有の重みＷ１ _ijを有しており、この重みＷ１ _ijによって各インク量Ｉ_jを重みづけて線形結合することにより得られる。また、各中間ユニットＵ_iは固有のバイアスｂ１ _iを有しており、同バイアスｂ１ _iが各インク量Ｉ_jの線形結合に加算される。ステップＳ３００では、すべての重みＷ１ _ijとバイアスｂ１ _iが初期設定される。初期段階では重みＷ１ _ijとバイアスｂ１ _iはどのように決めてもよく、例えば０を度数平均とした正規分布状に重みＷ１ _ijとバイアスｂ１ _iを分散させてもよい。 The value of each intermediate unit U _i is represented by the following equation (7).

Each intermediate unit U _i in the above equation (7) has a specific weight W1 _ij for each ink amount I _j, by linear combination with weighting each ink amount I _j by the weight W1 _ij Is obtained. Each intermediate unit U _i has a unique bias b1 _i , and the bias b1 _i is added to the linear combination of each ink amount I _j . In step S300, all weights W1 _ij and bias b1 _i are initialized. In the initial stage, the weight W1 _ij and the bias b1 _i may be determined in any way. For example, the weight W1 _ij and the bias b1 _i may be distributed in a normal distribution with 0 being the frequency average.

上記の（８）式において粒状性指数ＧＩは各中間ユニットＵ_iに対して固有の重みＷ２ _iを有しており、この重みＷ２ _iによって各中間ユニットＵ_iからの出力値Ｚ_iを重みづけて線形結合することにより得られる。同様にバイアスｂ２が加えられる。ステップＳ３１０では、各重みＷ２ _iとバイアスｂ２が初期設定される。なお、中間ユニットＵ_iと出力値Ｚ_iとの関係は下記の（９）式の伝達関数で表すことができる。

伝達関数は微分可能な単調増加連続関数であればよく、線形関数も適用することができる。本実施形態では、出力に非線形性を持たせるために非線形のハイパボリックタンジェント関数を設定する。むろん、同質の関数としてシグモイド関数を使用することもできる。ステップＳ３００では、すべての重みＷ２ _iとバイアスｂ２も初期設定される。重みＷ２ _iとバイアスｂ２についても初期段階ではどのように決めてもよく、ここでも０を度数平均とした正規分布状に重みＷ２ _iとバイアスｂ２を分散させてもよい。以上のようにして各パラメータを初期設定することにより、ニューラルネットワークＮＮＧの構造の概要が決定されたこととなる。ただし、各パラメータは適当に設定しただけであるため、これらを学習データＣＤ２によって学習させ最適化させる必要がある。 The finally obtained graininess index GI is expressed by the following equation (8).

Graininess index GI in the above (8) has a specific weight W2 _i for each intermediate unit U _i, weighted output values Z _i from the intermediate unit U _i by the weight W2 _i Obtained by linear combination. Similarly, a bias b2 is applied. In step S310, each weight W2 _i and bias b2 are initialized. The relationship between the intermediate unit U _i and the output value Z _i can be expressed by a transfer function of the following equation (9).

The transfer function may be a monotonically increasing continuous function that can be differentiated, and a linear function can also be applied. In the present embodiment, a nonlinear hyperbolic tangent function is set in order to give nonlinearity to the output. Of course, a sigmoid function can be used as a homogeneous function. In step S300, all weights W2 _i and bias b2 are also initialized. The weight W2 _i and the bias b2 may be determined in any way at the initial stage, and the weight W2 _i and the bias b2 may be distributed in a normal distribution with 0 as the frequency average. By initializing each parameter as described above, the outline of the structure of the neural network NNG is determined. However, since each parameter has only been set appropriately, it is necessary to learn and optimize these using the learning data CD2.

Therefore, in step S310, each parameter is optimized. Here, the parameters W1 _ij , b1 _i , W2 _i and b2 are optimized by an error back propagation method. In the back propagation method, an output (granularity index GI) with respect to an input (granularity evaluation ink amount set) in the learning data CD2 and an output when a granularity evaluation ink amount set is input to the learning neural network NNG. The parameters of each layer are sequentially determined by sequentially propagating an error from the graininess index _GINN to the previous layer.

On the other hand, in learning of the neural network NNR that predicts the spectral reflectance R (λ), the output (spectral reflectance R (λ) corresponding to the input (the ink amount set corresponding to the first representative point of attribute 2) in the learning data CD1). )) And the spectral reflectance R _NN (λ) that is output when the ink amount set corresponding to the first representative point of attribute 2 is input to the learning neural network NNR, is sequentially transferred to the previous layer. By propagating, the parameters of each layer are sequentially determined.

As a basic policy of optimization, the graininess index GI _NN predicted by the neural network NNG is learned by correcting the parameters W1 _ij , b1 _i , W2 _i , and b2 so as to minimize the above-described error. The value is close to the graininess index GI defined in the data CD2. However, if this policy is followed, if the granularity index GI defined in the learning data CD2 includes noise, the noise is also reproduced at the output of the neural network NNG. That is, it becomes overfitting. Therefore, an evaluation function E of the following equation (10) is prepared to suppress overfitting.

