JP2002245449A - Adaptive nonlinear mapping device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To obtain an adaptive nonlinear mapping method that efficiently performs accurate mapping. SOLUTION: The images of a mapping source and a mapping destination are segmented, the original images and the segmented images are divided into very small areas, and the characteristic quantity of each of the areas is extracted. In the image of the mapping destination, potential is formed in each characteristic quantity, and a motion vectors are obtained by the very small areas of the image of the mapping source in a potential steepest direction. Next, the obtained motion vectors are made to cooperate with each other between adjacent very small areas. A vector group representing the mapping relation between the image of the mapping source and the image of the mapping destination is found out by repeating the motion vector acquisition and the cooperation operation, and more accurate and more efficient mapping is carried out by performing mapping on the basis of the vector group.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for normalizing an image or extracting a change area between two images having different photographing conditions and photographing times in medical images, aerial photographs, satellite images and the like. The present invention relates to a non-linear mapping technique, which is an important technical element for performing the above, and particularly to an adaptive non-linear mapping method for forming an effective non-linear mapping by greatly improving a conventional method.

[0002]

2. Description of the Related Art Medical images such as MR (Magnet) of the brain
tic Resonance) images and CT (Compu)
In a ted Tomography image or the like, each image itself provides important medical information. However, if a plurality of images having different imaging conditions, imaging times, and the like can be compared and a change area can be accurately extracted, the density can be further increased. Useful medical information can be obtained. For example, in brain disease diagnosis,
By comparing MR images of the brain of an individual taken at a time interval of several years, the type of disease and its progress can be accurately grasped.However, MR images often differ due to aging of the device. Therefore, a mapping that absorbs this distortion is required.

When comparing a patient's brain with a standard brain,
Alternatively, when comparing a large number of human brains in a group, it is necessary to normalize the brain images to be compared, but the non-linear mapping technique is a useful means for that purpose.

[0004]

As an example of such a mapping technique, the inventors of the present invention disclosed in Japanese Patent No. 2860048.
No. "Nonlinear mapping device and its method" has been proposed. In this apparatus and method, the image of the mapping source is divided into minute regions, the minute region is slid on the image of the mapping destination to search for a portion having the same pixel value, and this operation is repeated over all the minute regions. Forming a mapping. However, in this technique, since the mapping destination is searched based on only the pixel value of the minute image, that is, only the black and white information, it is difficult to narrow down the area on the image of the mapping destination to which this minute region may correspond. As a result, it was necessary to search a wide area by exhaustive search. Therefore, the amount of calculation for the mapping becomes enormous, and particularly when a large image is mapped, a huge search time is required, and there is a drawback that the mapping is practically impossible.

[0005] Also, using only the pixel value as a reference,
Because the search for the corresponding part of the mapping is performed, when there is a slight difference in the shooting conditions between the two images, the mapping is formed in a place other than the original corresponding point, and the corresponding mapping destination can be found. In some cases, there was no problem, and accurate mapping could not be performed.

SUMMARY OF THE INVENTION An object of the present invention is to solve the above-mentioned drawbacks of the conventional nonlinear mapping technology and to provide an adaptive nonlinear mapping method capable of performing efficient and accurate mapping.

[0007]

According to a first feature of the present invention, before starting mapping between two images, each of the images is converted into a configuration attribute of the image (for example, in the case of a tomographic image of the brain, Is
Area segmentation (hereinafter, segmentation) according to brain gray matter, white matter, cerebrospinal fluid (CSF), etc.
The feature of.

As an example of the segmentation, the present inventors have already proposed a method in which a feature amount is extracted by combining a genetic algorithm and a neural network and the segmentation is executed in accordance with the extracted feature amount (Patent Document 1). No. 2873955). In this segmentation, an advanced image recognition system that also takes into account the surrounding image statistics can be used, and has the advantage of being less susceptible to changes in imaging conditions and noise.

Accordingly, for example, segmentation of each image is performed on the original image and the destination image according to the method of Japanese Patent No. 2873955, so that white matter becomes white matter. By performing the above, it is possible to configure a stable mapping system in which the boundaries between tissues and the like match with high accuracy.

According to a second feature of the present invention, at the time of forming a mapping, the attributes of each region of the mapping destination or the mapping source are represented by potentials, and each minute region of the mapping source having the corresponding attribute is represented by this potential. By using the gradient method to approach the corresponding mapping destination using the gradient method, an efficient nonlinear mapping is formed.

In the process of forming nerve conduction pathways in the human brain, NGF (never growth) is obtained from target cells.
a nerve attractant called factor) is released,
The growing nerve axon reaches the target cell depending on the concentration gradient of NGF. Therefore, in the second aspect of the present invention,
When forming a mapping, instead of searching in all directions on the image of the mapping destination, a potential corresponding to the NGF gradient is artificially generated, and a route from the mapping source to the mapping destination is formed along this potential. By doing so, exploration in an area directed to an unnecessary direction is omitted, and the original mapping destination can be found in a short time.

