JP3044915B2 - Color correction and color coordinate conversion method using neural network - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a color correction and color coordinate conversion method for inputting / outputting an image from a color scanner, a color printer, a CRT display, a color plotter or the like.

With the recent spread of color image input / output devices, devices for electronically handling color information, such as color scanners, color printers, CRT displays, and color plotters, are increasing. Since it is difficult to eliminate color distortion in input / output devices to be handled, a color correction device capable of performing accurate color correction in real time is required.

In addition, as a method of expressing colors, red (R), green (G), and blue (B) light are added in display on a CRT display, and cyan (C) is used in the field of printing. , Magenta (M), yellow (Y) using the reflected light of opaque color material using reflected light is used, but it is possible to print what is displayed, or to display what is printed, Also in this case, it is necessary to perform accurate color correction and the color coordinate conversion in real time.

[0004]

2. Description of the Related Art FIG. 6 is a diagram for explaining a conventional color correction method. FIGS. 6 (a1) to 6 (a3) show examples of a masking method, and FIG. 6 (b) uses a neural network. Shows the case.

[0005] As described above, there are an additive color expression and a subtractive color expression as color expression methods. The additive color expression is three colors of light ｛red
(R), green (G), blue (B)｝ is a method of expressing colors by superimposition.The subtractive color expression is three color materials ｛cyan (C), magenta (M), yellow (Y) This is a method of expressing color by superimposing the reflected light of｝.

In general, on a computer, for example, on a CRT display, the above-described red (R), green (G), and blue (B) lights are added, and in a printed matter, cyan (C), magenta ( M) and yellow (Y) reflected light of opaque color material using reflected light are used.

When inputting the original image, the cyan (C),
The reflected light of magenta (M) and yellow (Y) is measured by an input device. At this time, the input device causes distortion.
Also, the output device has the same color distortion, and the specified color is not reproduced.

Therefore, in order to correct the distortion, a correction function corresponding to the characteristic of the generated distortion is generated and corrected by a method called a masking method. Fig. 6 (a1), (a2),
(a3) shows a color correction method by the masking method, and FIG. 6 (a1) shows a predetermined function term and a masking coefficient in a matrix operation to obtain expected red (R), green (G ), Blue (B)
Can be obtained. Where N: the number of columns in the matrix, that is,
Here we show the number of terms, 3: the number of rows in the matrix, here red (R),
Indicates three rows of green (G) and blue (B), and indicates the number of rows of the matrix (here, one row may be used because a common function term is used to output each value). .

As described above, the correction function based on the masking method determines the form of the function such as the order and the number of terms in advance, and determines only the masking coefficient by the least squares method. Normally, a function using a nine-term function including a quadratic term called a quadratic masking method shown in FIG. 6 (a2) and a function using a constant are used.

In particular, when the accuracy of correction is required,
A function of 23 terms including a third-order term called a third-order masking method shown in FIG. 6 (a3) is used. If the number of terms used is quadratic or tertiary, the correction error decreases, but the amount of calculation increases accordingly.

The principle of the masking method is described in detail in, for example, "Image Processing for Color Reproduction", written by Hiroko Kodera, published by Kogaku Kogyo Shuppan, 1988, pp. 44-55. The correction function based on the above-mentioned masking method has a function form such as an order determined in advance, so that an optimum correction function cannot always be generated in all cases.

Therefore, as shown in FIG. 6 (b), it is known to generate this correction function using a neural network. For example, "" Color Correction of Color Hard Copy Using Neural Network ", Proc. Of the IEICE Spring Conference (1989), 199, shows that color correction at the time of printing is performed using a neural network. I have.

Since the neural network can acquire a non-linear transformation rule by learning, it is necessary to define a shape in advance as in the above-mentioned masking method,
A conversion rule more suitable for the distortion can be obtained by a method of generating a certain correction function.

In this case, red (R),
Green (G), blue (B), or cyan (C), magenta (M), yellow
The color correction system is trained by using the color data of the primary color such as (Y) and the corrected color data as the teacher data on the output side.

