JPH1169182A - Color correction method and color correction device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To perform color correction and color coordinate transformation with high accuracy by performing color coordinate transformation by a neural network that uses an error evaluation function which minimizes the mean value of errors in an output signal to perform color correction. SOLUTION: A neural network which uses an error evaluation function that minimizes a mean value of errors in an output signal performs color coordinate transformation and the color correction of an input image is performed. For instance, in a color image device, an input device 1 consists of a scanner or the like, and inputs the color data of an original image to a color correction device 2. The device 2 performs color correction processing of the color data that is inputted by a color correction means. An output device 3 comprises a printer or the like, records the original image which is undergone color correction processing on a recording medium paper or the like and outputs it. Here, the device 2 uses a neural network as a color correction means to perform color coordinate transformation.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a color correction processing by performing color correction and color coordinate conversion in a recording device such as a color copying machine or a color printer having a color correction function or a display device such as a color display. The present invention relates to a color correction method to be performed and a color correction device using the color correction method.

[0002]

2. Description of the Related Art Hitherto, as an image output apparatus for reading a color original image by a reading means such as a color scanner and outputting a reproduced image of the color original image by an output means such as a color printer, a color copying machine or a color printer has been known. There is an image processing device. In such a color image processing apparatus, even if the input image data is directly input to the output unit, in most cases, it is output as a duplicate image having a color different from the original image.

Various color correction techniques have been proposed in order to obtain a reproduced image whose color is faithfully reproduced from the original image. Representative color correction techniques include the 1990 Image Electronics Society Vol. 29, No. 3, "Color Correction Techniques for Faithful Color Reproduction".
Color correction processing by digitization using polynomial regression analysis, color correction processing by non-formulaization using three-dimensional interpolation using a look-up table, or color correction processing by non-formulation using a neural network I have.

[0004] Among the above-described color correction processes, in particular, a color correction process using a neural network can perform a desired color correction by simply providing color data even when a large number of colors are used in an original image. It is possible. In this method, the color data to be sampled is given to the neural network, and the neural network itself learns spontaneously so that the target color correction data can be obtained, and the neural network itself finds a processing procedure for color correction. This is due to the fact that

[0005] In addition, the neural network is
Since a large amount of data can be processed in parallel, it is suitable for color correction processing when the number of colors of the original image is large. As described above, an apparatus for performing a color correction process using a neural network is disclosed in JP-A-2-241.
No. 271 discloses a color correction device.

This color correction apparatus seeks an optimal solution for color correction by the lowest descent method based on the algorithm of back propagation (error propagation learning rule). This back propagation is excellent in practicality among neural network learning methods, and is generally implemented in a neural network.

The above-mentioned color correction apparatus uses a neural network which implements backpropagation, so that optimal color correction can be performed, and phenomena such as distortion of gradation reproduction caused by a small number of color samples. Can be reduced.

The neural network learning by the back propagation will be described in detail. First, as a first step, the number of intermediate layers between the input layer and the output layer and the number of units in each layer are set. . For example, the neural network in the following description is set to have n layers. That is,
The number of intermediate layers is n-2 layers excluding the input layer and the output layer. The number of units in the output layer is N.

Next, as a second step, learning data and initial values of all connection weights w between units in the neural network are given as input data to an input layer (first layer). Then, by using these input data, the output of the second layer (first intermediate layer), the third layer (second intermediate layer)
The output of ..., calculates the output of the n-1 layer (n-2 th intermediate layer), it will determine the output of the intermediate layers, the output o ⁿ from the output layer finally a second n layer .

At this time, a total sum I ^k _j of inputs to the j-th unit of the k-th layer is calculated by the following equation (1). Here, w ^{k−1, k} _{i, j} below is a coupling load between the i-th unit of the k−1th layer and the j th unit of the k-th layer, and the following o ^k−1 _i is the ^k- th unit. -1 is the output signal of the ith unit in the first layer.

I ^k _j = Σ _{i = 1} w ^{k−1, k} _{i, j} o ^k−1 _i (1) At the same time, the output signal o ^k _j in the j-th unit of the k-th layer is expressed by the following equation: It is calculated by (2). Note that f _j below is a function that determines input / output characteristics to the unit.
Differentiable by an independent variable sum I ^k _j of the input, and a non-decreasing function with respect to the sum I ^k _j input. As this f _j , for example, a sigmoid function or the like is used.

[0012] ^{_{o k j = f j (I}} k j) ... (2) Thus, as the forward-type process information processing via the intermediate layer from the input layer proceeds to the output layer in the back propagation.

[0013] Subsequently, as a third stage, the output signals obtained o ^k _j and, the k-th layer, which is provided as input data j
Compare with the teacher signal t ^k _j in the unit.
The teacher signal t ^k _j is a signal that is a correct answer of the input learning data. In other words, the teacher signal t ^k _j is an ideal value for the output signal o ^k _j which is a calculated value obtained from the input learning data.

[0014] In the case of this teacher signal t ^k _j and the output signal o ^k _j are the same, back to the second stage. At this time, the data of all connection weights in the neural network is not input data but data obtained by calculation and changed. On the other hand, if the teacher signal t ^k _j and the output signal o ^k _j do not coincide, to subsequent to the fourth step.

In the fourth stage, the output error δ ^k _j of the j-th unit of the k-th layer, which is the input data for the backward calculation processing of the output layer, is calculated. First, the output error [delta] ⁿ _j in the j-th unit of the n-layer (k = n) in the form of the output layer, the error evaluation function epsilon (t ⁿ _j, o ⁿ _j) and, from the above equation (2) It is calculated by the following equation (3).

[0016]

(Equation 2)

Here, the connection weight between the units of each layer in the neural network, for example, the connection weight w ^{k−1, k} _{i, of the i} ^- th unit of the ( ^k−1) -th layer and the j-th unit of the ^k- th layer _j of the change amount [Delta] w ^{k-1, k} _{i, j} is the output error [delta] ^k _j in the j-th unit of the k-th layer, the k-
It is calculated from the output signal ok ^-1 _i of the i-th unit in the first layer by the following equation (4).

