JPH03224076A - Graphic coloring system - Google Patents

Abstract

PURPOSE:To automatically set halftone coloring status by outputting or displaying a graphic colored from color information obtained by inputting a parameter to a neural network learned by a pair of parameter color information and with a neural network learning means. CONSTITUTION:The color information can be obtained by supplying the parameter of the pair 21 of parameter color information to the neural network 1. The neural network learning means 2 updates the weight value of the neural network 1 by using obtained color information and the color information of the pair of parameter color information. After such work is performed for several times by using plural pairs of parameter color information and weight is converged, it is recognized that learning is completed, and the color information in accordance with an arbitrary parameter can be obtained by using the neural network for which the learning is completed. In such a way, a color can be varied arbitrarily corresponding to the parameter.

Description

[Detailed Description of the Invention] [Summary] In a device for displaying figures, paintings, etc., the purpose is to arbitrarily change colors according to parameters, and the invention includes an input layer into which parameters are input, and the input layer. multiple intermediate layers that calculate and output weighted output values;
a neural network having a plurality of output layers that perform arithmetic processing on a plurality of values obtained by weighting the plurality of outputs of the intermediate layer and output the arithmetic results as color information;
means for storing a plurality of parameter color information sets in which parameters are associated with color information; an output value of an output layer obtained by applying the parameter of the parameter color information to an input layer of the neural network; and a value corresponding to the parameter; Neural network learning means for updating weight values of the neural network based on errors with color information, and coloring from color information obtained by inputting parameters to the neural network trained by the parameter color information set and the neural network learning means. The configuration includes a display output means for outputting or displaying the displayed figure.

[Industrial application field]

The present invention analyzes changes in colors in coloring work related to clothing design, etc., using parameters that affect those colors (sensory parameters, such as warm or cold, non-linear parameters, and year, month, and time such as seasons). This system relates to a system that arbitrarily displays various colors by viewing them as color changes with respect to parameters (such as minutes, etc.).

[Conventional technology]

2. Description of the Related Art Conventionally, systems have been implemented in the field of image processing such as animation that divide information on a screen into several areas and change the colors of the multiple areas based on given parameters. For example, consider the image in FIG. FIG. 9 is a diagram showing the scenery. Let the area of ■ be the sky, the area of ■ be the ground, and the area of ■ be the window of the house. The colors of these areas are changed according to the given time information. For example, the color of the landscape is different depending on the morning time (8:00 a.m. or 9:00 a.m.) and the evening time (5:00 p.m. or 6:00 p.m.). The sky is blue in the morning, but in the evening it is orange or red. In order to realize a system that changes the color of the area according to the time, conventional methods have been considered such as preparing color data corresponding to the time as a database or expressing it as a function.

[Problem to be solved by the invention]

However, if the above-mentioned method is prepared as a database, the amount of data in the database becomes large. For example, if the data is changed according to the time of day, the difference in the amount of data will not be large, but if it is made to change according to the season of -year, the amount of data will be enormous. If it is realized by a function, the function will be complicated.

Also, considering the case where the system is realized and modified,
If the method is prepared as a database, the amount of data to be corrected will be enormous. Furthermore, in the case of functions, it is necessary to ensure that there are no contradictions between the functions.

Therefore, it is difficult to automatically generate the coloring status of the target screen when the parameter values are different, and the user has to correct the coloring status each time depending on the change in the parameter status. In addition, the repair work was extremely difficult.

Therefore, an object of the present invention is to automatically set intermediate coloring conditions by allowing the user to appropriately set discrete coloring conditions without being concerned with the coloring condition in the initial state of the screen. The aim is to realize a coloring system that gives a unique coloring effect.

(Means for Solving the Problems) Fig. 1 is an explanatory diagram of the principle of the present invention.The present invention calculates an input layer into which parameters are input and a weighted value of the output value of the input layer. A neural network 1 having a plurality of intermediate layers that output, and a plurality of output layers that perform arithmetic processing on a plurality of values obtained by weighting the plurality of outputs of the intermediate layer, and output the calculation results as color information. , means 21 for storing a plurality of parameter color information sets in which parameters are associated with color information; an output value of an output layer obtained by applying the parameters of the parameter color information to the input layer of the neural network; a neural network learning means 2 for updating the weight value of the neural network based on the error with the corresponding color information; and a color obtained by inputting parameters to the neural network trained by the parameter color information set and the neural network learning means. The configuration includes a display output means 6 that outputs or displays figures colored from information.

