JP2003066539A - Color management system and method in photograph processor - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a color management system and a method for a photograph processor and capable of efficiently preparing a profile. SOLUTION: The color management system in the photographing processor for reading out an image formed on a positive film with a scanner 1 and preparing a photographing print by suitable coloring on the basis of the read picture data, is provided with a CCD scanner 1 for reading out a reference chart IT8 expressing a plurality of colors and a profile preparation device 20 for preparing a profile for converting the value of read picture data of each color into an objective reference value. The profile preparation device 20 is constituted of a feedforward hierarchical type neural network 22 based on a reverse propagation method of errors and learning is executed so that an input value approaches a reference value.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a color management system and method in a photographic processing device for reading an image on a photographic film by a reading device and producing a photographic print with an appropriate color based on the read image data. Regarding

[0002]

2. Description of the Related Art Such a photographic processing apparatus, for example, reads a frame image formed on a developed positive film by a CCD scanner (corresponding to a reading device) and operates a digital exposure engine based on the read image data. Is used to bake and expose the image. A photographic print can be obtained by developing and drying this paper.

However, even if a photographic print is made by using the value of the image data obtained by the CCD scanner as it is, a color shift occurs when compared with the image formed on the positive film. This color shift is caused by the fact that the sensitivity characteristics of the input device (CCD scanner, etc.), output device (digital exposure engine, etc.), and image forming medium (positive film) cannot be given in a simple form like a linear function. . That is, if the same image data value (R,
This is because even if G and B values are given, they are not represented by the same color because the sensitivity characteristics of the input device and the output device are different.

Therefore, a technique for making the colors of the input device and the output device as close as possible is known, and this technique is called color management. Therefore, the color management incorporated in the photo processing apparatus will be described with reference to FIG.

In FIG. 1, an image reading device is shown as an input device, and an image display device (monitor) and an image printing device (printer) are shown as output devices. A table (look-up table) representing one color for each color in the standard color gamut (CIE Lab) is called a profile. In this way, by expressing the characteristics of each device with a profile using the standard color gamut as a reference, it is possible to reproduce an appropriate color. The present invention particularly relates to a color management system for creating a profile interposed between an image reading device and a standard color gamut.

First, there is a standard chart (IT8) as a standard chart (calibration target) showing a standard color gamut. This standard chart is based on an input color target (Color Reflection Target).
t for Input Scanner Calib
has a formal name. As shown in FIG. 2, IT8 is a standard chart in which the vertical axis is A to L and the horizontal axis is 1 to 22 divided into a total of 264 colors. This standard chart is read by the image reading device. For example, 2
It is assumed that the value when a specific color of the 64 colors is read is (Xr, Yr, Zr). On the other hand, the above 26
For each of the four colors, a reference value (target value) when read is set in advance. If the reference value of the specific color is (Xt, Yt, Zt), the read value (Xr, Yr, Zr) is the reference value (Xt, Yt, Zt).
The table (profile) is created so that

[0007]

However, creating the above table complicates the work of the operator, and the labor for that is also great, and improvement is desired. For example, an 8-bit image has 16.7 million colors.
It is impossible to create a table for 6.7 million colors. In color management it is important to create a profile that can be applied for all colors with as little information as possible.

The present invention has been made in view of the above circumstances, and an object thereof is to provide a color management system and method for a photographic processing apparatus capable of efficiently creating a profile.

In order to solve the above-mentioned problems, a color management system of a photographic processing apparatus according to the present invention reads an image formed on an image forming medium by an image reading device, and based on the read image data. A color management system in a photographic processing device for creating photographic prints with appropriate colors, the image reading device reading a standard chart prepared in advance and expressing a plurality of colors, and the read color of each color. A profile creating device for creating a profile for converting a value of image data into a target reference value, wherein the profile creating device is configured by a neural network by an error back propagation method, and an input value is the reference value. It is characterized in that it is configured to perform learning so as to approach.