Δw ^{k−1, k} _{i, j} = ηδ ^{k k} o ^k−1 _i (4) η: proportionality constant and _{j in the} ^k- th layer other than the output layer (n-th layer) The output error δ ^k _j in the unit is calculated by the following equation (5) from the above equations (2) and (4).
The following N _{k + 1} indicates the number of units in the k + 1-th layer, and the following δ ^{k + 1} _s indicates the output error in the s-th unit in the k + 1-th layer. Similarly, the following w ^{k, k + 1}
_{j, s} indicates the coupling load between the j-th unit in the k-th layer and the s-th unit in the (k + 1) -th layer.

Δ ^k _j = f _j (I ^k _j ) (1-f _j (I ^k _j )) Σ ^{N k + 1} _{s = 1} (δ ^{k + 1} _s w ^{k, k + 1} _{j, s} ) = o _{^{k j (1-o k j}} ) Σ Nk + 1 s = 1 (δ k + 1 s w k, k + 1 j, s) ... (5) in other words, in the n-th layer is the output layer, the formula ( From the output error δ ⁿ _j in the n-th layer calculated in 3) and the above equation (2), the intermediate layer other than the output layer, the (n−1) -th layer
The output error δ ^n-1 _j in the layer is calculated by equation (5). Further, the output error δ ⁿ⁻¹ _{j in the (} ^{n−1) th} layer
From the equation (4), the unit of the (n-1) th layer and the (n-2) th layer are
The amount of change Δw ^{n−2, n−1} _{i, j} of the coupling load between the layer and the unit is calculated. Then, the change amount Δw of the coupling load
^{n-2, n-1} _{i, j} and the output error [delta] ⁿ _j in the n layer, the equation (4) because the output in the n-2 layer error [delta] ^n-2
_j is calculated.

The change amount Δw ^{k-1, k of the} coupling load as described above
_i, an operation of calculating a _j and output error [delta] ^k _j, the amount of change [Delta] w ^{1, 2} _i of coupling weight in units of the input layer (first layer) and the second layer (first intermediate _{layer), j} is Repeat until calculated.

Such information processing that proceeds from the output layer to the input layer via the intermediate layer is called backward processing in back propagation.

Finally, as a fifth stage, between the units of each layer calculated by this backward type processing (specifically, the i-th unit of the (k-1) -th layer and the j-th unit of the k-th layer) the amount of change in coupling load between) the Δw _^k-1, k _^i, each connection weight with ^{_{j w k-1, k i}} , performs a correction of _j, returns to the second stage. In the device that performs the color correction process using the neural network, learning by back propagation is performed in this manner.

[0023]

In THE INVENTION to be solved INVENTION method learning by back propagation as described above, in the case presented input pattern p of the input data, the error evaluation function ε from the output layer ^{_{(t n j, o n j}} ) As the teacher signal t ⁿ _j
Equation (6) is used which is a function using a difference between the output signal o ⁿ _j and.

[0024] _{^{ε (t n j, o n}} j) = E p = Σ N j = 1 (t n pj -o n pj) 2 ... (6) t n pj: output layer when presenting the input pattern p J ⁿ unit teacher signal on _pj : output signal of the j-th unit in the output layer when the input pattern p is presented Here, in the color correction device using the neural network, color correction and color coordinate conversion are performed. In this case, it is necessary to cause the neural network to perform learning so as to reduce the difference between the teacher signal and the output signal, that is, the average value and the standard deviation in the error of the output signal. This is because, if there is a variation in the color error to be corrected, in the obtained duplicate image, each reproduced color will have a difference that cannot be seen in the original image, and the color reproducibility of the entire duplicate image will decrease. To do that.

[0025] However, in the above formula (6), the input pattern when p was presented, j-th and the teacher signal t ⁿ _j at unit output signal o ⁿ _j and the difference (error in the output layer (the n-th layer) the sum of squares error evaluation function ε _{^{of) (t n j, o n}} j)
= E _p . Therefore, it can reduce the average value of the output error [delta] ⁿ _j of each unit in the output layer when presenting an input pattern p, has become unsuitable for reducing the standard deviation of the output error [delta] ⁿ _j .

That is, the error evaluation function ε of the above equation (6)
^{_{(t n j, o n j}} ) = for E _p is sum of squares, the error calculation by the equation (6) is intended to minimize the error of the overall output signal obtained from the output layer. Therefore, by using equation (6), in the case of performing the error evaluation, each of the output error of the output signal o ⁿ _j obtained from the respective units [delta] ⁿ _j
Mean value and standard deviation cannot be evaluated.

SUMMARY OF THE INVENTION The present invention has been made in view of the above problems, and an object of the present invention is to provide a color correction apparatus using a neural network for learning by backpropagation. An object of the present invention is to provide a color correction method capable of performing color correction and color coordinate conversion with higher accuracy than ever before by using an optimal function for conversion, and a color correction apparatus using the color correction method.

[0028]

According to a first aspect of the present invention, there is provided a color correcting method, comprising: a neural network using an error evaluation function that minimizes an average value of an error in an output signal; It is characterized in that the color of the input image is corrected by color coordinate conversion.

According to the first aspect of the present invention, in a neural network learning by back propagation in which an error is propagated to a previous layer, if the error evaluation function is used as the error evaluation function, the output of the neural network is output. The average value of the errors in the signal can be minimized. Therefore, in the color correction method using the neural network, it is possible to minimize the average color difference at the time of color coordinate conversion, and it is possible to perform the color correction processing better than before.

According to a second aspect of the present invention, in order to solve the above-mentioned problem, in addition to the first aspect of the present invention, in addition to the color correcting method of the first aspect, the number of units of N units in the output layer of the neural network is increased. a teacher signal for the j-th unit of which a t ⁿ _j, if the output signal from the j-th unit was o ⁿ _j, the error evaluation function is expressed by the following equation, error evaluation function epsilon (t ⁿ _j, o ⁿ _j ) = √ ｛Σ ^N _{j = 1} (t ⁿ
is characterized by a function expressed by _j -o ⁿ _j) ^2}.