[For production]

The parameters of the parameter color information set 21 are given to the neural network 1 to obtain color information. The neural network learning means 2 updates the weight values of the neural network 1 using the obtained color information and the color information of the parameter color information set.

The above operation is performed multiple times using multiple parameter color information sets. When the weights converge, the learning is completed, and color information corresponding to an arbitrary parameter is obtained using the neural network for which the learning has been completed.

If corrections are to be made, the corrections can be easily made by preparing a new set of parameter color information and performing additional learning. Unlike conventional methods, there is no need for the user to rewrite data files or correct functions.

〔Example〕

FIG. 2 is a configuration diagram of an embodiment of the present invention.

This device is a device that changes the colors shown in FIG. 9 according to given time information. In the figure, 1 is a neural network, 2 is a neural network operation unit, 3 is a neural network execution unit, 5 is a parameter input unit, 6 is a display device, 7 is a parameter input device, 8 is a color information extraction device, and 9 is an image input unit. It is a device.

Data input from an image input device 9 (scanner, etc.) is extracted by a coloring information extraction section 8 for each preset area. For example, in the case of Figure 9, ■,■
, ■ The color of the region is extracted. The operation of the coloring information extraction section 8 is shown in the flowchart of FIG.

Step 1 Receives input data from an image input device.

The data at this time is input as RGB dot information.

Step 2 A single neutral color representation is obtained by examining the ratio of RGB in the area of the data in which the RGB is mixed.

Step 3 Convert the color components after the conversion into numerical values using RGB. This numerical conversion is performed using 256 gradations (0 to 255). for example,
For purple, it is (255, 0, 255).

The data obtained by the coloring information extraction section 8 is sent to the neural network learning pattern generation section 4. The neural network learning pattern generation unit 4 associates the color information with time information.

Time information shall be input by the user. Specifically, this neural network learning pattern generation section 4 can be considered as an editor.

Here, a set of parameters and color information is created.

FIG. 8 shows a set of time and color information. For example, 1
On the first day of the month, the color of the sky, the color of the soil■, and the color of the window■ are respectively
RGB is (0,185,166).

(103,122,135), (21,157°152
).

In this way, a plurality of sets of parameters and color information are prepared. The groups are ranked 30th to 50th.

Next, this obtained information is stored in the memory area 21 (FIG. 3) of the neural network operating section 2.

In FIG. 3, 1 is the neural network controller shown in FIG. 2, 3 is the neural network execution section shown in FIG. 2, and 2 is the neural network operation section.

It is assumed that the neural network operation unit 2 performs learning of the neural network 1.

In the neural network 1 shown in FIG. 3, 11 is a network, and 12 is a memory area, which is an area for storing weight values and data used for intermediate calculations. It is assumed here that the 11 network is realized by software, and its flowchart is shown in FIG. Also, the fourth
The figure shows the configuration of the input layer, intermediate layer, and output layer.

In FIG. 4, h-1 and h-2 are input layers, and the year, month, and day are input to hl. Time is input to h-2. i-
1, i−n is the intermediate layer, and the output of the input layer has a weight Wih
is multiplied. The weights on each connection are different.

J-11, . . . , J-33 are output layers.

J-11, J-12, and J-13 output RGB of the sky color ■. J-21, 1-22, J-23 are soil color ■R
Output GB. J-3LJ-32°J-33 is assumed to output RGB of the window color ■.

The operation will be explained according to FIG.

Step 51 xl, ,=Σ)'ps+Wth is determined (w=h uses the one in the memory 12). (1
) formula (1+exp(-x pi+θi)) (21
Formula Step 52 ypi=1/ However, h: i: p: θ1: Wl: x91: Unit number of h layer (input layer) Unit number of i layer (middle layer) Pattern number of input signal Unit 1 of i layer Dark value h - Weight value of internal connection between 4 layers Sum of products of inputs from each unit of the layer to the 1st unit of the i layer y: Output from the h unit of the h layer for the input signal of the p pattern ypt: 1 of the i layer for the input signal of the 1)th pattern
Output step 53 from the unit No. 53: Ypi is stored in the Ypi storage area of the memory 12.

Step 54 The processes of steps 51 to 53 are carried out on the previous intermediate layer node I-
After obtaining each output Yp of 1, I-N, the process ends.

In steps 51 to 54, Ypffi is determined for each node -1°1-N.

Next, the processing in steps 51 to 54 described above is performed using the output of the intermediate layer.