そして、最適化対象のパラメータｐを変動させつつ評価関数Ｅをパラメータｐによって偏微分することにより同評価関数Ｅの勾配を求め、その勾配の絶対値が最も小さくなるパラメータｐの値を最適化後のパラメータｐとする（勾配法）。これにより、最適化対象のパラメータｐの変動において最も評価関数Ｅが小さくなるパラメータｐを特定することができる。なお、パラメータＷ１ _ij，ｂ１ _i，Ｗ２ _i，ｂ２のうち最適化対象のパラメータをパラメータｐと表記するものとし、最適化対象のパラメータｐは出力から近い順に順次設定される。一通りすべてのパラメータＷ１ _ij，ｂ１ _i，Ｗ２ _i，ｂ２が最適化されると、同様の処理を所定回数または評価関数Ｅが所定の閾値を下回るまで繰り返す。これにより、パラメータＷ１ _ij，ｂ１ _i，Ｗ２ _i，ｂ２間の交互作用も反映させつつ、評価関数Ｅを徐々に小さく収束させていくことができる。

Then, the gradient of the evaluation function E is obtained by partially differentiating the evaluation function E with the parameter p while changing the parameter p to be optimized, and the value of the parameter p having the smallest absolute value of the gradient is optimized. Parameter p (gradient method). Thereby, it is possible to specify the parameter p having the smallest evaluation function E in the variation of the parameter p to be optimized. Of the parameters W1 _ij , b1 _i , W2 _i , and b2, the parameter to be optimized is expressed as a parameter p, and the parameter p to be optimized is sequentially set in order from the output. When all the parameters W1 _ij , b1 _i , W2 _i , b2 are optimized, the same processing is repeated a predetermined number of times or until the evaluation function E falls below a predetermined threshold value. Thereby, the evaluation function E can be gradually converged while reflecting the interaction between the parameters W1 _ij , b1 _i , W2 _i , and b2.

すなわち、すべての粒状性評価用インク量セット（ｎは粒状性評価用インク量セット番号であり、ｎ＝１〜Ｎとする。）についての粒状性指数ＧＩ_NN，ＧＩの２乗誤差によって誤差関数Ｅ_Dが表される。誤差関数Ｅ_Dが含まれる評価関数Ｅを最小化させることにより、ガマットを網羅するあらゆるインク量セットについてカラーパッチの実評価により得られた学習データＣＤ２の粒状性指数ＧＩとニューラルネットワークＮＮＧにて予測する粒状性指数ＧＩ_NNとのずれを最小化させることができる。なお、所定回数最適化を繰り返しても評価関数Ｅが所望する閾値を下回らない場合には、中間ユニットＵ_iの数Ｉを増加させて、フィッティング能力を向上させてもよい。逆に、異常に少ない最適化回数でも評価関数Ｅが所望する閾値を下回った場合には、中間ユニットＵ_iの数Ｉを減少させて、フィッティング能力を抑制してもよい。 Incidentally, E _D is an error function for evaluating the error of the graininess index GI _NN and GI, and is expressed by the following equation (11).

That is, an error function is obtained by the square error of the graininess indexes GI _NN and GI for all the graininess evaluation ink amount sets (n is a graininess evaluation ink amount set number, where n = 1 to N). E _D is represented. By minimizing the evaluation function E including the error function E _D, prediction is made by the granularity index GI of the learning data CD2 obtained by the actual evaluation of the color patch and the neural network NNG for all ink amount sets covering the gamut. The deviation from the graininess index _GINN to be performed can be minimized. If the evaluation function E does not fall below the desired threshold even after the optimization is repeated a predetermined number of times, the number I of intermediate units U _i may be increased to improve the fitting ability. Conversely, if the evaluation function E falls below the desired threshold even with an unusually small number of optimizations, the number I of intermediate units U _i may be reduced to suppress the fitting ability.

上記（１２）式においては最適化対象のパラメータｐ_sの２乗和によって抑止関数Ｅ_Wが表される。なお、添え字ｓ（ｓ＝１〜Ｓ）は同種のパラメータｐの数を意味し、例えば重みＷ２ _iが最適化対象のパラメータｐとされた場合には、ｉ（ｉ＝１〜Ｉ）がｓ（ｓ＝１〜Ｓ）に相当する。上記（１２）式によれば、抑止関数Ｅ_Wを含む評価関数Ｅを最小化させることにより、パラメータｐ_sを０に拘束させることができる。ＮＮにおいて重みＷ１ _ij，Ｗ２ _iの絶対値が大きくなると、出力される粒状性指数ＧＩ_NNの変動曲線の屈曲が急峻となる。このような場合、ノイズを含む異常な教師信号（学習データＣＤ２の粒状性指数ＧＩ）の影響を受けている可能性が高い。従って、抑止関数Ｅ_Wによって重みＷ１ _ij，Ｗ２ _iを０に拘束させることにより、粒状性指数ＧＩ_NNの屈曲を平滑化し、ノイズを含む学習データＣＤ２の粒状性指数ＧＩに起因するオーバーフィッティングを抑止することができる。 On the other hand, E _W is a deterrence function for inhibiting the granularity index GI _NN predicted by the neural network NNG from being overfitted with the granularity index GI of the learning data CD2 based on the actual evaluation. Represented by an expression.