According to the above method, the j-th output layer of the neural network is the n-th output layer.
Th output error [delta] ⁿ _j are opposite calculation processing data for the unit, the error evaluation function ε (t ⁿ _j,
is calculated by the o ⁿ _j). This error evaluation function ε
_{^{(t n j, o n j}} ) is the error evaluation function may be minimized difference between the N teacher signal obtained as a component t ⁿ _j and the output signal o ⁿ _j, i.e. the average color difference.

Therefore, the error evaluation function ε (t ⁿ _j , o
ⁿ _j ), the neural network is learned by back propagation using the output error δ ⁿ _j , thereby effectively minimizing the average color difference in the color coordinate conversion, and performing the color correction process more satisfactorily. It can be carried out.

According to a third aspect of the present invention, there is provided a color correcting method for correcting a color by a neural network using an error evaluation function for minimizing an average value and a standard deviation of errors in an output signal. Thus, the color of the input image is corrected.

According to the third aspect of the present invention, in a neural network learning by back propagation in which an error is propagated to a previous layer, if the above error evaluation function is used as the error evaluation function, the output of the neural network is output. The average value of the error in the signal can be minimized, and the variation of the error can be reduced, that is, the standard deviation of the error can be minimized. Therefore, in the color correction method using the neural network, the average color difference at the time of color correction can be minimized, and the standard deviation of the color difference can be minimized. Therefore, concentration of the color difference only on a certain specific representative color is suppressed, and the color correction process can be performed better than in the related art.

According to a fourth aspect of the present invention, in order to solve the above-mentioned problem, in addition to the third aspect of the present invention, in addition to the color correcting method of the third aspect, of the N units in the output layer of the neural network, If the teacher signal for the j-th unit and t ⁿ _j, and the output signal from the j-th unit and o ⁿ _j, the error evaluation function is expressed by the following equation, error evaluation function epsilon (t ⁿ _j, o ⁿ _j ) = ｛Σ ^N _{j = 1} (t ⁿ _j −
o ⁿ _j ) ² ｝ ² .

According to the method of the fourth aspect, the j-th output layer of the neural network is the n-th output layer.
Th output error [delta] ⁿ _j are opposite calculation processing data for the unit, the error evaluation function ε (t ⁿ _j,
is calculated by the o ⁿ _j). This error evaluation function ε
_{^{(t n j, o n j}} ) further sum of squares of the difference between the teacher signal t ⁿ _j obtained as N components the output signal o ⁿ _j 2
Since it is a power, the error evaluation function can minimize the standard deviation of the error in addition to the average value of the error.

Therefore, the error evaluation function ε (t ⁿ _j , o
By learning the neural network by back propagation using the output error δ ⁿ _j calculated by ⁿ _j ), the average color difference and the standard deviation of the color difference in the color correction are effectively minimized, and further better. Color correction processing can be performed.

According to a fifth aspect of the present invention, in order to solve the above-mentioned problem, in addition to the third aspect of the present invention, in addition to the color correcting method of the third aspect, among the N units in the output layer of the neural network, a teacher signal for the j-th unit and t ⁿ _j of, if the output signal from the j-th unit was o ⁿ _j, is the error evaluation function, the following equation,

[0039]

(Equation 3)

This is characterized in that it is a function represented by

According to the method of the fifth aspect, the j-th output layer of the neural network is the n-th output layer.
Th output error [delta] ⁿ _j are opposite calculation processing data for the unit, the error evaluation function ε (t ⁿ _j,
is calculated by the o ⁿ _j). This error evaluation function ε
_{^{(t n j, o n j}} ) further the difference between the sum of squares of the teacher signal t ⁿ _j obtained as N components the output signal o ⁿ _j gamma
Since it is a power and 1 <γ <2, it becomes an error evaluation function that can suppress the output pattern in the neural network from falling into a local solution in addition to minimizing the average value and the standard deviation of the errors. I have.

Therefore, the error evaluation function ε (t ⁿ _j , o
ⁿ _j ), the learning of the neural network by back propagation using the output error δ ⁿ _j minimizes the average color difference and the standard deviation of the color difference in color correction, and the obtained output pattern is localized. A solution can be suppressed, and a more effective color correction process can be performed.

According to a sixth aspect of the present invention, there is provided a color correcting apparatus including a color correcting means for executing the color correcting method according to any one of the first to fifth aspects. It is characterized by having.

According to the configuration of the sixth aspect, in order to perform color correction using a neural network that performs learning by back propagation, a method of color correction, that is, color coordinate conversion, color correction, or the like is used. An error evaluation function capable of performing an optimum error evaluation is used. Therefore, the color correction device is a color image processing device,
It is possible to perform better color correction processing than before.

[0045]

BEST MODE FOR CARRYING OUT THE INVENTION

[Embodiment 1] An embodiment of the present invention will be described below with reference to FIGS. 1 to 3 and FIG.
Note that the present invention is not limited by this. In the present embodiment, in the color image processing apparatus,
A color original image is input by an input device as an input signal (hereinafter, referred to as an RGB signal) as a representative color of red (R), green (G), and blue (B), and input data of the original image is CI.
A description will be given of a case where the color of an input image is corrected by converting into an E (International Certification Committee) L ^* a ^* b ^* color system.

As shown in FIG. 2, the color image processing apparatus according to this embodiment includes an input device 1, a color correction device 2, and an output device 3. The input device 1 includes a scanner or the like, and inputs color data of an original image to the color correction device 2. The color correction device 2 performs a color correction process on the color data input by the color correction unit. The output device 3 is composed of a printer or the like, and records the color-corrected original image on a recording medium such as paper and outputs it.

In the color correcting device 2, a neural network 2a is used as a color correcting means for performing color coordinate conversion. The neural network 2a is a network including three layers, an input layer 21, an intermediate layer 22, and an output layer 23, as shown in FIG.

Each of the input layer 21 and the output layer 23 includes three units (also called neurons) 20 constituting a network. This is because the input data of the RGB signal is subjected to the color coordinate conversion as the output data of the L ^* a ^* b ^* color system. Therefore, the neural network 2a has three inputs and three outputs. Also,
The intermediate layer 22 includes eight units 20 described above. The units 20 are coupled to each other with a variable coupling load as indicated by line 24.