The calculation method in steps 55 to 57 is as follows from steps 51 to 57 above.
It is similar to 54.

As a result, J-11,..., J-33 are R from the output layer.
You can get GB data.

Since the weight, Wik+wJ1, is initially stored with a random value, the outputs of J-11, ..., J-33 will be the same RGB even if the time data shown in Fig.
data cannot be obtained. Therefore, we use the back propagation method to learn the weights.

It is assumed that the back propagation method is performed by the neural network operating section 2 shown in FIG.

FIG. 6 is a flowchart when this is realized by software.

The operation will be explained below. Bank propagation method (D
, E. Rumelhart, G. E. Hinton,
and R,J.

Williams, “Learning Inter
nal Representations by Err
orPropagation, PARALLEL
DISTRIBUTED PROCESSING, Vo
l, 1.1) I), 318-364. The M
ITPress, 1986) will first be described in detail.

In the bank propagation method, the weight value Wih and the threshold value θ of the hierarchical network are adaptively and automatically adjusted and learned using an error feedbank. (1)
(As is clear from Equation 21, the weight values W, h and the threshold value θ,
It is necessary to perform the adjustment at the same time, but this is a difficult task that interferes with each other. Therefore, the present applicant previously applied for [Patent Application No. 1984-333484
As disclosed in "Network Configuration Data Processing Apparatus" filed on December 28, 2013, "1" is always stored in the h layer on the input side.
We proposed that by providing a unit that outputs and assigns a threshold value θ as a weight value to the output, the threshold value θ1 is incorporated into the weight value wihO, and the darkness value θ is treated as a weight value. By doing this, the above equation (11T21 becomes Xpi=Σ3'phWih (3)
Formula)'pt=1/(1+exp(Xpl))
(4) Formula (3) (412) By analogy, the following (51 (61 formula) is obtained.

That is, xpj: Σyp positive W j i (
51 formula'jp== 1 / (1+exp(Xpj))
(Formula 61, where j: unit number W of the j layer, i: weight value of internal connection between the i-j layers xpJ: sum of products of inputs from each unit of the i layer to the first unit of the j layer) 'pjCpth 3 of the j layer for the input signal of the pattern
First, a random number in an appropriate range is used as the initial value of the weight, and each time the cent of the learning pattern is presented, the weight is incrementally improved according to the weight update rule shown below. This process is repeated until the error becomes less than the desired value.

Basically, the idea of the steepest descent method is used as the weight update rule.

First, the sum of squares E of the error between the teacher input vector and the network output vector for each pattern.

and define E as their sum. This teacher human power vector is stored in the memory 12 in association with the time R
This is GB data.

Step 61 Step 62 E-ΣEp However, the teacher signal E2 to the j-th unit of the j-th layer of the dpjep-th pata-th pata-to-human power is called the energy of the node work for the p-th pattern input, and E is the energy of the network. Apply the steepest descent method to this energy.

In the case of the steepest descent method, the t-th weight update amount ΔWjt(t) is determined as follows.

Wji Wik However, wH,: Weight between the hth layer unit and iNith unit wH: Weight between the ith layer i unit and jNjth unit ε (>0): Learning constant (learning rate)
t: Number of Updates In the propagation method, an equation with a momentum term added as shown below is used as a weight update rule for the purpose of accelerating convergence.

However, ε (>O): learning constant (learning rate
)α(>Q): Momentum
m) t: Number of updates The bank propagation method reduces E by repeating this weight update, and finally brings it below the value for which the error is sought.

Next, a method of calculating partial differentials of energy E with respect to Wl, etc. for all patterns required when executing hack propagation will be explained below.

First, we will show the processing equations of each unit in the network.

Xpi-Σ yphW=h Xpj=Σ)'ptWji y,,1. : Output from the h-th unit of the h layer (input layer here) for the p-th pattern input value Ypi: Output from the i-th unit of the i layer (here the middle layer) for the 1)-th pattern input value yl, J: Output from the j-th unit of the j-th layer (here, the middle layer) for the p-th pattern input value xpi: Summation Xpj' for the p-th pattern input to the i-th unit of the first layer The total sum for the p-th pattern human power to the unit of .Here, the partial differential of the energy for each pattern with respect to y□ is determined.

Obtain δ ypJ. The value that came out here is 62. Put it as. That is, δ)
'pJ. Using this δ2, the partial differential of energy with respect to Xpj for each pattern is expressed as follows.

Using this formula, the energy Wj for each pattern is
The partial differential with respect to , is obtained as follows.