In the above equation (12), the suppression function E _W is represented by the square sum of the parameters p _{s to be} optimized. The subscript s (s = 1 to S) means the number of parameters p of the same type. For example, when the weight W2 _i is the parameter p to be optimized, i (i = 1 to I) is This corresponds to s (s = 1 to S). According to the above equation (12), the parameter p _s can be constrained to 0 by minimizing the evaluation function E including the inhibition function E _W. When the absolute values of the weights W1 _ij and W2 _i increase in the _NN , the curve of the output granularity index GI _NN becomes steep. In such a case, there is a high possibility of being affected by an abnormal teacher signal including noise (granularity index GI of learning data CD2). Therefore, by constraining the weights W1 _ij and W2 _i to 0 by the inhibition function E _W , the curvature of the granularity index GI _NN is smoothed, and overfitting caused by the granularity index GI of the learning data CD2 including noise is inhibited. can do.

なお、上記（１３）式と上記（１４）式におけるγは下記（１５）式によって表される。

Here, α and β in the above equation (10) can be regarded as coefficients (hyper parameters) for adjusting the weights of the error function E _D and the inhibition function E _W in the evaluation function E. The hyper parameters α and β are given by the following equations (13) and (14).

Note that γ in the above formula (13) and the above formula (14) is expressed by the following formula (15).

Note that λ _s in the above equation (15) is an eigenvalue of an S row × S column Hessian matrix obtained by second-order differentiation of the error function E _D with the parameter p _s to be optimized. This eigenvalue λ _s can be said to reflect the variation in the slope of the error function E _D with respect to the parameter p _s . When the inclination variation of the error function E _D is large, the hyper parameter α is increased, and the contribution of the inhibition function E _W is increased. On the contrary, when the inclination variation of the error function E _D is small, the hyper parameter β is increased, and the contribution of the error function E _D is increased.

That is, when the granularity index GI _NN predicted by the neural network NNG can rapidly follow the granularity index GI of the learning data CD2 based on the actual evaluation or cannot follow it, the surrounding learning data CD2 Is likely to be abnormal (large influence of noise). In this case, the weight of the inhibition function E _W is increased. As a result, it is possible to prevent excessive fitting to an abnormal graininess index GI, and to obtain an output of the neural network NNG with high smoothness. It is desirable to update the hyper parameters α and β when the parameter optimization has progressed to some extent.

When the parameters W1 _ij , b1 _i , W2 _i , b2 and I can be set as described above, the structure of the neural network NNG is determined and the preparation of the granularity converter GC is completed. On the other hand, the structure of the neural network NNR that predicts the spectral reflectance R (λ) can be similarly optimized. However, since the output of the neural network NNR is constituted by the spectral reflectance R (λ) of each wavelength, an error averaged for each wavelength is propagated back to the error. As a result, the neural network NNR can be constructed so that the spectral reflectance R (λ) for each wavelength is predicted with high accuracy as a whole.

B. Creation of Lookup Table In the above, various converters RC, CC, GC as a printing result prediction apparatus and preparation thereof have been described. In the following, a lookup table is created using various converters RC, CC, GC. A lookup table creation device and a print control device that performs color conversion using the created lookup table will be described. In this specification, the lookup table may be abbreviated as LUT.

  FIG. 14 shows the configuration of a computer 110 that specifically implements the lookup table creation method of the present invention. In the figure, a computer 110 has hardware resources such as an HDD 111, a CPU 112, and a RAM 113, and these can execute processing according to program data 111a. In the computer 110, each converter RC, CC, GC can be used. Here, it is only necessary to be able to use each converter RC, CC, GC, and it is not necessary to have a process for preparing each converter RC, CC, GC described above. In the case of the spectral printing model converter RC, the conversion algorithm RC2 includes a prediction algorithm based on the cell division Yule-Nielsen spectral Neugebauer model and spectral reflectance data D that gives the spectral reflectance R (λ) of each node used in the model. As long as it can be used in If the granularity converter GC, the neural network NNG may be usable. Of course, the computer 10 and the computer 110 may be the same computer, and the converters RC, CC, and GC may be prepared and used consistently.

  In the computer 110, the LUT creation unit LM is executed in addition to the converters RC, CC, and GC, and the process of creating the LUT by the LUT creation unit LM is specifically realized. The LUT creation unit LM includes an initial LUT generation unit LM1, an evaluation function setting unit LM2, a smoothness degree calculation unit LM3, an optimization unit LM4, and an LUT generation unit LM5. Hereinafter, the details of the profile creation processing executed by each of the modules LM1 to LM4 will be described based on the flow.

FIG. 15 shows the flow of profile creation processing. In step S400, an initial ink profile IP is created. The ink profile IP is an absolute color space CIELAB color space (L ^* a ^* b ^*) and CMYKlclm space is an ink amount space _{(d c, d m, d} y, d k, d lc, d lm) Is a LUT that defines a plurality of representative lattice points. In the creation of the initial ink profile IP, for example 17 ^three sets of random ink amount set from among the effective region as described above _{(d c, d m, d} y, d k, d lc, d lm) generating a. The 17 ^three sets of the ink amount set is the initial value of the ink amount set for LUT of the present invention, it will be optimized by processing described later. Since the initial LUT ink amount set is finally optimized, it may be generated in any way in the initial stage.

