JPH04153888A - Correspondence device by hop field type neural network - Google Patents

Abstract

PURPOSE:To enable a correspondence process in a practical process time by scanning output values of respective units and storing addition results in registers divisionally by row-directional and column-directional results, performing update calculation based on a specific update rule, and completing the update calculation when the results reaches specific standards. CONSTITUTION:When the output values V(i,j) of the respective units in a correspondence relation table storage part 2 are updated and calculated, the output values V(i,j) are scanned and their addition results are stored in the 1st register Li3 and 2nd register Lj4. Then a correspondence relation calculation part 5 refers to the values stored in the 1st register Li3 and 2nd register Lj4, and updates and calculates the output values V(i,j) of the respective units in the correspondence relation table storage part 2 according to a specific update rule. Then the update calculation results are evaluated and when the results meet the specific standard, the update calculation is completed. Consequently, the correspondence process wherein the results do not depend upon the correspondence order becomes possible and the effective correspondence results are obtained in a practical process time.

Description

[Detailed Description of the Invention] [Summary] Regarding a device that determines a one-to-one correspondence of elements between two sets based on feature quantities extracted from each element of two sets, a correspondence processing algorithm using a conventional search is provided. In this paper, we aim to solve the problem that the results depend on the order of the correspondence processing, and it may not be possible to obtain practical correspondence results, and it is difficult to modify the obtained correspondence results. In a correspondence device using a Hopfield neural network that determines the one-to-one correspondence of each element in a set, there is a storage unit for a similarity matrix representing the degree of correspondence of each element in a set, and an output of each unit of the network. It has a value storage section, two registers, and an update calculation section for the output value of each unit, and when updating the output value of each unit, it scans the output value of each unit and updates the output value in the row and column directions. The system is configured by providing means for storing the addition results in registers in each direction, performing update calculations according to a predetermined update rule, and terminating the update calculations when the results reach a predetermined standard.

[Industrial application field]

The present invention relates to a device that determines a one-to-one correspondence between elements when there are two sets, based on feature amounts extracted from each element of the sets.

An example of the field of application of the device according to the present invention is automatic coloring of animation cells. To create animation, still images (
Although a large amount of cell drawings are required, the labor costs for coloring are extremely high, so automation of the work is desired. In response to this request, the cell image is automatically colored by finding a one-to-one correspondence between closed areas surrounded by outlines between two cell images and painting the corresponding closed areas with the same color. can do. This correspondence device is used to obtain a one-to-one correspondence between closed regions.

At this time, the present correspondence device regards the closed regions of the cell images as elements of correspondence, and determines the correspondence based on the feature amount of the shape extracted from each closed region.

[Conventional technology]

A correspondence determining algorithm used in a conventional correspondence processing device will be described below.

Let the respective sets of processing targets be Fa and Fb. Here, Fa + Fb are Na, Nb elements, e
, and ej.

Fa= (e: l i = 1.-, Na)
(1) Formula Fb= (eJl j=1.-, Nb
) (2) In automatic coloring of expression animation,
The set of closed regions directly in the two cells are Fa and Fb,
The closed regions correspond to e8 and eJ.

In the mapping process, feature values have already been extracted for each element in sets Fa and Fb, and based on the difference in feature values, the correspondence between element e in set Fa and element ej in set Fb is determined. It is assumed that a similarity matrix S, which is a degree of certainty, has been determined.

S= fs (i, j)l i=1. =-, Na; j
=1..., Nb) (3
)2 Each element of the similarity matrix S, S(i, j) is e, and e,
It is set so that the higher the corresponding certainty, the larger the value.

The result of the correspondence processing is expressed as a set P of ordered pairs pk of corresponding element numbers between sets Fa and Fb.

P= (pml p, t= (i, j); = 1...
・, to; i=1. ...-, Na; j=1. -, N
b;) (4) Equation (4) shows that the ordered pair pm = (1, J) has a correspondence relationship between e and e. However, e, and ej
Since the correspondence is one to l, the order pair C of the elements of the set P must satisfy the following equation.

(However, pkεP, pm+EP, l:'hzEP) In the conventional method, in order to obtain the result of the correspondence processing (set P) from the similarity matrix S, (5) a search process that maximizes the evaluation function within the range that satisfies 2 is performed. Let's do it. As an evaluation function, the element S(i, j) of the similarity matrix corresponding to the corresponding elements is
Use the sum of

An algorithm for determining the set P by searching using this evaluation function E is shown in FIG.