The operation and effect of the system with this configuration are as follows.
It is as follows. First, the standard chart is read by the image reading device. As this standard chart, for example,
IT8 is selected. A profile creation device is created that creates a profile for converting the read image data value of each color into a target reference value. Then, this profile creating device is configured by a neural network by the error back propagation method. The neural network is a technology that is constructed by using an artificial element that imitates a neuron of a living body, and is a technology that has received a great deal of attention in recent years.

264 colors are specified for the above-mentioned IT8. On the other hand, when colors are represented by 8-bit image data, the number of colors is 16.7 million.
It can be estimated by a neural network.

Here, values such as (R, G, B) and (X, Y, Z) are selected as input values of the neural network, and learning is performed so that these input values approach the reference value. As a result of diligent study, the present inventors have been able to confirm that a neural network can be used to secure the same level of quality as that of a profile based on a lookup table that has been conventionally performed. Further, learning by the neural network can be automatically performed by software. Therefore, profile creation can be performed more efficiently than when creating a table. as a result,
It is possible to provide a color management system for a photographic processing device that can efficiently create a profile. As a preferred embodiment of the present invention, the profile creation device defines a value obtained by dividing an error value of an output value with respect to the reference value by the reference value as an error rate,
An example is one configured to perform learning based on the error rate. Regarding the color misregistration, the magnitude of the numerical error does not necessarily match the magnitude of the apparent error. In particular, for a color with low lightness, the numerical value as data is small and it is easily affected by an error. Therefore, even if the magnitudes of the errors are exactly the same, the color shift with lower lightness is larger than the color shift with higher lightness. Therefore, a value obtained by dividing the error value of the output value with respect to the reference value by the reference value is defined as an error rate, and learning is performed based on this error rate. As a result, learning can be converged in the same way as in the case of a color having a low lightness, as in the case of a color having a high lightness.

As another preferred embodiment of the present invention, in performing the learning, the input value is classified into a plurality of categories on chromaticity coordinates, and the learning is performed for each of the categories. I can give you what you have.

For example, when IT8 is used as a standard chart, if all 264 colors are learned at once,
The error between the output value and the reference value becomes large. Therefore, it may take time for the learning to converge. Therefore, the chromaticity coordinates (for example, xy chromaticity coordinates) are classified into a plurality of categories, and learning is performed for each category. Thereby, learning can be performed efficiently.

As still another preferred embodiment of the present invention,
In each of the above-mentioned categories, a boundary portion is set to overlap, input values near the boundary are overlapped and learned, and learning results in both categories are averaged at a predetermined ratio.

In the case of classifying into a plurality of categories, if learning is performed without overlapping the boundary part of each category, there is a possibility that discontinuity occurs in the learning result near the boundary of the category. Therefore, the boundary portion of the category is set redundantly.
Then, the input values in the vicinity of the boundary of this overlapping portion are learned by overlapping. The learning results in both overlapping categories are averaged at a predetermined ratio. This makes it possible to maintain the continuity of learning results near the boundaries of categories.

As still another preferred embodiment of the present invention,
The predetermined ratio may be set based on the radial length of a circle centered on the point of saturation 0 on the chromaticity coordinate or the size of the central angle. The radial direction of the circle centered on the point of saturation 0 is along the saturation direction, and the size of the central angle is the size of the angle along the circumferential direction of the circle, that is, the size along the hue direction. Corresponding to Therefore, by setting a predetermined ratio based on the length in the radial direction and the size of the central angle, it is possible to maintain the continuity of the learning result with high accuracy.

[0017]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS A photographic processing apparatus in which a color management system according to the present invention is preferably used will be described with reference to FIG.

In FIG. 3, a CCD scanner 1 functioning as an image reading device is provided. The CCD scanner 1 reads an image on a positive film F (an example of a photographic film) which is an image forming medium, and measures the intensity of transmitted light, which is the light emitted from the reading light source 2 and transmitted through the positive film F. Thus, the image data can be acquired. The positive film F is conveyed by the conveying roller 3 at a constant speed, and the image data is taken in line by line. Therefore, although the CCD scanner 1 is a line sensor, it may be an area sensor instead.

The image processing section 2 processes the acquired image data and transfers it to the exposure engine 3. This image processing unit 2 is provided with a profile creation device 20 according to the present invention.
It is equipped with a part that functions as.