  Next, in step S410, the evaluation function setting unit LM2 sets an evaluation function. That is, an optimization guideline described later is set. Since the ink profile IP defines the correspondence between the CIELAB color space, which is an absolute color space, and the ink amount space with respect to a finite number of lattice points, correspondence relationships other than the lattice points are predicted in the future by interpolation processing. The Rukoto. In general, when a grid point arranged in order in each color space calculates a color positioned between them by interpolation calculation, interpolation can be performed without greatly changing the interpolation accuracy depending on the local position of the space. Therefore, by adopting an index for smoothing the grid point arrangement in the CIELAB color space as an optimization guideline in the present embodiment, interpolation calculation is performed with high accuracy at the time of ink profile IP creation and color conversion after creation. It becomes possible to do. As a result, it is possible to create an ink profile IP that can suppress the occurrence of tone jump and can obtain a printed matter in which gradation changes smoothly.

  In the ink profile IP, in order to realize as wide a color reproducibility as possible, the grid points should be distributed over the entire color reproduction gamut that the printer 20 can reproduce with the ink set (CMYKlclm). Therefore, by adopting an index that secures the color reproduction gamut in the CIELAB color space as an optimization guideline in the present embodiment, an ink profile IP capable of realizing a wide color reproducibility can be realized. With the above guidelines, an optimal distribution of lattice points in the CIELAB color space can be specified.

However, even determine the optimal grid point in the CIELAB color space, the ink amount set corresponding to the lattice point by the ink profile _{IP (d c, d m,} d y, d k, d lc, d lm) of It cannot be determined uniquely. The ink profile IP defines the correspondence between the CIELAB color space, which is an absolute color space, and the CMYKlclm space, which is an ink amount space, but the correspondence between the CIELAB color space and the CMYKlclm space has a unique relationship. It is not. That is, even if one L ^* a ^* b ^* value is determined in the CIELAB color space, an ink amount set (d _c ,) that realizes a print result in which the L ^* a ^* b ^* value can be reproduced under a certain light source. _{d m, d y, d k} , d lc, it is impossible to uniquely determine d _lm). For example, since K ink and CMY ink have a relationship capable of color separation, the same L ^* a ^* b ^* can be reproduced even if the color separation ratio is changed in a certain light source. The same applies to the relationship between C ink and lc ink and between M ink and lm ink.

Therefore, at the same time determine the optimal grid point in the CIELAB color space, the ink amount set corresponding to the lattice point _{(d c, d m, d} y, d k, d lc, d lm) necessary to is also optimized There is. For example, the separation ratio between K ink and CMY ink cannot be determined uniquely even if the L ^* a ^* b ^* value in the CIELAB color space is determined. However, if dark K ink is generated in the highlight area, the graininess is reduced. It will stand out. Therefore, if the pointer of optimizing the improved granularity, the highlighted area L ^* a ^* b ^* a d _k = 0 for values ink amount sets _{(d c, d m, d} y, d _k , d _lc , d _lm ). Conversely, the color separation ratio between K ink and CMY ink cannot be determined uniquely even if the L ^* a ^* b ^* value in the CIELAB color space is determined, but a lot of composite gray is used with CMY ink whose spectral reflectance is not flat. In this case, the light source dependency of color becomes a problem. Therefore, the grid points of the ink profile IP should be optimized using the improvement of color constancy as a guideline for optimization. Furthermore, the ink amount sets _{(d c, d m, d} y, d k, d lc, d lm) by applying the guidance to reduce overall size of the optimal ink profile in terms of the ink running costs IP can be created.

上記の（１６）式において、評価関数Ｅ_pは５個の項をｗ₁〜ｗ₄の重み係数によって加算した値であり、各項がそれぞれ上述した格子点を選択する指針に基づいて設定されている。（１６）式の第１項は、上述した粒状性コンバータＧＣ（ニューラルネットワークＮＮＧ）によって得られる粒状性指数ＧＩを最大値で除算することによって正規化し、所定の重み係数ｗ₁を乗算したものとなっている。なお、ψはインク量セット（ｄ_c，ｄ_m，ｄ_y，ｄ_k，ｄ_lc，ｄ_lm）を表し、評価関数Ｅ_pの添え字ｐ（ｐ＝１〜１７³）は注目する格子点の識別符号を示している。粒状性指数ＧＩは小さい方が良好な画質となるため、（１６）式の第１項が小さくなるほど最適であるといえる。 As described above, in order to create the optimum ink profile IP, it is preferable to optimize the grid points in consideration of various factors, and to set the optimum color separation rule in consideration of all factors. It is difficult. Therefore, in the present embodiment, an evaluation function E _p that can simultaneously evaluate these elements is set based on the above-described guidelines (step S410). Specifically, the evaluation function E _p is set by the following equation (16). In addition, the light source used when calculating the evaluation function E _p is also set.