That is, the above-mentioned problem is solved by converting the image A into the image B
An adaptive non-linear mapping method for mapping to the image A
A segmentation step of performing segmentation of the images A and B based on the configuration attributes of the two images; and a division step of dividing the images A and B and the image A ′ and the image B ′ after the segmentation into minute regions. A feature amount extracting step of extracting a feature amount of the divided minute region, a step of forming a potential for each of the extracted feature amounts of the image B, and a step of forming the potential of the image A in the steepest potential direction of the image B. A moving step of moving an area, a coordinating step of coordinating a moving vector at the time of the moving of the image A between adjacent minute regions in the image A, and the moving step and the coordinating step are alternately repeated, and Obtaining a group of vectors representing the mapping relationship of the image B; and Comprising the step of performing a mapping to the image B, it is possible to solve the adaptive nonlinear mapping method.

The above object is also achieved in an adaptive nonlinear mapping method for mapping an image A to an image B, wherein the images A and B
A segmentation step of performing segmentation based on the configuration attributes of the two images,
A dividing step of dividing the image B and the image A ′ and the image B ′ after the segmentation into minute regions;
A feature value extracting step of extracting a feature value of the divided small area; a step of forming a potential for each of the extracted feature values of the image A; A moving step of moving the minute area, a coordinating step of coordinating a moving vector at the time of the movement of the image B between adjacent minute areas in the image B, and the moving step and the coordinating step are alternately repeated to obtain the image A. And a step of calculating a vector group representing a mapping relationship between the image A and the image B, and a step of executing mapping of the image A to the image B based on the determined vector group.

The above object is further achieved by an adaptive nonlinear mapping method for mapping an image A to an image B.
Dividing the image B into minute regions; extracting a characteristic amount of the divided minute regions; forming a potential for each of the extracted characteristic amounts of the image B; Moving the minute area of the image A in the steepest potential direction;
A coordination step of coordinating a movement vector during the movement of the image A between adjacent micro regions in the image A;
A step of alternately repeating the moving step and the coordination step to obtain a vector group representing a mapping relationship between the image A and the image B, and a step of executing the mapping of the image A to the image B based on the obtained vector group. The problem is solved by an adaptive nonlinear mapping method comprising:

The above object is further achieved by an adaptive non-linear mapping method for mapping an image A to an image B.
A segmentation step of performing segmentation based on the configuration attributes of the two images;
A dividing step of dividing the image B and the image A ′ and the image B ′ after these segmentation into minute regions;
A feature value extraction step of extracting a feature value of each of the divided minute regions; a movement vector calculation step of calculating a movement vector of the minute region based on the extracted feature values; and a movement vector calculation step. A step of coordinating the obtained motion vector between adjacent minute regions in the image A; and a step of repeatedly repeating the motion vector calculation step and the coordination step to obtain a vector group representing a mapping relationship between the image A and the image B. Performing the mapping of image A to image B based on the determined vector group.

[0016]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The adaptive nonlinear mapping method of the present invention will be described below in detail with reference to embodiments. FIG. 1 is a flowchart showing the first embodiment of the present invention. In the present embodiment, for example, the MR tomographic image A of the brain matches each of the microstructures with the MR tomographic image B of the same person's brain taken at a different time from the image A. The purpose is to perform a non-linear mapping that does not occur.

In FIG. 1, first, steps 1A and 1B
Then, the image A of the mapping source and the image B of the mapping destination are input.
Next, in steps 2A and 2B, the input images A and B
Each of the pixel values of the image, various statistics defined in relation to the surroundings, and global a priori information on the target image (for example, each layer of the muscle layer constituting the head, skull, gray matter, white matter Segmentation (region division) into a plurality of elements (features) based on, for example, the adjacency order relationship of the elements.

In the next steps 3A, 3B, the original images A, B and the segmented images A ', B'
Is divided into minute regions (for example, lattice regions). Each divided small area contains multiple elements (for example, white matter and gray matter)
May be included or only a single element may be included, but based on this component pattern and the global positional relationship where each minute area is placed in the entire image,
The feature amount of each minute area is extracted (steps 4A and 4A).
B). It should be noted that pixel values are obtained as feature values from the minute regions obtained from the original images A and B.

Next, in the channel on the image B side, a potential is formed around the minute area for each feature amount obtained above (step 5B). This potential is
Assuming that the distance from the minute region is x, it is defined that x monotonically decreases or monotonically increases.

Next, in step 6, each minute area of the image A is slightly moved in the steepest direction of the potential formed by the corresponding feature of the image B for each feature. At this time, as shown in step 7, each minute area of the image A cooperates with the movement direction of the surrounding minute area with respect to the movement vector at the time of movement to avoid tearing of the entire image.