The data of the teacher signal used for learning of the neural network 2a is data of the L ^* a ^* b ^* color system obtained by measuring a predetermined color chart with a spectrophotometer. Further, the learning data is obtained by inputting the predetermined color chart by the input device 1 and obtaining RGB data.
This is signal data. Further, as a learning rule, the lowest descent method based on a known back propagation is used.

The construction of the neural network 2a is performed by the following procedure. First, as procedure 1,
A predetermined color chart (assuming that M colors are used in this color chart) is measured by a spectrophotometer, and the obtained L ^* a ^* b ^* color system data is used as teacher signal data. Next, as a procedure 2, the same color chart is input to the color correction device 2 as RGB signal data by the input device 1 and the input data is used as learning data. Finally, as a procedure 3, the neural network 2a is trained by back propagation using the teacher signal data and the learning data.

The procedure 3 will be described in more detail. As shown in FIG. 1, the learning by the back propagation is basically performed in six steps (hereinafter, steps are abbreviated as S). First, as S1, an input pattern of input data composed of the data of the teacher signal and the learning data created in the above procedures 1 and 2 is presented. On the basis of this input pattern, the forward processing of the neural network is performed.

Next, as S2, the output error of each unit 20 of each layer in the neural network 2a is calculated. To calculate the output error, a function ε (t ^k _j , o ^k _j ) of a difference between the teacher signal t ^k _j to the j-th unit 20 of the k-th layer and the output signal o ^k _j from the unit 20 is obtained. Used as an error evaluation function. This error evaluation function ε (t ^k _j ,
o ^k _j), the intermediate layer of the neural network 2a 22
Or the j-th unit 20 ^k _j in the output layer 23
It has become a possible differential in the output signal o ^k _j from.

The function f _j that gives the input / output relationship in each unit 20 is a sigmoid function as shown in FIG. 6, and is represented by the following equation (11a) or (11b).

F _j (x) = 1 / (1 + exp (−x)) (11a) or f _j (x) = 1 / (1 + exp (−0.5 ^* x)) (11b) signal o ^k _j is the sum I ^k _j input to the j-th unit of the k-th layer, the sigmoid function f _j
From the following equation (12), it is calculated in the forward processing of the neural network 2a in S1.
Note that the sigmoid function f _j can be differentiated by the sum I ^k _j of the inputs, which are independent variables, and the sum I ^k of the inputs
It is a non-decreasing function for ^k _j .

[0055] ^{_{o k j = f j (I}} k j) ... (12) In the above-described neural network 2a, the output layer 23
(In the n-th layer, the neural network 2a, n =
J-th unit 20 ⁿ _j error evaluation function epsilon (t ⁿ _j for evaluating the error of 3), and o ⁿ _j), from the equation (12),
The output error δ ⁿ _j of the j-th unit 20 ⁿ _j in the output layer 23 is given by the following equation (13). Incidentally, I ⁿ _j represents the sum of the input in the j-th unit 20 ⁿ _j of the output layer 23.

[0056]

(Equation 4)

Further, based on the output error δ ⁿ _j , S3
Then, the connection weight between the units 20 of each layer in the neural network 2a is calculated. For example, the k-th
The change amount Δ of the connection weight w ^{k−1, k} _{i, j} between the _i ^- th unit 20 ^k−1 _i in the first layer and the j-th unit 20 ^k _j in the ^k- th layer
w ^{k−1, k} _{i, j} is the j-th unit 20 ^k _{j of the} ^k- th layer
From the output error [delta] ^k _j, and the output signal o ^k-1 _i in i-th unit 20 ^k-1 _i of the k-1 layer in, is calculated by the following equation (14).

Δw ^{k−1, k} _{i, j} = ηδ ^k _j o ^k−1 _i (14) η: proportionality constant and an expression for calculating an output error in a layer other than the output layer 23 (n-th layer) Is the j-th unit 20 of the k-th layer
Since it Expressed as a general expression for calculating the output error [delta] ^k _j in ^k _j the equation (12) and equation (14) is calculated by the following equation (15). The following N _{k + 1} indicates the number of units in the (k + 1) th layer, and the following δ ^{k + 1} _s indicates the _s in the (k + 1) th layer.
The output error in the 20th unit 20 ^{k + 1} _s is shown. Similarly, the following w ^{k, k + 1} _{j, s} indicates the coupling load between the j-th unit in the k-th layer and the s-th unit in the k + 1-th layer.

Δ ^k _j = f _j (I ^k _j ) (1−f _j (I ^k _j )) Σ ^{N k + 1} _{s = 1} (δ ^{k + 1} _s w ^{k, k + 1} _{j, s} ) = o ^k _j (1− ^ok _j ) Σ ^{N k + 1} _{s = 1} (δ ^{k + 1} _s w ^{k, k + 1} _{j, s} ) (15) For example, in the output layer 23 (third layer), (1
The output error δ ³ _j is calculated in 3). This output error δ ³
_j and the above equation (12), the output error δ ² _j in the intermediate layer 22 (second layer) other than the output layer 23 is _expressed by the following equation (1).
It is calculated by 5). Further, from the output error δ ² _j in the intermediate layer 22, the connection between the i-th unit 20 ¹ _i of the input layer 21 (first layer) and the j-th unit 20 ² _j of the intermediate layer 22 is calculated by Expression (14). Load change amount Δw ^1,2
_{i and j} are calculated.

[0060] Thus, in backpropagation, the error evaluation function _{^{ε (t n j, o n}} j) a so as to minimize the j-th unit of the k-th layer 20 ^k output error in _j [delta] ^k _j Is calculated.

Based on the change amount Δw ^{k−1, k} _{i, j} of the connection load obtained in this way, the connection load w
Modify ^{k-1, k} _{i, j} . This is the backward processing in the back propagation.

Next, in S5, it is determined whether all patterns have been presented as output patterns in the neural network 2a. At this time, if it is determined that all the patterns have not been presented, the process returns to S1 to present a new pattern and perform the forward processing. on the other hand,
If it is determined that all patterns are presented,
Proceed to the last S6.