This decision algorithm is a process of finding a set P that has the maximum evaluation function E by sequentially adding and deleting correspondences that satisfy (5) 2 to a set P whose initial state is an empty set. It shows. Also, the set Ck is an ordered pair Pi
It represents a candidate set of ordered pairs that can be added to the set P when the correspondences up to +−+ have been determined. The set Ck is
Even if the element is added to the set P, the set P is a set of elements that satisfies equation (5), and for example, when the correspondence relationship in Figure 7 is an element of the set P, the element of the correspondence relationship is The set Ck is composed of the elements inside the mouth excluding the intersecting hatched parts.

Since this search is a type of NP-complete problem, it is known that as the number of elements Na and Nb increases, the amount of calculation required for the exhaustive search increases exponentially.

Therefore, in order to obtain correspondence results (set P) in a practical calculation time, set an allowable processing time, terminate the search within that time, and select the optimal result among the correspondence results already obtained as set P. It is necessary to output it as .

[Problem to be solved by the invention]

As explained above, in the prior art, which elements are matched between sets is determined depending on the set P of the correspondence results determined up to that point. For example, as shown in Figure 7, an ordered pair showing the correspondence of elements of a set with the first element in the row,
If the second elements are arranged and displayed as a column, and the correspondences already registered in the set P are displayed with "○" marks, the similarity information in the diagonally shaded area in Figure 7 will be calculated using the * of the algorithm in Figure 6. Unless the correspondence corresponding to the "O" part is removed from the set P in step B, it will be ignored in subsequent correspondence processing.

Therefore, with the sequential response processing algorithm, practical response results may not be obtained depending on the setting of the allowable processing time.

Furthermore, with the conventional technology, it is difficult to modify a part of the correspondence result. For example, when removing the correspondence pk· from the set P, it is necessary to redetermine the correspondences after p,· (P m, k>k''). Conversely, add the correspondence 26· to the set P. Then, since there is a possibility that the set P does not satisfy equation (5) after the addition, a more complicated correction process than when removing the correspondence is required.

The present invention has been made in view of the above-mentioned problems, and it is an object of the present invention to provide a response device that performs response processing in a practical processing time and that allows easy modification of response results.

[Means to solve the problem]

According to the invention, the above objects are achieved by the means specified in the claims.

That is, the present invention has two sets Fa=(ei11=1 to Na) and Fb=(e
In a correspondence device using a Hopfield neural network that determines the one-to-one correspondence of each element between j l j = 1 to Nb), the correspondence between element ei in set Fa and element ej in set Fb A similarity matrix storage unit (1
) and a correspondence table storage unit (2) that stores output values V(i, j) of each unit of a Hopfield neural network consisting of Na x Nb units.
and the address (i) of the corresponding relationship table storage unit (2).
. It has a correspondence calculation part (5) that performs update calculation of the a force value V(i, j) of each unit in the table storage part (2), and Output value V(i
, j), the output value V(i, j) of each unit in the correspondence table storage unit (2) is scanned and the first register Li(3) and the second Register L
The means for storing the addition result in j(4) and the correspondence calculation unit (5) refer to the values stored in the first register L i (3) and the second register L j (4). means for updating the output value V(i, j) of each unit in the correspondence table storage unit (2) according to a predetermined update rule, and when the update calculation result is evaluated and a predetermined standard is reached. This is a corresponding device using a Hopfield neural network, which is provided with a means for terminating the update calculation at the end of the update calculation.

Further, regarding the invention as claimed in claim 2, in the correspondence device using the Hopfield neural network as claimed in claim 1, in order to correct the correspondence result, the output value V(i, j ) and a means for prohibiting updating of the output value V(i, j) of the corrected unit to the correspondence calculation unit (5), and restarting the corresponding device. This is a correspondence device using a Hopfield neural network that obtains a corrected correspondence result.

Furthermore, regarding the invention as claimed in claim 3, together with the corresponding device using the Hopfield neural network as claimed in claim 1 or claim 2, an image input unit ( 11), a label image storage unit Q2) that adds and stores an identification label to a closed area in the image for each color image, and what color each label area of one image (Fa) was. a label-color correspondence table storage unit (13a) indicating the label image storage unit G2; a region feature storage unit (15) that stores the region features extracted by the unit (14);
), the region feature storage unit (similarity generation unit 06 that calculates similarities for all combinations of each label region from the region features of the two images stored in 19), and the correspondence relationship determined by the corresponding device. , the colored image creation unit (19) refers to the label/color correspondence table storage unit (13a) of one image (Fa) and colors the label area of the other image (Fb) with the corresponding color. This is an automatic coloring device for animated images.