In the above equation (16), the evaluation function E _p is a value obtained by adding five terms by weighting factors w ₁ to w ₄ , and each term is set based on the above-described guidelines for selecting the lattice points. ing. The first term of the equation (16) is obtained by normalizing the graininess index GI obtained by the above-described graininess converter GC (neural network NNG) by dividing by the maximum value and multiplying by a predetermined weighting factor w _1. It has become. Incidentally, [psi ink amount sets _{(d c, d m, d} y, d k, d lc, d lm) represents, subscript p of the evaluation function E _{p (p} = 1~17 ³⁾ the lattice point of interest The identification code is shown. The smaller the graininess index GI, the better the image quality. Therefore, it can be said that the smaller the first term of the equation (16), the more optimal.

(16) The second term of the color constancy index CII obtained by the color converter CC normalized by dividing by the maximum value, and is obtained by multiplying a predetermined weight coefficient w _2. The color constancy index CII is obtained by inputting an arbitrary ink amount set (d _c , d _m , d _y , d _k , d _lc , d _lm ) into the spectral printing model converter RC. ) Is further converted by the color converter CC, and can be said to be a function of ψ. (16) the third term in is made shall smoothness degree evaluation index SI obtained by smoothness calculator LM3 normalized by dividing by the maximum value, obtained by multiplying a predetermined weighting coefficient w _3. The smaller the color constancy index CII, the better the image quality. Therefore, the smaller the second term in the equation (16), the more optimal. The light source from which the color converter CC calculates the color constancy index CII is set in step S410. For example, the comparison light sources L1 and L2 are set to D50 light, the F11 light is set, and the standard light source Ls is set to D65 light.

平滑程度評価指数ＳＩは、注目する格子点から互いに逆向きのベクトルの距離が等しく、方向が正反対に近いほど値が小さくなるようにしてある。 FIG. 16 schematically illustrates the smoothness degree evaluation index SI. In the figure, ◯ indicates the position of a plurality of lattice points in the CIELAB space, and ● indicates the lattice point of interest (the lattice point for which the evaluation function _Ep is calculated) among the lattice points. When the position vector of the lattice point of interest is Lp and the position vectors of the six lattice points adjacent to the lattice point are L _{a1 to} La ₆ , the smoothness degree evaluation index SI is expressed by the following equation (17). .

The smoothness degree evaluation index SI is set such that the distance between vectors in the opposite direction from the lattice point of interest is equal, and the value becomes smaller as the direction is closer to the opposite direction.

As shown in FIG. 16B, lines connecting adjacent lattice points (such as a line passing through the lattice points indicated by the vector L _a1 to the vector L _p to the vector L _a2 ) are close to a straight line, and the lattice points are evenly arranged. Since the arrangement of lattice points in the CIELAB color space tends to be smoothed as the degree of smoothness increases, the degree of smoothness increases as the smoothness degree evaluation index SI shown in Expression (17) decreases. L ^* a ^* b ^* values in the CIELAB color space, by sequentially converting the ink amount sets _{(d c, d m, d} y, d k, d lc, d lm) the spectral printing model converter RC and color converter CC Can be obtained. The light source for which the color converter CC calculates the L ^* a ^* b ^* value uses the D65 light set as the standard light source Ls in step S410. Accordingly, it can be said that the smoothness evaluation index SI specified by the position vector in the CIELAB color space in the above equation (17) is a function of ψ. The smaller the smoothness degree evaluation index SI, the higher the interpolation accuracy can be expected. Therefore, it can be said that the smaller the third term of the equation (16), the more optimal. Next, the fourth term of the equation (16) indicates whether or not the position vector L _p of the target lattice point is close to a specific color.

  FIG. 17 shows the color reproduction gamut of the printer 20 in the CIELAB color space. As shown in the figure, the color reproduction gamut of the printer 20 is determined in advance by the hardware specifications of the printer 20 and the ink set, and colors can be reproduced within this range. Accordingly, the grid points of the ink profile IP must be present in the entire color reproduction gamut in the CIELAB color space. For this reason, some grid points need to be constrained to the outer surface, ridgeline, or vertex of the color reproduction gamut. By adding the color difference ΔE between the color of the color reproduction gamut on the outer surface, the ridgeline, and the vertex and the lattice point as the fourth term of the equation (16), and performing optimization, the lattice point exists in the entire color reproduction gamut. Can do. Note that the fourth term of the equation (16) is also specified by the position vector in the CIELAB color space, and thus can be said to be a function of ψ. As the lattice point showing the same color as the outermost color of the color reproduction gamut is included, the widest color reproducibility can be realized. Therefore, it can be said that the smaller the fourth term in the equation (16), the more optimal.

(16) Section 5 of the is the ink amount set _{(d c, d m, d} y, d k, d lc, d lm) the total value of those normalized. Thereby, the grid point can be optimized by the evaluation function E _p taking into account the total consumption of ink. (16) Section 5 also, the ink amount set depending _{(d c, d m, d} y, d k, d lc, d lm) , it is possible that a function of [psi. Note that T _{Duty in the} equation (16) is a value corresponding to the limit of the amount of ink that can adhere to the recording medium. The smaller the amount of ink, the better the running cost. Therefore, the smaller the fifth term in equation (16), the more optimal.