The exposure engine 3 outputs the exposure light to the paper P.
By irradiating (corresponding to a photographic light-sensitive material),
An image is printed and exposed on the emulsion side of the paper P. Paper P
Are stored in a roll shape in the paper magazine 4.
The paper P pulled out from the paper magazine 4 is cut into a print size by the paper cutter 5.
The paper P cut into the print size is conveyed at a constant speed by the exposure conveyance roller 6, and the image is sequentially printed and exposed by the exposure engine 3. The paper P on which the image is printed and exposed is subjected to predetermined processing in the development processing section 7 and the drying processing section 8 to produce a photographic print.

Further, a controller 9 for controlling each section in the photo processing apparatus is provided. A keyboard 10 for inputting various data and a monitor 11 for displaying a read image and the like are connected to the controller 9.

<Regarding Profile Creating Device> Next,
A profile creation device used in the color management system of the present invention will be described. As already described in FIG. 1, it is necessary to create a profile representing the sensitivity characteristic of the CCD scanner. This profile is created using a feedforward hierarchical neural network based on the backpropagation method. A neural network is constructed using artificial elements that simulate the neurons of a living body.
This is a technology that has received a great deal of attention in recent years. The structure of the neural network is conceptually shown in FIG.

The hierarchical type refers to a network in which an arbitrary number of intermediate layers are present between the input layer and the output layer. Two intermediate layers are shown in the illustrated example. Feedforward is a network in which there is no coupling between units in the same layer and all units can receive input only from the previous layer and send output only to the next layer. That is, the signal is configured to flow in one direction from the input side to the output side. The error back-propagation method is one of learning methods for a feedforward hierarchical neural network. That is, the relationship between the input value and the output value is changed by changing the connection weight or the threshold value of the units forming the network. Appropriate initial values are given to parameters (coupling weights and threshold values) inside the network, and then small adjustments to the parameters are repeated so that the input / output error becomes small.
As a result of this iterative process, learning is performed so that the output value converges (approaches) to the reference value.

The handling of color data described in the present invention will be described. The image data captured by the CCD scanner 1 can be represented by each value of (B, G, R). In order to express each value of (B, G, R) by the sensitivity seen by human eyes, each value of (X, Y, Z) may be used. Here, (B, G, R) and (X, Y,
There is a relationship between Z) as in the following equations 1 and 2.

[0025]

[Equation 1]

[Equation 2] Also, in order to represent the color of an object that cannot be represented in XYZ, the lowercase quantities xy are used. This is represented by a ratio in which (X, Y, Z) is 1 when summed, and z is omitted. The xy chromaticity coordinates are shown in FIG. 5, which is called the CIE xy chromaticity diagram. The chromaticity diagram has a bell shape, and the CIE Lab chromaticity diagram shown in FIG. 6 is obtained by correcting the bell shape.

Next, the block configuration of the image processing unit 2 will be described with reference to FIG. The CCD scanner 1 originally reads the image on the image forming medium, but reads the image of the standard chart IT8 when performing learning by the neural network. The standard chart IT8 represents a standard color gamut, and represents 264 colors as shown in FIG. The conversion unit 21 converts (B, G, R) of the image data read by the CCD scanner 1 into (X,
Y, Z). Further, a cube root (X ^1/3 , Y ^1/3 , Z ^1/3 ) is calculated from this (X, Y, Z).
For convenience, (X ^1/3 , Y ^1/3 , Z ^1/3 ) is ^replaced by (X, Y,
Z) Expressed as ^1/3 . The value of the cube root is used because it matches well with the sensitivity curve of the human eye. In the case of a positive film, it is said that the values obtained by raising the cube root match well.

It should be noted that (B, G,
R) image data as it is in the neural network 2
2 may be input to the inside of the neural network 22 so that the conversion to (X, Y, Z) and the conversion to the cube root are performed.