Hereinafter, the corresponding devices according to claims 1 and 2, which are the basis of the present invention, will be further explained.

Compatible bag 1 of the present invention is based on the use of a neural network for solving the traveling salesman problem proposed by J. J. Hopfield (J. J. Hopfielda
nd D, W, Tank (1985), “Neura
l computation inoptimizat
ion problems, Biol, Cyb
ernetics52, 141-152. ”).

This Hopfield type neural network is known to be able to approximately solve NP-complete problems in a short time, but in order to make it into an actual device, the connection relationship between each unit is calculated using a value called the synaptic load. Requires a large amount of memory to store.

Therefore, in the compatible device of the present invention, the compatible device is realized without requiring a large amount of memory by the following method.

That is, FIG. 1 shows the principle configuration diagram of the present invention,
1 is a similarity matrix storage unit that stores the similarity matrix S(i, j), 2 is a correspondence table storage unit that stores the output values of each unit of the Hopfield neural network, and 3 is the output value V (i.

A first register (Li) that stores the total sum of j)
, 4 is a second register (Cj) that stores the sum of the output values V (i, j) in the column direction, and 5 is an update calculation of the output values V (i, H) according to a predetermined update rule (described later). It represents a correspondence calculation section.

As shown in FIG. 1, in the correspondence device of the present invention, the correspondence is expressed by the output value of the unit placed in each cell in the correspondence table storage section 2.

For example, element e in set Fa and element e in set Fb,
The correspondence relationship of is expressed as the output value of Uniso t-(i, j) ■(i, j
) is expressed as follows.

The value of V(ij) takes a real value between "0.0 and 1.0" because the correspondence relationship is not determined during the correspondence process. The condition shown in equation (5) is that only one unit output in both the vertical and horizontal directions on this correspondence table is "1".
This corresponds to a value of "0", and other units have a value of "0.0°".

Further, in the correspondence calculation section 5, the initial value of the unit output is set appropriately, the unit output value is repeatedly updated by the following calculation, and the correspondence is determined from the unit output value when converged.

Here, A and B are constants determined a priori.

Correcting the correspondence result is performed by changing the output value of the unit corresponding to the correspondence to be corrected and repeating the calculations of equations (8,1)2 to (8,4) again.

Further, in the invention as set forth in claim 2, the output value V(i, j) of the unit in the correspondence chifur storage section 2 is modified as follows.

e, and ej are compatible → Not compatible: V (i, j) Corrected to -0,0 e, and eJ are not compatible → Corresponding: V (j, j) = Corrected to 1.0 (8, 1) 2 The calculations of (8, 4) 2 are repeated using the unit output after the correspondence has been corrected as the initial value. However, the correspondence calculation section 5 does not update the corrected unit output values.

[Operation] In the correspondence device of the present invention, based on the similarity matrix S(i, j) in the similarity matrix storage unit 1, the output (JV(i,

The update calculation of j) is performed to determine the correspondence relationship, but the update rules of equations (8.1) 2 to (8, 4) used for the above update calculation end up determining the value of the following evaluation function F. The output value V(i,j) of the unit is changed so as to decrease the output value V(i,j).

F=A [Σ(ΣV(i, j) −1) ” +Σ(
ΣV(i, j) 1)”] B(ΣΣS(i, j)
V(i, j)) (9) where A and B have the same values as in equation (8,2).

The first term on the right side of equation (9) is set to be the minimum when the condition shown in equation (5) is satisfied, and the second term is set as (6).
It is minimum when the expression is maximum.

Therefore, the convergence value of the unit output by repeated calculation of equations (8,1)2 to (8,4) represents the result of the corresponding processing.

Further, in the case of the invention described in claim 2, unlike the prior art, the handling device of the present invention does not perform the handling processing sequentially, so there is no possibility that a practical solution cannot be obtained depending on the order of handling. Also, for the same reason, when modifying the correspondence, after making partial modifications, (8, 1) 2 ~ (8
, 4), it is possible to obtain a correspondence result that restores the one-to-one correspondence relationship broken by the correction.

Further, the invention as claimed in claim 3 is an example in which the corresponding apparatus as claimed in claim 1 or 2 is applied to an automatic coloring apparatus for animation images, and the details thereof will be explained in the embodiment section.