Further, the above color correction, for example, red (R), green
(G), Blue (B) From expression, cyan (C), magenta (M), yellow
(Y) Color coordinate conversion such as conversion to expression was performed separately.

[0016]

In the above-mentioned conventional masking method, as described above, when the order used in color correction increases, that is, when the number N of items increases, the error approaches the expected value and the error decreases. However, since the amount of calculation is significantly increased, the above-described secondary masking method is usually used in the normal case. Therefore, there is a need for a color correction method that can improve the accuracy without significantly increasing the calculation time.

In the conventional color correction method using a neural network, only primary colors are used as input and output. However, in order to improve accuracy without increasing the weight calculation amount in the neural network, It was necessary to determine what terms could be entered and use more appropriate input terms.

In view of the above-mentioned conventional disadvantages, the present invention provides a method for performing color coordinate conversion and color correction at the time of inputting / outputting an image from a color scanner, a color printer, a CRT display, a color plotter, etc., using a neural network. It is an object of the present invention to provide a method capable of performing color correction and color coordinate conversion capable of improving accuracy by reducing the amount of calculation in the neural network, specifically, the amount of weight calculation. is there.

[0019]

FIGS. 1 to 3 show an embodiment of the present invention. FIG. 1 (a) shows a case where primary or higher order color data is input. FIG. 1B shows a case in which the scale of the neural network is reduced while suppressing a decrease in the accuracy of color correction and color conversion. FIG. 2 shows a case in which the neural network is divided for each output.
FIG. 2 shows a case in which, in the method of FIG. 2, a color element having a large error is set to one output, and the other components are learned by a neural network having a plurality of outputs. The above problems can be solved by a color correction and color coordinate conversion method using a neural network configured as follows.

(1) When inputting / outputting an image, a color correction of the input / output device is performed by using a neural network which has learned the relationship between the color distortion of the image and the accurate color by the input / output device.
It is a method of performing color coordinate conversion, and from primary color data,
Higher-order color data is created, learned as input data of the neural network, and color correction,
It is configured to perform color coordinate conversion.

(2) A method for performing color correction and color coordinate conversion by the neural network according to the above item 1, wherein after learning, only the combination of input items having a large effect on the output among the input patterns at the time of learning. Is selected, a neural network is formed, learning is performed again, and color correction and color coordinate conversion are performed using the learning result.

(3) A method for performing color correction and color coordinate conversion by the neural network according to item 1 above, wherein after learning, input items having a large effect on output are separately selected from input patterns at the time of learning. Then, a neural network is selected, learning is performed again, and color correction and color coordinate conversion are performed using the learning result.

(4) A method for performing color correction and color coordinate conversion by the neural network according to item 2 or 3, wherein after selecting an input item, a part of the learned weight is extracted and re-learned. It is constituted so that.

(5) A method for performing color correction and color coordinate conversion by the neural network according to the above items 1 to 4, wherein the teacher data on the output side is set to a value after the color coordinate conversion.

[0025]

[0026]

[0027]

[0028]

[0029]

According to the present invention, in the input device,
Various types of color data, such as terms used in secondary masking and tertiary masking, generated from primary colors are used as input data. For example, a color chart is a value measured by a spectral reflectometer, red ( The measured values of R), green (G), and blue (B) are trained on the neural network as teacher data on the output side. << See FIG. 1 (a) >> By using various types of data, the amount of calculation increases, but the accuracy of the output result increases. At this time, the weight W after the learning is set.

Here, in the present invention, after the learning,
It is checked which of the input terms in the color data used for the learning has a strong effect. And while maintaining the output accuracy,
In addition, in order to reduce the amount of calculation, only input terms that have a large effect on output are selected, and those terms are used as input data, and learning is performed again. Since the number of items of input data is reduced, the amount of calculation is reduced, and learning is performed using only input terms that have a large effect on output. Therefore, the influence on output data is reduced, that is, a decrease in accuracy can be reduced. .

That is, the neural network learning method according to the present invention has a large effect on the output result, even if the calculation is performed by deleting the input term having little effect on the output data, ie,
It focuses on the fact that there is no change in accuracy and the amount of calculation can be reduced.