All sections constituting the evaluation function E _p as described above, the ink amount sets _{(d c, d m, d} y, d k, d lc, d lm) with is represented by a function of, as smaller Lattice points are optimal. Accordingly, in step S400, only the initial grid points are appropriately determined. Therefore, even if the evaluation function E _p is calculated by paying attention to each grid point, the evaluation function E _p does not become a small value. Accordingly, in step S420, the optimization unit LM4 optimizes each lattice point so as to minimize the evaluation function E _p . Which term is to be optimized is determined by the above-described weighting factors w _{1 to} w ₅ . Therefore, it is desirable to set which item should be emphasized when creating the ink profile IP, and to set the weighting factors w _{1 to} w ₅ based on it. For example, if it is desired to create an ink profile IP with good running cost even at the expense of image quality, the weight coefficient w ₅ should be set large. In addition, it is desirable to increase the weighting factor w ₄ for the grid points that should be located on the outer surface of the color reproduction gamut to increase the constraint on the outer surface.

In particular, in step S420, the ink amount sets which minimize the evaluation function E _p for each grid point _{(d c, d m, d} y, d k, d lc, d lm) are sequentially calculated. For example, locally moving the ink amount set from the position of the initial ink amount set in the ink amount space, we calculate for each grid point an ink amount sets which minimize the evaluation function E _p at that time. Thus, the position of the lattice points in the ink amount space is modified in a direction that minimizes the evaluation function E _p. Furthermore, likewise locally moving the ink amount sets the position of the corrected, will calculated for each grid point an ink amount sets which minimize the evaluation function E _p at that time. By repeating the aforementioned processing (for example, 200 times), and finally it is possible to optimize the lattice point evaluation function E _p becomes extremely small for each grid point. It should be noted that the optimization of the grid point may be completed by performing the above process a specified number of times, or the optimization of the grid point may be completed when the value of the evaluation function E _p falls below a predetermined threshold value.

In this optimization process, it is necessary to calculate the evaluation function E _p for the ink amount sets (d _c , d _m , d _y , d _k , d _lc , d _lm ) that are sequentially updated. each converter RC mentioned above, CC, by utilizing the GC and smoothness calculator LM3, successively, corresponding to the respective ink amount sets _{(d c, d m, d} y, d k, d lc, d lm) The spectral reflectance R (λ), the graininess index GI, the color constancy index CII, and the smoothness evaluation index SI are calculated. According to the above optimization, evaluation function E _p ink amount set which is excellent in graininess, color constancy and running costs by _{(d c, d m, d} y, d k, d lc, d lm) lattice points of At the same time, the distribution of the grid points in the CIELAB color space is also optimized. This is because the evaluation in the CIELAB color space is also incorporated into a part of the evaluation function E _p in the third and fourth terms of Expression (16). In the present embodiment, since the optimization of the grid points in the CIELAB color space and the optimization of the grid points in the ink amount space can be performed simultaneously, the processing efficiency is high. In the present embodiment, the grid point optimization technique disclosed in Japanese Patent Application Laid-Open No. 2006-197080 can also be applied. In this case, a virtual force in the direction in which the evaluation function E _p is 0 is applied to each lattice point in the ink amount space, and the position of the lattice point in the ink amount space is converged to a steady state by the force. .

As each grid point as described above is optimized, corresponding to the ink amount set of the optimized lattice points at step _{S430 (d c, d m,} d y, d k, d lc, d lm) The calculated L ^* a ^* b ^* values are calculated by the spectral printing model converter RC and the color converter CC (D65 light). Then, to each other corresponding L ^* a ^* b ^* values and the ink amount set _{(d c, d m, d} y, d k, d lc, d lm) ink profile IP that describes the correspondence between the LUT generator LM5 Create.

In step S440, a color conversion profile CP is created based on the ink profile IP. The color conversion profile CP is an LUT that converts, for example, image data in which the color of each pixel is expressed in the sRGB color space into image data in the ink amount space in the printer 20. Since the sRGB color space has a corresponding relationship (sRGB profile SP) with the CIELAB color space based on the CIE standard, the sRGB color space is determined by the L ^* a ^* b ^* values of each grid point defined in the ink profile IP. RGB values and the ink amount sets _{(d c, d m, d} y, d k, d lc, d lm) identifies a correspondence between, it is possible to LUT of. At that time, interpolation processing is performed, but since the distribution in the CIELAB color space of the grid points defined by the ink profile IP is smoothed by the above-described optimization, high interpolation accuracy can be realized. It is desirable that the sRGB profile SP is also optimized by the above-described smoothness degree evaluation index SI (see Japanese Patent Application Laid-Open No. 2006-197080). Since the gamut of the sRGB color space in the CIELAB color space and the color reproduction gamut of the printer 20 are different, gamut mapping is appropriately performed. Here, an example in which an sRGB color space, which is another absolute color space, is combined with the ink profile IP is taken as an example, but a device-dependent source color space that depends on the input device and the ink profile IP are combined. A device link profile may be created. When the color conversion profile CP can be created, the color conversion unit CM refers to the color conversion profile CP and sets the print image data representing the color of each pixel in the sRGB color space as an ink amount set (d _c , d _m , D _y , d _k , d _lc , d _lm ). Furthermore, the ink amount sets _{(d c, d m, d} y, d k, d lc, d lm) than that sequentially executes a halftone process or microweave processing on the image data of, generating print data Can do. By outputting the generated print data to the printer 120, printing based on the print image data can be executed.