Nao, Hopfield type neural network calculations are:
Originally, the following equation is used to calculate U(i, j).

U(i, j)←ΣΣT (i+j, i'+j')V(
i, j)? (i, j) (9,1) However, in the corresponding processing handled by the present invention, T(i+j+i'
, j') and I (i, j) are calculated using the following equations.

T(11J11'|j') is the synaptic load, and in order to preserve this directness, ((Na×Nb)"X32bit)
memory is required. (For example, 4 MB of memory is required to associate 10 elements with each other.) However, T(1+j+IZJ") in equation (9,2)
Since the value of can be easily obtained by dividing the cases using the following equation, there is no need to calculate and store the value in advance.

Therefore, U(i
, j), a large amount of memory for storing T(i, j+i'+j'') can be omitted.

(Embodiment) FIG. 2 is a diagram showing an embodiment of the present invention, and shows an automatic coloring device for animation images using the present invention.

This automatic coloring device includes an image input section 11, a label image storage section 12, a label-color correspondence table storage section 13a, a feature extraction section 14, a region feature storage section 15, and a similarity generation section 1.
6, a correspondence processing section 17, a correspondence correction section 18, and a colored image creation section 19.The inputs to this automatic coloring apparatus are a colored image a and a line image to be colored.

The operation of this automatic coloring device will be explained below.

(1) The image input unit 11 reads the color image a and the line image S into the computer using an image input device such as a camera,
Perform different processing on each.

For the color image a, a label image Fa is created in which a label (i=1..., Na) is attached to each closed region colored in the same color, and this is stored in the label image storage unit 12. The information is stored in the section 12a. Further, a label-color correspondence table indicating which color each label area has is created and stored in the label-color correspondence table storage section 13a.

For line images, after removing the background, we extract closed regions surrounded by lines, and label each closed region (j=1.
..., Nb) is added to create a label image Fb.

Thereafter, the label image Fb is stored in the storage section 12b in the label image storage section 12, and the process is ended.

(2) After the image input unit 11 completes the image reading operation, the region feature extraction unit 14 is activated. The region feature extraction section 14 extracts region features for each label head region from the label images stored in the label image storage section 12 and stores them in the region feature storage section 15 .

There are six types of region features to be extracted: area, perimeter, complexity, chain code histogram, center of gravity position, and principal axis of inertia direction. The area is the number of pixels in the label area on the frame memory, the perimeter is the number of pixels on the outline of the label area, the complexity is the square of the perimeter divided by the perimeter, and the chain code histogram is the number of pixels in the principal axis of inertia direction. The outline of the label area when aligned with the joint axis of the label area is represented by chain codes in 8 directions, and the frequency table is accumulated for each code from 0 to 7. The center of gravity position is the coordinate value on the frame memory of the center of gravity of the label area, and the inertia. The ball axis direction is the counterclockwise angle formed by the principal axis of inertia of the label area and the horizontal axis of the image. The area feature extraction unit 14 transfers the area features of the label image Fa stored in the label image storage unit 12a to the area feature storage unit 15a.
The area features of the label image Fb stored in b are respectively stored in the one area feature storage unit 15b.

(3) The similarity creation unit 16 uses the area feature of the label i area in the label image Fa stored in the area feature storage unit 15a and the label j in the label image Fb stored in the area feature storage unit 15b. From the region feature of the region, all i and j
The similarity, S(i, j), is calculated for the combination of . In this embodiment, the above six types are used as area features to determine label i area and label j? Calculate the similarity between il regions. The value of S(i,j) takes a real value from [0 to 1], and is set to take a value closer to 1 as the difference between the area features of the two areas is smaller. The calculated similarity value is sent to the correspondence processing unit 7.

(4) As shown in FIG. 3, the correspondence processing section 17 includes a similarity matrix storage section 21, a correspondence table storage section 22, a frame memory scanning section 27, a register section (Li) 23, and a register section (Cj) 24. , a correspondence evaluation section 26, and a correspondence calculation section 25.

That is, the similarity sent from the similarity creation unit 16 is stored in the similarity matrix storage 21 and the correspondence table storage 22. Both the similarity matrix storage unit 21 and the correspondence table storage unit 22 are configured with a Na x Nb x 32-bit frame memory, and the address (i, j)
S(i,j) is stored in . The reason why the similarity is stored in the correspondence relation table storage unit 22 is to use the similarity matrix as the initial value of the correspondence table.