The neural network 22 has 3 inputs and 3 inputs.
It is an output feedforward hierarchical neural network. Both the input value and the output value are expressed by (X, Y, Z) ^1/3 . On the other hand, 26 represented in the standard chart IT8
A reference value (also called a target value or a teacher value) is preset for each of the four colors. Based on the difference between the reference value (X, Y, Z) ^1/3 and the output value (X, Y, Z) ^1/3 , learning by the error back propagation method is performed.

The cube operation unit 23 performs the operation of raising (X, Y, Z) ^1/3 to the cube and returning it to the original (X, Y, Z). The conversion unit 24 converts (X, Y, Z) into (B, G, R). Further, γ correction is performed by the γ correction unit 25, and this image data is transferred to the exposure engine.

<Learning by Neural Network> Next, learning by the neural network 22 will be described in detail. First, the standard chart IT8 is read by the CCD scanner 1. In addition, this standard chart IT
8 is defined by the standard. For example, ANS
It is defined in the I standard. As described above, the reference value is preset for each of the 264 colors shown in the standard chart IT8. This reference value data needs to be input to the storage device in the image processing unit 2 in advance.
Moreover, this reference value is given as a value of (X, Y, Z). Note that in the present embodiment, since the evaluation is performed with the value rooted to the third power, the value of (X, Y, Z) ^1/3 is input and stored in the storage device in advance.

Next, learning is carried out by inputting (X, Y, Z) ^1/3 or (R, G, B) to the neural network 22, and the learning is performed at once for the 264 colors represented in the standard chart IT8. When learning, the amount of calculation becomes enormous and learning takes time. Therefore, in the present invention, instead of learning the 264 colors at once, it is divided into 10 categories (= groups) and learning is performed for each category. This will be described.

In the CIE xy chromaticity diagram of FIG. 5, each category is expressed using xy coordinates as follows.

Category 0: x <(1/3) 4x−1 <y <(1/2) x + (1/6) 1: x <(1/3) x <y <−x + (2/3) 2: x <(1/3), y> (1/3) x ≧ (1/3), 2x− (1/3) <y 3: x ≧ (1/3) (1/2) x + ( 1/6) <y 4: x ≧ (1/3) − (1/20) x + (7/20) <y <x 5: x ≧ (1/3) −x + (2/3) <y < (1/10) x + (3/10) 6: x ≧ (1/3) y <− (1/2) x + (1/2) 7: x ≧ (1/3), y <− (5 / 2) x + 7/6 x <(1/3), y <(23/10) x− (13/30) 8: (x− (1/3)) ² + (y− (1/3)) ² ≦ 0.00006 9: (x− (1/3)) ² + (y− (1/3)) ² ≦ 0.00004 The categories shown above are illustrated. When expressed in CIE Lab of No. 6, it is shown as follows. The angle display is (x
= 1/3, y = 1/3) (point of saturation 0) is displayed as the origin.

Category 0: 206.6 degrees to 256.0 degrees (width 50.0 degrees) 1: 135.0 degrees to 225.0 degrees (width 90.0 degrees) 2: 63.4 degrees to 180.0 Degree (width 116.0 degrees) 3: 26.6 degrees to 90.0 degrees (width 63.4 degrees) 4: -2.9 degrees to 45.0 degrees (width 47.9 degrees) 5: -45. 0 degree-5.7 degree (width 50.7 degree) 6: 270.0 degree-333.4 degree (width 63.4 degree) 7: 246.5 degree-291.8 degree (width 45.3 degree) 8: Within radius 0.00006 9: Within radius 0.00004 As described above, learning is performed by applying the neural network to each category. This speeds up learning
And, it will be possible to do it with high accuracy.

Further, in the categories described above, the area of the region is not uniform but varies. It is easy to understand the width of the categories expressed on the CIE Lab chromaticity coordinates, but the widths of the categories are not the same. This is because when the 264 colors of the standard chart IT8 are divided into 10 categories, it is adjusted so that each category has approximately the same number of colors (patches). This is because the learning time of each category can be equalized, and as a result, the learning time can be shortened.