The movement of the minute area of the image A along the gradient of the potential and the cooperative operation with the surrounding movement vector, that is, steps 6 and 7, are based on the degree of matching between the image A and the image B (for example, the square error of the pixel value). Is continued several times to several tens of times (step 8) until the sum of the two becomes smaller than a preset threshold value or the preset number of rounds of the loop is reached (Step 8). The mapping is realized (step 9).

Although the first embodiment describes a two-dimensional image, the same applies to three-dimensional data defined by voxel values (Voxel values) in a three-dimensional space such as brain MR data. Obviously the procedure is applicable.

In the first embodiment, segmentation is first performed on a target image, and then a minute area is moved by potential formation. However, depending on the type and quality of the target image, the segmentation part in the first half of the processing is omitted, the feature amount (for example, pixel value) of the image is directly extracted from the original image, and the potential is formed according to the feature amount. It is also possible to move the minute area along the potential gradient. This embodiment will be described in detail in the description of the third embodiment of the present invention.

Further, when the nonlinearity between the image A and the image B is small (that is, when the required movement amount of the minute area is sufficiently small), the "potential formation / gradient" in the latter half of the processing of FIG. Instead of “movement by pixel”, a search between pixel values or a norm of a vector formed from the result of pixel value and segmentation, which is used in a conventional method (for example, a technique disclosed in Japanese Patent No. 2860048) It is also possible to collate and search for. This embodiment will be described in detail in the description of the fourth embodiment of the present invention.

Hereinafter, the execution of the mapping based on the above-described segmentation and potential formation will be described in more detail with reference to an embodiment. In the following, an operation example of the multi-potential method for realizing a non-linear mapping (image A to image B) between two similar images A and B will be described.

First, the image B is subdivided into an M × N area in a grid pattern, and for each of the obtained micro areas, the characteristic amount of the micro image is extracted. In the following example, a simple binary image is used for simplicity of description, and the size of each minute region is 2 × 2 pixels as shown in FIG. Therefore, when the feature amounts are not aggregated, 2 ⁴ = 16 image elements (feature amounts)
Can be held. Examples of these elements are shown in FIG.
Is shown in

Each of these image elements on the image B forms a unique potential over the entire image. At this time, the potential at a point at a distance x from the center of each image element is equal to that distance. It has a scalar quantity that is inversely proportional to the square. That is, the potential φη created by the ηth image element is expressed by the following equation, where Kη is a positive constant.

Φη = Kη / 2 × ²

The maximum gradient at the point of the distance x of this potential is obtained by differentiating the above equation with respect to x, and fη = −Kη / x ³ .

On the other hand, the image A is similarly subdivided into minute regions, and the attribute η of each image element is given. Each image element on the A side senses only the potential φη of the corresponding attribute, and moves slightly in the maximum direction of the potential gradient. This movement is shown in FIG. In this figure, the movement trajectories of two image elements are displayed.

In the above example, since the amount of one movement is limited to adjacent pixels, the movement direction is limited to only the horizontal and 45 degrees directions, but generally moves in the maximum potential gradient direction. Shall be. In addition, in order to prevent the phase structure from being disturbed due to the individual movement of the image elements, a cooperative operation between the movement vectors is performed at each movement. This method includes, for example, a method of obtaining an average value and a method of obtaining a median (median value).

The potential generated by each of the minute regions constituting the image B is the same as the potential of the same attribute from a plurality of points of the image B when the image is simple or when the type of the feature amount defined in each minute region is small. May be generated. When a small image is moved by being guided by such a potential having a plurality of sources, there are places where the 1: 1 mapping relationship cannot be maintained (see FIG. 4).

In such a case, it is necessary to take measures such as increasing the types of the feature values or taking into account the relative positional relationship in each image as a part of the feature value. Also, in order to avoid the case where the corresponding image element cannot be found due to the difference in the way of cutting the minute area, the B side creates a potential even when the grid is shifted vertically and horizontally, and the sum of these cases is calculated. Create a potential.

The mapping method according to the present potential method is particularly effective when realizing a mapping in which one image is greatly distorted and a large moving amount is required. For example, A, B on the image
Assuming that the position between them is n pixels away, the conventional two-dimensional search range needs to search for a square area centered on point A, that is, an order of 2n ² pixels. In this method, even when repetition of movement to an adjacent pixel having the largest number of steps is used, point B can be reached with the number of searches on the order of 8n. When the ratio of the number of steps is n = 100, n / 4 = 25, and the minute area can be moved from A to B with about 1/25 steps. Furthermore, when a method such as the steepest descent method is applied to the gradient of each potential,
Further higher speed is possible.