(5) After the storage of the similarity is completed, the correspondence processing unit 17 repeats the following steps (a) to (C) until the correspondence evaluation unit 26 issues a termination signal.

(a) The frame memory scanning section 27 first scans the Na registers of the register section (Li) 23 and the register section (Cj
) 24 are set to "0".Next, the correspondence table storage unit 2 is scanned by rask and the address (
The value of i, j) and V(i, j) are added to the values of register (Li) 23 and register (Cj) 24 and saved.

In parallel with this rask scan, the frame memory scanning section 27 performs a rask scan on the similarity matrix storage section 21 to obtain S(i,j).
and V(i,j) and S(i.

Calculate the sum of products W of J).

W= (ΣΣS (i, j)V (i, j)) (1
When the 0 type past scanning is completed, the frame memory scanning section 2
7 sends the activation signal and the value of W to the correspondence evaluation section 26 and ends the process.

('b) When the correspondence evaluation section 26 receives the activation signal, it calculates the value of W, the register section (Li) 23, and the register section (Cj).
24, the next evaluation function value F is calculated.

-1,0) 2] -BXW (12j
where L(i) is the value of register (Li) 23, C(
j) is the value of the register (Cj) 24, and A and B are constants determined a priori. The correspondence evaluation section 26 monitors the value of F, and when the value of F has not decreased compared to the previous calculation, it generates an end signal to terminate the processing of the correspondence processing section 17, and the correspondence correction section 18 Start.

In other cases, the correspondence calculation unit 25 is activated.

(C) As shown in FIG. 4, the correspondence calculation section 25 is composed of a processing prohibition flag storage section 25a and a calculation section 25b.

The processing prohibition flag storage section 25a is Na X Nb X 1
The value of 0 is normally stored in the bit memory, and the content is rewritten by the correspondence correction unit 18.The calculation unit 25b stores the content (
When it is "1", no processing is performed, and when it is "0", the next processing is performed by referring to the correspondence table storage section 22, register section (Li) 23, and register section (Cj) 24. Calculate V(i,j) using the formula.

■ (U(i, j) in equation 19 is calculated using the following equation.

U(i, j)←−A ((L (i) +C(j)−
2・■(i, j) -0,5) +B-3(i, j
) (14) A and B in equation (5) take the same values as in equation (3). The calculated V(i, j) is sequentially stored at the address (i, j) of the correspondence chifur storage section 22. V(i, j
), the correspondence calculation section 25 restarts the frame memory scanning section 27. (Repeated calculations up to this point.) (6) As shown in FIG. 5, the correspondence correction section 18 is composed of a label image conversion section 18a, a frame memo 8b, a display 18c, a tablet 18d, and a correction processing section 18e. . The label image conversion unit 18a refers to the correspondence table storage unit 22 and the label image storage unit 12b, and creates a label image Fb′ in which the label j of the label image Fb is rewritten to a label i corresponding to the label j, The data is stored in the frame memory 18b. The display 18c displays the label image Fa on the label image storage section 12a and the label image Fb' on the frame memo 8b in an arbitrary color corresponding to each label. The operator compares the label images Fa and Fb' and displays the display 18.
The correspondence relationship to be corrected is specified using the tablet 18d linked with c. The correction processing section 18e operates the correspondence table storage section 22 and the processing prohibition flag storage section 25a based on the correspondence relationship to be corrected. Correspondence between area j in label image Fa and area j in label image Fb (i, j
), write “1.0” and “1” to the addresses (i, D, respectively) of the correspondence table storage unit 22 and processing prohibition flag storage unit 25a, and add the correspondence relationship (i, j).
When deleting the correspondence relationship (i, Since it is specified to change the correspondence relationship (i'', j) to the correspondence relationship (i'', j), the correspondence relationship (i, j) is removed and the correspondence relationship (i'', j) is added at the same time. Correspondence table storage After the operations of the section 22 and the processing prohibition flag storage section 25d are completed, the correspondence correction section 18 restarts the correspondence processing section 17. If the correspondence is not corrected, the correspondence correction section 18 restarts the colored image creation section 19. Start.

(7) The colored image creation unit 19 refers to the label-color correspondence table storage unit 13a and creates a frame memo U18b.
The label area of the label image Fb" stored in the label image Fb" is filled in with the color corresponding to the label and output as an automatic coloring result.The output result is saved in a recording device such as MT (I tape) and used for creating animation. Ru.