Further, in each category, the area is not completely divided, but the boundary portion is set to overlap. This is because if the categories are set without setting the overlapping part, discontinuity occurs in the learning result near the boundaries between the categories. Therefore, the overlapping part is set in the category, and the colors of the overlapping part (near the boundary) are learned in both categories. The learning results in both categories are averaged (mixed) at a predetermined ratio.
I am trying to do it. This will be described with reference to FIG. For data belonging to both category (1) and category (2), the mixture ratio is set based on the angle θ (center angle) viewed from the point of saturation 0 (the center of the circle). This point will be described. The radius direction of the circle is along the saturation direction. Further, the direction of the central angle is the direction along the circumferential direction and the hue direction.

First, the range of category (1) is θ _1a <θ <
_Set to θ _1b . Then, learning is performed by the neural network, and the output in category (1) is d ₁ = d ₁ (r, θ)
Suppose that it was obtained in. In addition, the range of category (2) is θ _2a
<Θ <θ _2b . Then, learning is performed by the neural network, and the output in category (2) is d ₂ = d ₂
It is assumed that it is obtained by (r, θ). In this case, d ₁
Assuming that the mixing ratio of (1) and (d ₂ ) is (1−t): t, t is determined so that t = (θ−θ _2a ) / (θ _1b −θ). As a result, category (1)
And the output of category (2) can be continuously connected.

On the other hand, with respect to the low saturation data belonging to the categories (8) and (9), the mixture ratio can be determined based on the size of the circle in the radial direction. For example,
It can be averaged according to the ratio of the radius magnitudes.
By averaging in this way, discontinuity in the learning result can be eliminated.

FIG. 9 shows a comparison of the results with and without overlap learning near the category boundary. The top of FIG.
This is the result when there is no overlap learning near the category boundary, and discontinuity is seen at the category boundary. On the other hand, the lower part of FIG. 9 shows the results when there is overlapping learning near the category boundaries, and it can be seen that the learning results show continuity.

When data is input to the neural network 22, the first output data which has not been learned at all is the same as the input data. The output data output value and the reference value are compared with each other. In the present invention, learning is performed by introducing the concept of an error rate. The error rate is calculated by the following formula.

Error rate = (reference value-output value) / reference value In other words, the error rate is expressed as an absolute error rather than a relative error due to a simple (reference value-output value). Regarding the color misregistration, the magnitude of the numerical error does not necessarily match the magnitude of the apparent error. In particular, for a color with low lightness, the numerical value as data is small and it is easily affected by an error. Therefore, even if the magnitudes of the errors are exactly the same, the color shift with lower lightness is larger than the color shift with higher lightness. Therefore, the error rate is defined as described above, and learning is performed based on this error rate. As a result, learning can be converged in the same way as in the case of a color having a low lightness, as in the case of a color having a high lightness.

Although learning is performed based on the error rate, in the present invention, learning is performed based on the error back propagation method as already described. Here, when the back propagation is performed from the output layer, an operation according to the error rate is performed.
In practice, the allowable value is determined in advance, and when the error rate does not exceed it, it is halved and then propagated back to reduce the influence on the update of the unit coupling weight.

As described above, the neural network 22
Is used to perform learning until the output value reaches a predetermined level. To check the learning result, follow the procedure below.

First, the print P1 is created using the learned data. With this image data, (X, Y, Z) ^1/3 output from the neural network 22 is cubed by the cube calculation unit 23 to calculate (X, Y, Z). This is further converted by the conversion unit 24 into (B, G, R) image data. This is γ-corrected by the γ-correction unit 25, then transferred to the exposure engine and the image is printed and exposed on the paper to obtain the photographic print P1. On the other hand, using the standard value (X, Y, Z) ^1/3 as it is,
Multiply operation, (R, G, B) conversion, and γ correction are performed to obtain a photographic print P2. The learning result can be confirmed by comparing the photographic prints P1 and P2. As a result of actually making and comparing the photographic prints P1 and P2, the present inventors were able to confirm that the degree of coincidence is indistinguishable as far as human eyes can see. . That is, it was confirmed that the learning by the neural network 22 was sufficiently effective and practical.

<Other Embodiments> (1) In this embodiment, (B, G, R) is replaced by (X, Y,
However, the present invention is not limited to this.
Also, as the standard chart, a chart other than IT8 may be used. Further, although learning is performed by (X, Y, Z) ^1/3 , the learning is not limited to this. For example, learning may be performed using a logarithmically compressed numerical value.