Similarly, energy W3 for each pattern. ,
Compute the partial derivative with respect to . First, find the partial differential of the energy E for each pattern with respect to the output ypi of the first layer, σYpi dX pj a Y pi
get. Furthermore, i of the energy Ep for each pattern
Calculating the partial differential with respect to the total sum Xpi to the scrap crabite, δXpi δypJ δxp, -Σ δpj Y pj (1-3' pi)
Y p□ (L y p=) is obtained. By calculating the partial differential of the energy for each pattern with respect to the weight between h-i layers using this equation, we obtain 2 expressed by the following sum of products.

W86 0 Xpi 6Wih The partial differential due to the energy weight for the fully-input cover pattern is expressed as the sum of the partial differential due to the energy weight for each pattern.

δWJI a W j i δWih δWih Steps 63 and 64 are executed as described above.

Next, in steps 66 and 67, 2 (71(8)) is executed. StesonoPro 8 Adds ΔW and □ to the weight Wji before updating to update Wji. Step 69 The same process as in step 68 is performed on Wi.

Step 70 W: Store H, W, h in the memory 12.

Step 71 Steps 63-70 are performed for all weights.

Step 72: Give the data in the memory 21 to the neural network 11 a number of times greater than the number of data, and from the result, step 61 to 7
Perform operation 1.

When the error (Ep) becomes less than or equal to a predetermined number, it is considered that convergence is achieved, and learning is terminated.

Through the above processing, arbitrary month, day, and time data are entered into the neural network 1 whose weight values have been obtained through learning from the parameter input device 7 (specifically, a keyboard or the like). The parameter input unit 5 receives the data, passes it to the neural network execution unit 3, sends the month, date and time data to the neural network 1, and sends the RGB
Get data.

The color of the picture shown in FIG. 9 is changed according to the RGB and outputted to the display device 6.

When making corrections, all you have to do is add data and re-learn.

In the embodiment, time and color information are associated, but it is also possible to express warm, cold, etc. numerically and associate it with color. Furthermore, although the information is displayed on a display, it may also be output using a printer.

As explained above, the present invention can be modified in various ways from the viewpoint of the gist of the present invention, and the present invention does not exclude them.

〔effect〕

As explained above, since the present invention uses a neural network to output colors, it can be easily learned.

Further, according to the present invention, the user does not need to set the coloring situation in detail while drawing a figure, and by specifying only sketch-like line drawing information and rough coloring information, speeding up of drawing can be achieved.

[Brief explanation of drawings]

Fig. 1 is an explanatory diagram of the principle of the present invention, Fig. 2 is an embodiment of the invention, Figs. 3 and 4 are explanatory diagrams of a neural network,
5 is a flowchart of the neural network, FIG. 6 is a flowchart of the neural network operation section 2, FIG. 7 is a flowchart of the preprocessing section 8, FIG. 8 is an explanatory diagram of parameter color information, and FIG. 9 is a conventional example and a flowchart of the preprocessing section 8. It is an example of an output screen in an example. 1 is a neural network, 2 is a neural network operation unit, 3 is a neural network execution unit, 5 is a parameter input unit, 6 is a display device, 7 is a parameter input device, 8 is a color information extraction device, and 9 is an image input device. . 1 Explanation of the principle of the present invention Figure 1. Example Flowchart of the preprocessing section 7th prison January date] [Time delay] ■ ■ [Sky color] (Top color) (RG B) (RG B) ■ ... [Color of will] (RG β) 1.1 12'00 4, + 17100 01g5 +66 103122135202148
20 121 +gs 41'21157152 50158162 Explanatory diagram of parameter color information Fig. 8 Side view of output screen in conventional example and embodiment

Claims (1)

[Claims] An input layer into which parameters are input, a plurality of intermediate layers which calculate and output weighted output values of the input layer;
A neural network (1) comprising a plurality of output layers that perform arithmetic processing on a plurality of values obtained by weighting the plurality of outputs of the intermediate layer and output the arithmetic results as color information.
and means (21) for storing a plurality of parameter color information sets in which parameters and color information are associated with each other, and an output value of an output layer obtained by applying the parameters and color information to the input layer of the neural network. , a neural network learning means (2) for updating a weight value of the neural network based on an error between the color information corresponding to the parameter; and inputting parameters to the neural network trained by the parameter color information set and the neural network learning means. A figure coloring method characterized by comprising a display output means (6) for outputting or displaying a figure colored from the color information obtained.