C. Summary and variations:
As described in section A, in the present invention, in the first stage, printing of the primary color gradation patch, measurement of the spectral reflectance R (λ), and the color patch only for a small number of first representative points are used. Printing and spectral reflectance R (λ) are measured. After that, the spectral reflection of the second representative point is obtained by the neural network NNR in which the spectral reflectance of the gradation patch of the primary color and a part of the stable reflectance of the first representative point are constructed as learning data CD1. Predict the rate R (λ). Thus, it is possible to obtain a spectral reflectance R (λ) of a node (first representative point + second representative point) sufficient for prediction of the spectral reflectance R (λ) by the cell division Yule-Nielsen spectral Neugebauer model. it can. Since the color patch printing and the spectral reflectance R (λ) are measured only for the primary color gradation and a small number of first representative points, the workability of printing / measurement can be improved, and an efficient printing result can be obtained. Prediction and LUT creation can be realized.

  In the above description, the spectral reflectance is predicted. However, it is also possible to predict other printing results by each process of the present invention. For example, the color as the printing result may be predicted. When the spectral reflectance R (λ) of the first representative point is obtained, the color under the desired light source is specified by the color converter CC, and a neural network is created using the specified color as learning data. Then, the color of the second representative point is specified by the neural network, an LUT defining the correspondence between the first representative point and the second representative point and the color is created, and further, interpolation prediction by the LUT (second prediction step) It is also possible to predict a color corresponding to an arbitrary ink amount set. In this embodiment, the graininess index GI is predicted only by the neural network NNG. However, the graininess index GI is predicted by two-stage prediction, and the amount of ink for graininess evaluation that is actually printed / scanned. It is also possible to reduce the number of sets. Of course, color constancy CII can also be predicted by the method of the present invention.

  Further, the prediction method at each prediction stage is not limited to the neural network or the cell division Yule-Nielsen spectroscopic Neugebauer model. Of course, when the spectral reflectance R (λ) is predicted, the spectral reflectance R (λ) of the second representative point is first predicted by the cell division Yule-Nielsen spectral Neugebauer model, and then an arbitrary ink amount set is set. May be predicted based on the spectral reflectances R (λ) of the first representative point and the second representative point.

  Furthermore, in the above-described embodiment, the print result is predicted by the computer 10 connected to the printer 20, but the print result may be predicted by the printer 20 or another device. If the printer 20 can predict the print result of itself, the printer 20 alone can create a color conversion profile or the like according to the usage status of the printer 20. Further, as in the above-described embodiment, when the computer 10 that has prepared the converters CR, CC, and CG is different from the computer 110 that creates the color conversion profile CP, the spectral reflectance data D prepared by the computer 10 is obtained. For example, it may be distributed via the Internet to a computer 110 or a printer having no spectral reflectance measurement environment. As a result, the computer 110 or printer used by a general user who has received the spectral reflectance data D prepared by the computer 10 can create the color conversion profile CP.

Furthermore, the evaluation function E _p set when creating the color conversion profile CP is merely an example, and other evaluation functions E _p can be used. For example, an evaluation function E _p having only the smoothness degree evaluation index SI may be used. Furthermore, an index for evaluating image quality other than the color constancy index CII of the graininess index GI may be incorporated as an index for evaluating the image quality. Further, as in the pamphlet of International Publication WO 2005/043884, image quality optimization and grid point arrangement smoothing may be performed at different stages. Even in this case, the present invention can be applied because the spectral reflectance R (λ) is predicted when printing with an arbitrary ink amount set using the spectral printing model converter RC.

  The ink set used for printing is not limited to the above-described CMYKlclm. For example, spectral reflectance when printing with O (orange) ink, R (red) ink, V (violet) ink, G (green) ink, lk, llk (light black) ink, DY (dark yellow) ink, etc. R (λ) can also be predicted by the method of the present invention. Since each of these inks has a specific spectral reflectance specification, the spectral reflectance R (λ) can be predicted according to the ink amount set. In printing using many ink types, the number of nodes of the cell division Yule-Nielsen spectral Neugebauer model necessary for prediction of the spectral reflectance R (λ) increases in a multiplier manner. The effect is great.