(2) The image forming medium is not limited to a positive film, and the present invention can be applied to other photographic films such as negative films, printed matter, and prints.

(3) The number of categories can be set arbitrarily. The degree of overlap can also be set as appropriate. (4) The profile between the output device and the standard color gamut (CIE Lab) can be similarly created by the neural network. (5) The color management system according to the present invention can also be applied to printer calibration.

[Brief description of drawings]

[Figure 1] Diagram showing the concept of color management

FIG. 2 is a diagram showing a standard chart IT8.

FIG. 3 is a schematic diagram showing the configuration of a photographic processing device.

FIG. 4 is a diagram showing a configuration of a neural network.

FIG. 5 is a diagram showing CIE xy chromaticity coordinates.

FIG. 6 is a diagram showing CIE Lab chromaticity coordinates.

FIG. 7 is a diagram showing the configuration of a profile creation device.

FIG. 8 is a diagram illustrating a mixture ratio between adjacent categories.

FIG. 9 is a diagram for comparing results with and without overlapping learning.

[Explanation of symbols]

1 CCD scanner 2 Image processing unit 20 Profile creation device 21 Converter 22 Neural network 23 3rd power calculator 24 Converter 25 γ correction unit

   ─────────────────────────────────────────────────── ─── Continued front page    (72) Inventor Koji Nitta             No. 1 at 579 Umehara, Wakayama City, Wakayama Prefecture             Inside Ritsu Koki Co., Ltd. (72) Inventor Koichi Kugo             No. 1 at 579 Umehara, Wakayama City, Wakayama Prefecture             Inside Ritsu Koki Co., Ltd. (72) Inventor Koji Kita             No. 1 at 579 Umehara, Wakayama City, Wakayama Prefecture             Inside Ritsu Koki Co., Ltd. F term (reference) 2H106 AA72 AA80 AA85 BA27 BA52                       BA55 BA64                 5B057 AA11 BA02 CE18 CH01 CH07                       CH08                 5C077 LL16 LL17 MM27 MP08 PP31                       PP32 PP36 PP46 PQ15 PQ23                       SS01 TT02                 5C079 HB01 HB05 HB08 HB11 MA04                       MA10 MA13 NA13 PA03 PA08

Claims (6)

[Claims] 1. A color management system in a photographic processing device for reading an image formed on an image forming medium by an image reading device and creating a photographic print with an appropriate color based on the read image data. And an image reading device that reads a standard chart that is prepared in advance and expresses a plurality of colors, and a profile creation device that creates a profile for converting the read image data value of each color into a target reference value. The profile creating apparatus is configured by a neural network by an error backpropagation method, and is configured to perform learning so as to bring an input value closer to the reference value. system. 2. The profile creating device defines a value obtained by dividing an error value of an output value with respect to the reference value by the reference value as an error rate, and is configured to perform learning based on the error rate. The color management system in the photographic processing apparatus according to claim 1. 3. When performing the learning, the input values are classified into a plurality of categories on chromaticity coordinates, and the learning is performed for each of the categories. Alternatively, the color management system in the photographic processing device according to item 2. 4. The boundaries of each category are set to overlap each other, and input values near the boundaries are overlapped and learned, and learning results in both categories are averaged at a predetermined ratio. The color management system in a photographic processing apparatus according to claim 3, wherein 5. The predetermined ratio is set based on a radial length of a circle centered on a point of saturation 0 on the chromaticity coordinate or a size of a central angle. The color management system in a photographic processing apparatus according to claim 4. 6. A color management method in a photographic processing device for reading an image of a photographic film with a reading device and creating a photographic print with an appropriate color based on the read image data, which is prepared in advance. A step of reading a standard chart representing a plurality of colors, and a step of creating a profile for converting the read image data value of each color into a target reference value, and creating the profile. In the step, a color management method in a photographic processing apparatus, wherein learning is performed so as to bring an input value closer to the reference value by using a neural network by an error back propagation method.