In S6, it is determined whether or not the neural network 2a ends the learning. If the termination condition is not satisfied, the process returns to S1 to repeat the learning. On the other hand, if the termination condition is satisfied, the learning is terminated.

Incidentally, in the above S2, it is general that the evaluation of the color coordinate conversion is evaluated based on the magnitude of the average color difference. Therefore, minimizing the average color difference is the purpose of the neural network 2a that performs color coordinate conversion.

[0065] In this case, usually, in the learning method by the back propagation, the above error evaluation function ε (t ⁿ _j, o ⁿ
_j ), when the input pattern p is presented, the following equation (1)
The function shown in 6) is used.

Ε (tⁿ _j, Oⁿ _j) = E_p= Σ^N _{j = 1}(tⁿ _pj-Oⁿ _pj)^Two… (16) tⁿ _pj: J number of the output layer when the input pattern p is presented
Eye unit teacher signal oⁿ _pj: J number of the output layer when the input pattern p is presented
Output signal of the eye unit This equation (16) is the output when the input pattern p is presented.
Unit 20 of layer 23 ⁿ _jTeacher signal t atⁿ _pjAnd out
Force signal oⁿ _pj, Ie, the teacher signal tⁿ _jAnd output
Signal oⁿ _jIs a function based on the error of
tⁿ _jAnd output signal oⁿ _jAnd the sum of the squares of the differences
You.

However, when this equation (16) is used, the sum of the errors in the output data from the output layer 23 can be minimized, but the average value of the errors (in this embodiment, the average Learning that minimizes (color difference) cannot be performed by the neural network.

[0068] Therefore, the input pattern error evaluation functions that can minimize the average color difference between the measured and calculated values of the color chart in _{^{p ε (t n j, o}} n j) is necessary to set the resulting. For this purpose, first, the average color difference is expressed by the following equation (17).

[0069]

(Equation 5)

[0070] Therefore, the error evaluation function that the average color difference represented by the formula (17) and the minimum _{^{ε (t n j, o n}} j)
, I.e., the output signal from the output layer 23 o ⁿ the error evaluation function that minimizes the mean value of the error of _{j ε} (t ⁿ _j, o
ⁿ _j ) = E _p2 is newly set. That is, the above equation (1
Error evaluation function epsilon (t ⁿ _j represented by 6), the o ⁿ _j), is replaced in the following equation (18) is a function based on the above equation (17). Note that the number of units 20 in the output layer 23 is N = 3.

[0071]

(Equation 6)

Therefore, from this equation (18), the output layer 23
Considering L ^* , which is the _first unit 20 ³¹ , the following expression (19) is obtained.

[0073]

(Equation 7)

[0074] The second unit of the output layer 23 20 ³ ₂
In it a ^* and third unit 20 ³ is ₃ b
^* Is the same as the above equation (19). as a result,
The above equation (13) can be expressed by the following equation (20).

[0075]

(Equation 8)

As described above, the output error δ ⁿ _j , which is the data for the backward calculation processing for the j-th unit 20 ⁿ _j in the output layer 23 (n-th layer), is an error evaluation that minimizes the average color difference. function _{^{ε (t n j, o n}} j) because it is calculated by = E _p2, by performing learning by back propagation using the output error can be an average color difference to a minimum.

As described above, in the color correction method according to the present embodiment, the average color difference is obtained by using the above equation (18) in the neural network for learning by back propagation in which the error is propagated to the previous layer. Can be minimized. As a result, the color correction apparatus according to the present embodiment using the above-described color correction method can perform good color coordinate conversion in the color image processing apparatus.

[0078] It should be noted that the error evaluation function _{^{ε (t n j, o n}} j)
Is not limited to E _p2 represented by the above equation (18), but may be any function that can optimally evaluate an error in color coordinate conversion. Further, the color correction device 2, the error evaluation function _{^{ε (t n j, o n}} j) function as are several sets, depending on the type of the original to be converted color coordinates,
The configuration may be such that an optimal function is appropriately selected from the plurality of functions.

Embodiment 2 Another embodiment of the present invention will be described below with reference to FIG. 1 and FIGS. 4 to 6. For the sake of convenience, members having the same functions as those shown in the drawings of the first embodiment are denoted by the same reference numerals, and description thereof will be omitted.

In this embodiment, as shown in FIG. 4, in the color image processing apparatus shown in the first embodiment,
The difference is that a color correction device 4 is used instead of the color correction device 2. The color correction device 4 color-corrects the RGB signals input from the input device 1 and converts them into cyan (C), magenta (M), and yellow (Y) density signals (hereinafter, referred to as CMY signals). Thus, the color of the input image is corrected.

In the color image processing apparatus according to the present embodiment, similarly to the first embodiment, the input device 1 comprises a scanner or the like, and the color data of the original image is
To be captured. The output device 3 is composed of a printer or the like, and records the color-corrected original image on a recording medium such as paper and outputs it.

In the above-described color correction device 4, a neural network 4a is used as color correction means for performing color coordinate conversion. As shown in FIG. 5, the neural network 4a is a network including four layers: an input layer 41, intermediate layers 42 and 43, and an output layer 44.

Each of the input layer 41 and the output layer 44 includes three units 40 constituting a network. This is for color correction of input data of the RGB signal as output data of the CMY signal. Therefore, the neural network 4a has three inputs and three outputs. Also, the intermediate layers 42 and 43 are
Are provided. The units 40 are connected to each other by a variable connection load as indicated by a line 45.

The data of the teacher signal used for learning of the neural network 4a is obtained by combining all the divided CMY signals when the minimum value and the maximum value of the CMY signal are divided at equal intervals. is there.
The learning data is CMY selected as a teacher signal.
The signal data is output by a printer, and the output color sample is captured by an input device 1 such as a scanner.
This is the data of the GB signal. Further, as the learning rule of the neural network 4a, the lowest descent method based on a known back propagation method is used.