At this time, as a method of determining the magnitude of the influence of each input term, even if the value of the input term in the neural network after learning is not input, the output value is larger than the original output result. When there is no change, a method is adopted in which the influence of the input term is small.

Specifically, there are the following two types of methods for selecting an input term that has a strong influence on the output side. Method 1: One is a selection method based on an error for each combination of input terms. Each has little effect when used separately, but when used together, it is effective when the effect on output data increases.

For example, it is assumed that learning is first performed with N input terms. Then, there are 2 N combinations of input data. From all the combinations, a combination of input terms having a large influence on the output side is selected in consideration of the calculation time depending on the number of inputs.

Specifically, when learning is performed using the N input terms, the above 2 N combinations are input to the learned neural network. Each pattern has a term for inputting a value and a term for not inputting a value (inputting “0”). With reference to the error between the output data and the teacher data for the 2 N power combinations, a combination with a small error, that is, a high-precision combination is extracted, and then the value of the input term becomes “0”. By selecting a combination that has many terms, that is, a calculation time that is short, for example, a combination that has several items of input and obtains almost accurate output data, the effect on the output data is large, and It is possible to extract a combination of input terms with a small amount of calculation.

Method 2: The second method is a selection method based on the influence of each input item. This is a method in which, using the above-mentioned 2 N combinations, a term having a large influence among the N terms is selected, and some of them are selected.

Specifically, a plurality of input patterns having a small error from the teacher data of the output data, that is, a plurality of high-precision input patterns are extracted from among the output data corresponding to the above 2 N combinations. In the term for which the value is entered, ie,
By selecting only those items whose input terms are not “0” and that are frequently used, it is possible to extract input terms that have a large effect on output data.

Next, when learning is performed again using only the input term selected as described above as an input term, a random number may be used as an initial value of the weight W. , A weight W corresponding to the selected input term may be extracted, and the weight may be re-learned as an initial value.

At this time, since learning is performed using only the selected term, an input unit that is not used in the weight W is created. If the connection to the unit without input is not used and the weight W of the other connection is used as it is, the learning time can be reduced as compared with the case of learning from the beginning by using the weight by random numbers. Can be.

In the above learning, as a teacher data on the output side, the result of the color correction once is subjected to color coordinate conversion. For example, red (R), green (G), blue (B) → cyan (C), magenta By using the color data after (M) and yellow (Y)｝, the color correction and the color coordinate conversion can be performed at the same time. By performing such learning, the calculation time can be further reduced. can do. For example, there is a case where the input is red (R), green (G), and blue (B), and the output data is X, Y, and Z, which are values in an XYZ color system that is easy to handle with numerical values.

[0041]

[0042]

[0043]

[0044]

[0045]

[0046]

[0047]

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 to 3 are diagrams showing an embodiment of the present invention, FIGS. 4 and 5 are diagrams showing an experimental example of the present invention, and FIG. 4 is a diagram showing a higher-order input term. FIG. 5 shows an experimental example in which learning is performed with the number of items set to three or four, and FIG.
An experimental example in the case of learning by dividing into outputs is shown.

The present invention uses a neural network which learns the relationship between the color distortion of an image and an accurate color by an input / output device at the time of inputting / outputting an image. In the method of performing the conversion, high-order color data is created from the primary color data, learning is performed as input data of the neural network, and color correction and color coordinate conversion are performed using the learning result. In a method of performing color correction and color coordinate conversion by inputting higher-order color data to a neural network, after learning, only a combination of input terms having a large effect on output is selected from input patterns at the time of learning. Means for constructing a neural network, re-learning, performing color correction and color coordinate conversion using the learning result, and separately selecting input terms having a large effect on output from input patterns at the time of learning. And-option, configure the neural network, means for performing learning again, also, a single neural network of the multi-input multi-output, each first output, or,
The means for learning by dividing into one output and another plural output neural network is a means necessary for carrying out the present invention. Note that the same reference numerals indicate the same object throughout the drawings.