It is a block diagram which shows the structure of the printing result prediction apparatus of this invention. It is a flowchart of a spectral printing model converter preparation process. It is the figure which showed the primary color typically taking CMY space as an example. It is a figure explaining a mode that a 1st representative point and a 2nd representative point are set. It is a figure which shows spectral reflectance data. It is a figure explaining the neural network NNR. It is a figure which shows a spectral Neugebauer model. It is a figure which shows a cell division | segmentation Yule-Nielsen spectroscopic Neugebauer model. It is a figure which shows a mode that a color is specified from a spectral reflectance. It is a flowchart of a granularity converter preparation process. It is a figure which shows the process which calculates the granularity index | exponent GI. It is a figure explaining the neural network NNG. It is a flowchart of a neural network creation process. It is a block diagram which shows the structure of the lookup table preparation apparatus. It is a flowchart of a profile creation process. It is a figure explaining a smoothness degree evaluation index. It is a graph which shows the color reproduction gamut of a printer.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 10,110 ... Computer, 11, 111 ... HDD, 12, 112 ... CPU, 20 ... Printer, 30 ... Spectral reflectometer, 40 ... Scanner, RC ... Spectral printing converter, CC ... Color converter, GC ... Graininess converter, D: Spectral reflectance data, CD1, CD2: Learning data.

Claims (17)

A printing result prediction method for predicting a printing result formed by adhering a plurality of types of ink on a recording medium,
A first measurement step of performing printing with a plurality of ink amount sets that make gradation with respect to the ink amount for each ink, and measuring the print result;
A first setting step of setting a plurality of representative coordinates in the ink amount space as first representative points;
Printing with an ink amount set corresponding to the first representative point, and measuring a printing result;
A second setting step of setting, as the second representative point, another representative coordinate different from the first representative point in the ink amount space;
A first prediction step of predicting a print result printed with the ink amount set corresponding to the second representative point based on the measurement results in the first measurement step and the second measurement step;
A second prediction step of predicting a print result printed with an arbitrary ink amount set based on the measurement result in the first measurement step, the measurement result in the second measurement step, and the prediction result in the first prediction step; ,
A printing result prediction method comprising:   The printing result prediction method according to claim 1, wherein the plurality of gradation ink amount sets are created so that color differences are substantially uniform.   The printing result prediction method according to claim 1, wherein the first representative point is set so as not to include coordinates measured in the first measurement step.   The printing method prediction method according to claim 1, wherein prediction methods used in the first prediction step and the second prediction step are different from each other.   The prediction method used in the first prediction step uses a neural network in which learning is performed based on the measurement results of the first measurement step and the second measurement step. Item 5. The printing result prediction method according to any one of Items4. In the first setting step, lattice points that are uniformly distributed in the ink amount space are generated, and the first representative point is selected from the lattice points.
The first representative point is selected so that the ink amount set to be used by the printing apparatus has a higher density in the vicinity of the outer edge of the effective area distributed in the ink amount space than in the vicinity of the center of the effective area. The printing result prediction method according to claim 1, wherein the printing result is predicted.   The printing result prediction method according to claim 6, wherein a ratio of the number of the first representative points and the number of the second representative points inside the effective area is about 1: 3.   The print result prediction method according to claim 1, wherein the print result is a spectral reflectance on the recording medium.   The color specifying step of specifying a color when a desired light source is irradiated on the printing result, which is the spectral reflectance on the recording medium, is further provided. The printing result prediction method according to the item. A method for creating a look-up table that defines the correspondence between a representative ink amount set in an ink amount space centered on the ink amount for a plurality of types of ink and the colors when printing is performed using the ink amount set Because
Printing is performed with a plurality of ink amount sets that make gradation with respect to the ink amount for each ink, and the spectral reflectance of the printing result is measured. The spectral reflectance when printing is performed with the ink amount set corresponding to the first representative point which is a representative coordinate of the first representative point, and in another representative coordinate different from the first representative point in the ink amount space. Spectral reflectance when printed with an ink amount set corresponding to a second representative point is predicted based on the spectral reflectance obtained by both the above measurements, and further printed with an arbitrary ink amount set. Spectral reflectance prediction step of predicting the rate based on the spectral reflectance obtained by the both measurements and the spectral reflectance obtained by the prediction,
A color specifying step for specifying a color when a desired light source is irradiated with respect to the spectral reflectance predicted in the spectral reflectance prediction step;
The color when printing with a plurality of representative LUT ink amount sets generated in the ink amount space and irradiating a desired light source is specified by the spectral reflectance prediction step and the color specifying step, and the color Creating a lookup table that defines the correspondence between the LUT ink amount set and the LUT ink amount set;
A method for creating a lookup table, comprising:   In the creating step, the LUT ink amount set is optimized so that the degree of smoothness of the distribution in the color space of the color acquired by the spectral reflectance prediction step and the color specifying step is increased. The creation method according to claim 10.   12. The creation method according to claim 11, wherein the LUT ink amount set is optimized so that image quality is improved when printing is performed with the LUT ink amount set.   At least one of the image quality is obtained by acquiring the spectral reflectance when printing with the LUT ink amount set by the spectral reflectance prediction step, and at least two types of the spectral reflectance in the color specifying step. The creation method according to claim 12, wherein the color difference is when the light source is irradiated.   The creation method according to claim 12 or 13, wherein at least one of the image quality is graininess when printed with the LUT ink amount set.   15. The LUT ink amount set is optimized so that a total of the ink amounts constituting the LUT ink amount set satisfies a predetermined condition. The creation method described in the section.   A lookup table created by the lookup table creation method according to any one of claims 10 to 15.   A print control apparatus that performs color conversion of print image data using the lookup table created by the lookup table creation method according to any one of claims 10 to 15.

Images