The construction of the neural network 4a is performed by the following procedure. First, as procedure 1,
729 colors, which are colors of each of the nine steps in the CMY signal, are used as teacher signal data. Next, as the procedure 2, the data of the selected 729 colors is output by the output device 3. Subsequently, as a procedure 3, the output color sample is input to the input device 1.
Thus, the input data of the RGB signals of 729 colors is set as learning data. Finally, as a procedure 4, the neural network 4 uses the data of the teacher signal and the learning data.
a is learned by back propagation.

The procedure 4 will be described in more detail. As in the first embodiment, as shown in FIG. 1, the learning by the back propagation basically consists of 6 steps.
Is performed in three S.

First, as S1, an input pattern of input data composed of the data of the teacher signal and the learning data created in the above procedures 1 to 3 is presented. On the basis of this input pattern, the forward processing of the neural network is performed.

Next, as S2, the output error of each unit 40 in the neural network 4a is calculated.
The calculation of the output error, the same first embodiment, the function _{^{ε (t n j, o n}} j) is used as the error evaluation function.
Also, the functions that give the input / output relation in each unit 40 are the sigmoid functions f _j (11a) and (11b) as shown in FIG. 6, as in the first embodiment.

When the neural network 4a has n = 4
Since the layer, the error evaluation function _{^{ε (t n j, o n}} j) from the above function f _j, the equation output error [delta] ⁴ _j of the j-th unit 40 ⁿ _j in the output layer 44 (13 ). Similarly, the output error δ of the j-th unit 40 _{kj in} the ^k- th layer other than the output layer 44
^k _{j is} also calculated by the equation (15). Thus, in backpropagation, the error evaluation function ε (t ⁿ _j, o ⁿ
The output error of the unit 40 is calculated so as to minimize _j ).

Subsequently, as S3, the output error δ ^k _j
By using the output signal o ^k _j from unit 40 ^k _j and,
Each unit 40 (unit 40 ^k-1 _i and unit 4
0 ^k _j ), the change Δw ^{k−1, k} _{i, j} of the connection weight, which is the correction amount of the connection weight w ^{k−1, k} _{i, j} , is calculated by the above equation (15).
Is calculated using

On the basis of the change amount of the coupling load obtained in this manner, as S4, a backward type process in back propagation for correcting each coupling load is performed.

Next, in S5, it is determined whether all patterns have been presented as output patterns in the neural network 4a. At this time, if it is determined that all the patterns have not been presented, the process returns to S1 to present a new pattern and perform the forward processing. on the other hand,
If it is determined that all patterns are presented,
Proceed to the last S6.

At S6, it is determined whether or not the neural network 4a ends the learning. If the termination condition is not satisfied, the process returns to S1 to repeat the learning. On the other hand, if the termination condition is satisfied, the learning is terminated.

Incidentally, when the neural network 4a learns by the normal back propagation in the above S2, the error evaluation function is the above equation (16). Therefore, the sum of the errors in the neural network 4a according to the present embodiment is represented by the following equation (21).

E = { _p } (C _tp −C _op ) ² + (M _tp −M _op ) ² + (Y _tp −Y _op ) ² ｝ (21) Here, as in the present embodiment, When color correction for converting RGB signals into CMY signals is performed, it is necessary that the average color difference is small and the color difference is also uniformly distributed. That is, in the output signal of the neural network 4a, the average value of the errors must be small and the errors must be uniformly distributed, that is, the standard deviation of the errors must be small.

On the other hand, even when the average value of the above-mentioned errors is small, a state where the errors are concentrated in a specific color,
That is, if good color correction cannot be performed for a specific color, it is not preferable color correction. For example,
In the present embodiment, the error (Δ
C, ΔM, ΔY), two sets of conversion results such as the following set1 and set2 are obtained.

Set1: (ΔC, ΔM, ΔY) = ｛(3,3,
3), (0,0,0), (0,0,0)｝ set2: (ΔC, ΔM, ΔY) = ｛(3,0,0), (0,3,0), (0,
0,3)｝ In this case, in set1, errors (color differences) are concentrated in ΔC, but in set2, errors are seen in ΔC, ΔM, and ΔY. Therefore, it can be considered that set2 is better corrected for color than set1. However, when error evaluation is performed on set1 and set2 using the above equation (21), it is evaluated that an error occurs in set1 and set2 at the same level as described below.

E _set1 = E _set2 = 27 As described above, the equation (21) is _converted to the error evaluation function ε (t ⁿ _j ,
o ⁿ _j ), it cannot be expected that the neural network 4a will perform learning suitable for color correction. Therefore,
New error evaluation function ^{_{ε (t n j, o n}} j) as to set the following equation (22). Note that the number of units 40 in the output layer 44 is N = 3.

[0099] ^{_{E p3 = {Σ N j =}} 1 (t n j -o n j)} 2 = {Σ 3 j = 1 (t n j -o n j)} 2 = {(C tp -C op) ² + (M _tp −M _op ) ² + (Y _tp −Y _op ) ² ｝ ² (22) This equation (22) is obtained by calculating N components from the N units provided in the output layer 44. Teacher signal t ⁿ _j obtained as
It has a further square of the sum of squares of the difference between the output signal o ⁿ _j. Therefore, when the above equation (22) is used, in the case where an error of approximately the same magnitude occurs, the more uniformly the error is distributed, the smaller the result E of the error evaluation becomes.

For example, using this equation (22),
When the error evaluation is performed for t1 and set2, the result is as follows.

E _set1 = 729> E _set2 = 243 As described above, when the above set 1 and set 2 are compared, the level of error generation is higher in set 1 and is more unsuitable for color correction than set 2 Is evaluated. Thus, equation (22), rather than the formula (21), the error evaluation function ^{_{ε (t n j, o n}} j) which is suitable for color correction can be regarded as.

Here, regarding the above equation (22), the output layer 4
Considering the _first unit 40 ⁴ ₁ (outputting the density signal of C), the following expression (23) is obtained.

[0103]

(Equation 9)

The second unit 40 ⁴ ₂ of the output layer 44
(Output the density signal of M) and the third unit 40
⁴ ₃ (output of Y density signal) is also calculated by the above equation (23).
Is the same as

As a result, the above equation (13) can be expressed by the following equation (24).