Hereinafter, referring to FIGS. 1 to 3, FIG.
An experimental example of the color correction and color coordinate conversion method using the neural network of the present invention will be described with reference to FIG. First, in this experimental example, a case where a color original image is input by a scanner and color distortion correction of the result is performed will be described. The neural network used for distortion correction used this time is a three-layer hierarchical network as shown in FIGS. 1 to 3, and the number of units in the intermediate layer is six. The output side teacher data used for the learning of the neural network uses the value measured by the spectral reflectometer of the color chart, and the input side data is the same term as the secondary masking calculated from the above-mentioned scanner measurement value. (R, G, B, RG,
RB, GB, R ² , G ² , B ² ) are used, and based on the learning result, an input term that greatly affects the output is selected. As a learning rule, a known back propagation method was used.

1) Scanner input is performed using the 892 color chart as an original image. 2) Measure the same color original image with a spectral reflectometer. 3) As input-side data, the same terms (R, G, B, RG, RB,
GB, R ² , G ² , B ² ) are calculated.

4) The teacher data on the output side is used as the measured value of the above-mentioned spectral reflectometer. 5) Of the above 892 color input / output data, 456 colors
The neural network is trained, and the remaining 456 colors are used for evaluation of the neural network after learning. The weight after the learning is defined as W.

6) The nine items (R, G, B, R
For all combinations (2 9) of G, RB, GB, R ² , G ² , and B ² ), the error of the output data with the teacher data is checked, and the combination of the error with a small error, that is, the combination with good learning accuracy is determined. Select an input term. This is because if the value of the output data does not change much without inputting the value of an input term, it can be said that the influence of the input term is small.

7) As described above, out of the nine input terms, a combination of the terms having the smallest error in the case of three and four pieces described later was selected as the input terms having a large influence. In this case, the amount of calculation is smaller than that of the tertiary masking method, as is clear from the experimental results described later.

8) Only the combination of the selected terms is input and learned. In this case, the combination of the selected terms is used as an input, and the weight W obtained by learning using the nine input terms is not used.
Learning was performed 0 times.

9) After the learning is completed, the remaining 456
For the color evaluation data, the error between the teacher data and the output data of the neural network is compared. The experimental results are shown in FIG. As a result of the experiment, when four input items were selected, R, G, B, and GB were obtained, and when three input items were selected, R, G, and GB were obtained.

FIG. 4 shows the learning data in each case,
For the evaluation data, the error between the output data of the neural network and the teacher data and the average value are shown. The value at this time is, for example, La for convenience of the experiment.
b These are values of the color system, and each calculation amount is represented by a ratio.

In FIG. 4, ΔEL indicates a brightness error, ΔEab indicates a hue error, and ΔE indicates an average error (specifically, an added value) of the brightness and hue errors. I have.

As is apparent from FIG. 4, the use of the above-described color correction method according to the present invention, compared with the tertiary masking method,
The computational complexity is small and the accuracy is better than the tertiary masking method, indicating that this method is effective.

Next, another experimental example, that is, a case of learning using a neural network in which an output term is divided into one output and two outputs will be described with reference to FIG. Also in this case, a color original image is input by a scanner, and the color of the input image is corrected.

The neural network in this case is 3
It was a hierarchical network of layers, and the number of units in the intermediate layer was four. The teacher data on the output side used for learning of the neural network uses the values of R, G, and B, which are values obtained by measuring the color chart with a spectral reflectometer, and the input side data uses the measured values of the scanner. As a learning rule, a known back propagation method was used.

1) Scanner input is performed using the 892 color chart as an original image. 2) Measure the same color original image with a spectral reflectometer. 3) As the input data, R, which is the measured value of the scanner,
G and B values.

4) The teacher data on the output side is used as the measured value of the spectral reflectometer. 5) Of the above 892 color input / output data, 456 colors
The neural network is trained, and the remaining 456 colors are used for evaluation of the neural network after learning.

6) As a result of performing the color correction in the above learning, among the output values (R, G, B), particularly, for the one having a large error, a one-output neural network dedicated to the output correction is used. Do the learning.

In the specific example, since the error for red (R) was the largest, a neural network for the R correction was created and learning was performed. In this case, the learning data includes, as input-side data, the measurement values (R,
G, B), and red (R) value among the measured values of the spectral reflectometer was used as the teacher data on the output side.