[Effects of the Invention] As explained above, in the present invention, when performing correspondence processing, the Hopf 1eld is used which omits the memory for storing the value called the synaptic load indicating the connection relationship between each unit.
By using a type neural net, it is possible to perform correspondence processing in which the results do not depend on the order of correspondence, unlike the correspondence processing by searching in the conventional technology, and it is now possible to obtain effective correspondence results in a practical processing time. Become. Also,
Since no special processing is required to perform the correction, it becomes possible to realize a compatible device with a simple configuration that implements the correction processing.

[Brief explanation of drawings]

Figure 1 is a diagram showing the principle configuration of the present invention, Figure 2 is a diagram showing an embodiment of the present invention, Figure 3 is a diagram showing the configuration of the corresponding processing section, and Figure 4 is a diagram showing the configuration of the corresponding processing section.
The figure shows the configuration of the correspondence calculation section, FIG. 5 shows the configuration of the correspondence correction section, FIG. 6 is a diagram for explaining the prior art, and FIG. 7 shows an example of determining the set Ck according to the prior art. FIG. DESCRIPTION OF SYMBOLS 1... Similarity matrix storage unit, 2... Correspondence table storage unit, 3... Register (Li), 4... Register (Cj), 5... Correspondence calculation unit, 11... - Image input section, 12... Label image storage section, 12a... Storage section for image a, 12b... Storage section for image a, 13a
...Label/color correspondence table storage section, 14...
Region feature extraction unit, 15...Region feature storage unit, 15a.
... Storage unit for image a, 15b... Storage unit for image a, 16... Similarity creation unit, 17... Correspondence processing unit, 1
8... Correspondence correction section, 18a... Label image conversion section,
18b... Frame memory, 18c... Ice play, 18d... Tablet, 18e... Correction processing section, 19... Colored image creation section, 21... Similarity matrix storage section, 22. ...Correspondence table storage section, 23...
・Register (Li), 24...Register (Cj), 2
5... Correspondence calculation unit, 25a... Processing prohibition flag storage unit, 25b... Calculation unit, 26-Correspondence evaluation unit, 27... Frame memory scanning unit

Claims (3)

[Claims] (1) Two sets Fa=(ei|i=1~Na), Fb
In a correspondence device using a Hopfield neural network that determines the one-to-one correspondence of each element between = (ej | A similarity matrix storage unit (1) that stores a similarity matrix S (i, j) representing the corresponding degree of certainty, and an output value V of each unit of a Hopfield neural network consisting of Na×Nb units. (i, j), and the address (i, j) of the correspondence table storage unit (2).
), a first register Li (3) that adds and stores the value V(i, j) in the row direction, a second register Cj (4) that adds and stores the value V (i, j) in the column direction, and a correspondence table storage section. (
2), and a correspondence calculation unit (5) for updating the output value V(i, j) of each unit in the correspondence table storage unit (2). , j), the output value V(i, j) of each unit in the correspondence table storage section (2) is scanned and the first register Li(3) is
and means for storing the addition result in the second register Lj (4); and the correspondence calculation unit (5) stores the addition result in the first register Li (3
) and the values stored in the second register Lj (4), the correspondence table storage unit (2
), and a means for evaluating the update calculation results and terminating the update calculation when a predetermined standard is reached. A correspondence device using a characteristic Hopfield neural network. (2) In the correspondence device using the Hopfield neural network according to claim 1, in order to correct the correspondence result, the output value V (i, j) of the unit corresponding to the correspondence to be corrected is determined.
and a means for prohibiting updating of the output value V(i, j) of the corrected unit to the correspondence calculation unit (5), and restarting the corresponding device. A correspondence device using a Hopfield neural network characterized by obtaining a corrected correspondence result. (3) An image input unit (11) for reading two color images (Fa, Fb) into the device together with a corresponding device using the Hopfield neural network according to claim 1 or claim 2, and for each color image. A label image storage unit (12) that stores identification labels added to closed areas in the image, and a label-color correspondence table that shows what color each label area of one image (Fa) was. Storage part (1
3a), and a region feature extraction unit (14) that extracts features for each label region from the label images in the label image storage unit (12).
and a region feature storage section (15) that stores the region features extracted by the region feature extraction section (14), and each label is extracted from the region features of the two images stored in the region feature storage section (15). One image (Fa
An animation characterized by comprising a colored image creation unit (19) that refers to a label-color correspondence table storage unit (13a) of ) and colors the label area of the other image (Fb) with a corresponding color. Automatic image coloring device.