[0106] _{^{δ n j = 4 · √ (}} E p3) · (t n j -o n j) · o n j (1-o n j) ... (24) Thus, the output layer 44 (the n-th layer output error [delta] ⁿ _j are opposite calculation processing data for the j-th unit 40 ⁿ _j in), the error evaluation function that minimizes the mean and standard deviation of error in the color correction epsilon (t ⁿ _j ,
o ⁿ _j ) = E _p3 . Therefore, by performing learning by back propagation using this output error, it is possible to minimize the average value and the standard deviation of the errors in color correction.

As described above, in the color correction method according to the present embodiment, by using the above equation (22) as the error evaluation function in the neural network learning by the back propagation method for propagating the error to the previous layer, It is possible to minimize the average value of the errors and to reduce the variation of the errors, that is, to reduce the standard deviation of the errors. As a result, the color correction apparatus according to the present embodiment using the above-described color correction method can perform excellent color correction in the color image processing apparatus.

[0108] It should be noted that the error evaluation function _{^{ε (t n j, o n}} j)
Is not limited to E _p3 represented by the above equation (22), but may be any function that can optimally evaluate an error in color correction. Further, the color correction device 4, the error evaluation function _{^{ε (t n j, o n}} j) function as are several sets, depending on the type of the original to be color corrected, optimal from the plurality of functions It may be configured to appropriately select a suitable function.

[Embodiment 3] Another embodiment of the present invention will be described below with reference to FIGS. 1, 4 and 5. For the sake of convenience, members having the same functions as those shown in the drawings of the first embodiment are denoted by the same reference numerals, and description thereof will be omitted.

In this embodiment, the error evaluation function ε (t
ⁿ _j, as o ⁿ _j), except that the resulting output pattern is used functions like hardly fall into a local solution, are the same as those the second embodiment.

[0111] In the above color correction method in the second embodiment, in the learning of the neural network by back propagation as shown in FIG. 1, the error evaluation function _{^{ε (t n j, o n}} j) in S2 as an expression (22) In the method using equation (22), the error evaluation function ε (t
ⁿ _j, there is a possibility that the sharpness of the o ⁿ _j) is high.

The error evaluation function ε (t
ⁿ _j, using o ⁿ _j), neural network 4a
Is constructed, the output data obtained by the neural network 4a, that is, the output pattern for color correction may easily fall into a local solution. When the output pattern falls into a local solution as described above, the neural network 4a often cannot easily escape from the local solution.

[0113] Therefore, in this embodiment, the error evaluation function _{^{ε (t n j, o n}} j) instead of the equation as (22) to set the following equation (31).

[0114]

(Equation 10)

[0115] The equation (31), from the N units provided in the output layer 44, the square sum of the difference between the teacher signal t ⁿ _j obtained as N components the output signal o ⁿ _j Further, it is γ-th power, and 1 <γ <2.

Normally, neural network learning has a property that the higher the nonlinearity of an approximate function or error evaluation function, the more likely it is to fall into a local solution. Therefore, when the order of the error evaluation function is increased, the non-linearity of the error evaluation function increases, and the standard deviation of the error of each element cannot be reduced.

[0117] Therefore, in the equation (31), when the further power the square sum of the difference between the teacher signal t ⁿ _j obtained as N components the output signal o ⁿ _j, 1 <γ <2 Γ within the range is used as the order of the intermediate value. Thus, unlike the above formula which further squaring the square sum of the difference between the teacher signal t ⁿ _j and the output signal o ⁿ _j (22), never nonlinearity of the error evaluation function is high. Therefore, it is possible to prevent the output pattern obtained from the neural network 4a from easily falling into a local solution.

For example, when γ = 1.5, error evaluation is performed on set1 and set2 in the second embodiment using the above equation (31).

E _set1 = 140.3> E _set2 = 81.0 Thus, when the above set 1 and set 2 are compared, set 1 has a higher level of error generation than set 2 and is not suitable for color correction. It has been evaluated that there is some error, and the nonlinearity of the error evaluation function is not high. Therefore, in the equation (31), it is possible to suppress the output pattern obtained from the neural network 4a from easily falling into a local solution.

Therefore, the equation (31) is suitable for color correction, and the error evaluation function ε (t ⁿ _j ,
o ⁿ _j ).

Here, regarding the above equation (31), as in the second embodiment, considering each unit 40 ⁿ _j of the output layer 44 as the n-th layer, the above equation (13) is obtained by the following equation (13). 32).

[0122]

[Equation 11]

As described above, the output error δ ⁿ _j , which is the data for the backward calculation processing for the j-th unit 40 ⁿ _j in the output layer 44 (n-th layer), is the average value and the standard deviation of the errors in the color correction. while minimizing the resulting output pattern is calculated by the error evaluation function _{^{ε (t n j, o n}} j) = E p4 which suppresses falling into a local solution. Therefore, by performing learning by back propagation using this output error, it is possible to minimize the average value and standard deviation of errors in color correction, and to suppress the obtained output pattern from falling into a local solution. .

As described above, in the color correction method according to the present embodiment, by using the above equation (31) as the error evaluation function in the neural network learning by the back propagation method for propagating the error to the previous layer, It is possible to suppress the output pattern from falling into a local solution. As a result, the color correction apparatus according to the present embodiment using the above-described color correction method can perform excellent color correction in the color image processing apparatus.

[0125] It should be noted that the error evaluation function _{^{ε (t n j, o n}} j)
Is not limited to E _p4 represented by the above equation (31), but may be any function that can optimally perform error evaluation in color correction. Further, the color correction device 4, the error evaluation function _{^{ε (t n j, o n}} j) as in the second embodiment, a plurality of functions such as E _p3 and the E _p4 represented by formula (22) The function may be set so that an optimum function is appropriately selected from the plurality of functions according to the type of the original image to be color-corrected.

[0126]

According to the color correcting method of the present invention,
As described above, the color correction of the input image is performed by performing the color coordinate conversion by the neural network using the error evaluation function that minimizes the average value of the error in the output signal.