7) Remaining output (in this case, green (G), blue (B)
｝ Is collectively learned by a three-input two-output neural network. At this time, the same input data as in the above case 6) is used, and green (G) and blue (B) among the measured values of the spectral reflectometer are used as the output teacher data.

8) The evaluation is performed using the evaluation data. FIG. 5 shows the experimental results in this experimental example. In the conventional color correction method of three inputs and three outputs, the error for red (R) is green.
(G) and blue (B) were larger than the error, so we used two neural networks, one for red (R) correction and one for green (G) and blue (B) correction. They were trained 200,000 times.

FIG. 5 shows the error between the output data of the neural network and the teacher data and the average value of the learning data after learning and the evaluation data. Also in this case, for convenience of experiment, it is indicated by an error value in a known Lab color system.

As can be seen from the figure, the use of the color correction method according to the present invention makes it possible to increase the accuracy of color correction without significantly increasing the amount of calculation as compared with the conventional method (multi-input multi-output). You can see that it can be done.

Although the above experimental example is an experimental example of color correction, as described in the operation section, a value obtained by performing color coordinate conversion on a color after color correction is used as teacher data on the output side. It goes without saying that it may be used. In this case, color correction and color coordinate conversion can be performed by one neural network.

In the above experimental example, a neural network is constructed by selecting an input term, and an output term is divided into one output or one output and a plurality of outputs. In the above, an example in which a neural network for color correction with a plurality of outputs is constructed and learned has been described, but it goes without saying that these two types of methods may be combined. That is, an input term having a large influence on output data is selected, and the output term is divided for each output, or for one output and a plurality of outputs, to construct a neural network for the color correction, Alternatively, it goes without saying that a method of performing color coordinate conversion is also effective.

[0071]

As described above in detail, according to the color correction and color coordinate conversion method using the neural network of the present invention, the accuracy of color correction can be increased without significantly increasing the calculation amount, and at the same time. Color coordinate conversion can also be performed.

[Brief description of the drawings]

FIG. 1 shows an embodiment of the present invention (part 1).

FIG. 2 shows an embodiment of the present invention (part 2).

FIG. 3 shows an embodiment of the present invention (part 3).

FIG. 4 shows an experimental example of the present invention (part 1).

FIG. 5 shows an experimental example of the present invention (part 2).

FIG. 6 is a diagram illustrating a conventional color correction method.

[Explanation of symbols]

 ○ Each unit of neural network R, G, B Color of light used for additive color expression C, M, Y Color of reflected light used for subtractive color expression

(56) References JP-A-4-83471 (JP, A) JP-A-4-51670 (JP, A) JP-A-4-47475 (JP, A) (58) Fields surveyed (Int .Cl. ⁷ , DB name) G06T 1/00 G06T 5/00-5/50 G09G 5/02 H04N 1/46 H04N 1/60

Claims (4)

(57) [Claims] An input / output device for inputting / outputting an image;
A neural that learned the relationship between image color distortion and accurate color
Color correction, color
A method of performing coordinate transformation, which, after the learning, affects an output among input patterns at the time of the learning.
Select only combinations of input terms with large
Construct a neural network, learn again, and
Color correction and color coordinate conversion using
Color correction and color coordinate conversion methods using neural networks. 2. An image input / output device for inputting / outputting an image.
A neural that learned the relationship between image color distortion and accurate color
Color correction, color
A method of performing coordinate transformation, which, after the learning, affects an output among input patterns at the time of the learning.
The input terms with large
And learn again, and use the learning result to perform color compensation.
Neural network characterized by performing positive and color coordinate conversion
Color correction and color coordinate conversion method by network. 3. A neural network according to claim 1, wherein
Color correction and color coordinate conversion
After selecting the input terms, a part of the learned weights is extracted and
A neural network characterized by learning
Color correction and color coordinate conversion method. 4. The neural network according to claim 1, wherein :
This is a method of performing color correction and color coordinate conversion by a work, and it is characterized in that teacher data on the output side is a value after color coordinate conversion.
Color correction and color coordinates by neural network
Conversion method.