Therefore, in the above method, since the error evaluation function that minimizes the average value of the error in the output signal is used, the average color difference at the time of color coordinate conversion can be minimized. There is an effect that the color correction process can be performed well.

As described above, the color correcting method according to the second aspect of the present invention includes the following method in addition to the color correcting method according to the first aspect.
The teacher signal for the j-th unit of the N units in the output layer of the neural network is t
ⁿ _j and the output signal from the j-th unit is o ⁿ
If the _j, the error evaluation function is expressed by the following equation, error evaluation function _{^{ε (t n j, o n}} j) = √ {Σ N j = 1 (t n
a _j -o ⁿ _j) method is expressed as the ^2}.

[0129] Thus, in the above method, the error evaluation function _{^{ε (t n j, o n}} j) the average value of the error in the output signal can be minimized output error [delta] ⁿ calculated by
Learning of the neural network by back propagation is performed using _j . Therefore,
There is an effect that the average color difference in the color coordinate conversion can be effectively minimized, and the color correction process can be performed more favorably.

According to the color correcting method of the third aspect of the present invention, as described above, the color correction is performed by the neural network using the error evaluation function that minimizes the average value and the standard deviation of the error in the output signal. This is a method for correcting the color of an image.

Therefore, in the above method, the error evaluation function that minimizes the average value and the standard deviation of the error in the output signal of the neural network is used, so that the average color difference at the time of color correction is minimized, and Standard deviation can also be minimized. Therefore, it is possible to suppress the color difference from being concentrated only in a certain specific representative color, and it is possible to perform the color correction process better than before.

According to the color correction method of the fourth aspect of the present invention, as described above, in addition to the color correction method of the third aspect, the j-th unit among the N units in the output layer of the neural network is provided. ^{Let the} teacher signal for the unit be t ⁿ _j
And then, if the output signal from the j-th unit was o ⁿ _j, the error evaluation function is expressed by the following equation, error evaluation function _{^{ε (t n j, o n}} j) = {Σ N j = 1 (t ⁿ _j −
o ⁿ _j ) ² ｝ ² .

According to the method of the fourth aspect, the error evaluation function ε (t ⁿ _j , o) can minimize the standard deviation of the error in addition to the average value of the error in the output signal.
Learning of the neural network by back propagation is performed using the output error δ ⁿ _j calculated by ⁿ _j ). Therefore, there is an effect that the average color difference and the standard deviation of the color difference in the color correction can be effectively minimized, and the color correction process can be performed more favorably.

According to a fifth aspect of the present invention, as described above, in addition to the color correcting method of the third aspect, the j-th unit among the N units in the output layer of the neural network is provided. ^{Let the} teacher signal for the unit be t ⁿ _j
And then, if the output signal from the j-th unit was o ⁿ _j, is the error evaluation function, the following equation,

[0135]

(Equation 12)

This is a method that is a function represented by

[0137] Thus, in the above method, the output pattern in the neural network by using the output error [delta] ⁿ _j calculated by the error evaluation function which can prevent the falling into local minimum _{^{ε (t n j, o n}} j) Learning of the neural network by back propagation is performed. Therefore, it is possible to minimize the average color difference and the standard deviation of the color difference in the color correction, suppress the obtained output pattern from falling into a local solution, and perform more effective color correction processing.

A color correcting apparatus according to a sixth aspect of the present invention has a configuration provided with a color correcting means for executing the color correcting method according to any one of the first to fifth aspects as described above. is there.

Therefore, in the above configuration, in order to perform color correction using a neural network that performs learning by back propagation, a method of color correction, that is, an optimal error evaluation according to color coordinate conversion, color correction, and the like is performed. An error evaluation function that can be used is used. Therefore, the color correction apparatus has an effect that the color image processing apparatus can perform better color correction processing than before.

[Brief description of the drawings]

FIG. 1 is a flowchart illustrating a process of learning a neural network for color correction by back propagation in a color correction method according to an embodiment of the present invention.

FIG. 2 is a block diagram of a color correction device that executes the color correction method.

FIG. 3 is a schematic diagram illustrating a configuration of a neural network used in the color correction device of FIG. 2;

FIG. 4 is a block diagram of a color correction device that executes a color correction method according to another embodiment of the present invention.

FIG. 5 is a schematic diagram showing a configuration of a neural network used in the color correction device of FIG.

FIG. 6 is a graph showing a sigmoid function that gives an input / output state to a unit in each layer of the neural network.

[Explanation of symbols]

 2 color correction device 2a neural network (color correction means) 4 color correction device 4a neural network (color correction means) 20 units 23 output layers 40 units 44 output layers

Claims (6)

[Claims] 1. A color correction method for correcting a color of an input image by performing color coordinate conversion by a neural network using an error evaluation function that minimizes an average value of errors in an output signal. Wherein a teacher signal for the j-th unit of the N units in the output layer of the neural network, and t ⁿ _j, if the output signal from the j-th unit was o ⁿ _j, the error evaluation function is expressed by the following equation, error evaluation function _{^{ε (t n j, o n}} j) = √ {Σ n j = 1 (t n j
Color correction method according to claim 1, characterized in that a function represented by -o ⁿ _j) ^2}. 3. A color correction method for correcting a color of an input image by performing a color correction by a neural network using an error evaluation function that minimizes an average value and a standard deviation of errors in an output signal. Wherein a teacher signal for the j-th unit of the N units in the output layer of the neural network, and t ⁿ _j, the output signal from the j-th unit case of the o ⁿ _j, the error evaluation function is expressed by the following equation, error evaluation function _{^{ε (t n j, o n}} j) = {Σ n j = 1 (t n j -
o ⁿ _j) ^2} color correction method according to claim 3, characterized in that the function represented by ^2. 5. The teacher signal for the j-th unit of the N units in the output layer of the neural network, and t ⁿ _j, the output signal from the j-th unit case of the o ⁿ _j, the The error evaluation function is as follows: 4. The method according to claim 3, wherein the function is a function represented by: 6. A color correcting apparatus comprising a color correcting means for executing the color correcting method according to claim 1.