Next, the mapping using the segmentation result will be described. The purpose of this embodiment is to perform non-linear mapping of the brain MR image A in FIG. 5A so as to match the reference image B in FIG. 5B. In the present embodiment, the two images A and B are firstly classified into the brain parenchyma, muscle and skull, and cerebrospinal fluid.
Implement segmentation into two categories. Based on this classification result, each pixel is characterized by a discrete value s representing its attribute, and this value is combined with a vector x obtained by vectorially combining the original pixel value p.
This point is defined as a feature vector. That is, i and j are unit vectors orthogonal to each other, and α and β are positive real constants.

X = αpi + βsj Further, in order to judge the quality of the matching of the small areas of the two images A and B, the norm ε of the difference between the two vectors is evaluated as follows. ε = ‖x _A - x _B ||

FIG. 6 shows a composite vector x represented by a discrete value s and a pixel value p, which are the segmentation results.
The matching between the image A and the image B is executed by selecting an image having the minimum ε in FIG.

That is, ε = 0 means that the pixel values and the segmentation results of the two minute regions completely match.

FIG. 5 (c) shows the result of executing the mappings (a) to (b) according to the present method using the above-mentioned matching score criterion, and FIG. 5 (c) shows almost the same as FIG. It can be seen that a matched image has been obtained. The outline of the brain parenchyma marked with a white line in the figure represents the outline of the brain parenchyma in (b). Therefore, the contour coincidence between the mark and the brain substance on the image indicates the completeness of the mapping. On the other hand, in the conventional method using only pixel values, it can be seen that the brain parenchyma in FIG. 5A cannot move sufficiently and only an incomplete mapping can be realized (see FIG. 5D).

FIG. 7 is a flowchart showing the flow of processing according to the second embodiment of the present invention. In the present embodiment, instead of forming a potential on the mapping destination, a potential is formed on the mapping source side, and each minute area of the mapping destination determines from which source of the mapping source the mapping is received, based on the maximum inclination direction of the corresponding potential. It is characterized by going back to finding the corresponding mapping source. More specifically, the image of the mapping source and the image of the mapping destination are appropriately divided into regions, and the degree of correlation between each region of the mapping source and each region of the mapping destination is evaluated. The correlation distribution thus obtained is used in the same manner as the potential shown in the first embodiment. At this time, the processing time is reduced by evaluating the correlation between the divided regions of the mapping source and excluding the divided regions having many highly correlated divided regions, that is, the divided regions having no individuality from the search processing. .

Next, the operation of the second embodiment will be described with reference to the flowchart of FIG. First, the mapping source image A
And the destination image B (steps 11A and 11A).
B) For both images A and B, segmentation is performed based on the organization and image attributes (steps 12A and 12B), and then divided into minute regions like a grid (steps 13A and 13B).

Next, in steps 14A and 14B,
The image feature amount is extracted for each minute area of the images A and B. As the feature amount, for example, a value obtained by combining a pixel value and a discrete value indicating a segmentation result in a vector manner is used. In particular, for the image A, when the feature amounts calculated for each minute region in the image A have a large number of overlaps (that is, a large number of minute regions have substantially the same feature vector), Since it is difficult to regard the feature amount as an effective feature amount, the feature amount relating to these minute regions is rejected (step 15A, selection of effective feature amount).

Next, a feature amount that has been avoided is set as an "effective feature amount", and a potential that monotonically decreases with respect to the distance is formed around a minute area of the image A having this feature amount (step 16A).

Next, for each minute area of the image B, a search is made for which minute area of the image A is to be mapped. This search is performed in such a manner that the potential on the image A corresponding to each feature amount is traced back to the steepest gradient to obtain the image element of the mapping source (step 17, search for the receiving target area).

The points adjacent to each other on the image B are defined as receiving vectors (= movement vectors) having a close relationship based on the condition of continuous mapping that the areas adjacent to each other on the image A should be regarded as "accepting areas". A mutual operation is performed by a median value calculation or an average value calculation (Step 1).
8).

The above-described movement based on the potential (step 17) and the cooperation between the receiving vectors (step 18) are alternately repeated until a group of vectors representing a mapping relationship is obtained (step 19). A vector group representing a mapping relationship is obtained. Therefore, a mapping is performed in step 20 based on this vector group.

For an image element for which no correspondence is found because it does not have an effective feature, a continuous mapping is formed by a cooperative operation and an interpolation operation with surrounding motion vectors having an effective feature.

As in the first embodiment described above,
When the number of effective feature values is extremely small, the method according to the present invention is used in the conventional method (Japanese Patent No. 2860048) to automate the part where a small number of corresponding points (markers) are designated by human intervention and the last mapping. The formation can also be realized by a hybrid method using a search method using pixel values as in the conventional method, and as a whole system, an extremely accurate mapping system can be realized without human intervention.

FIG. 8 is a flowchart showing a third embodiment of the present invention. This embodiment is characterized in that a segmentation is not performed on an image of a mapping source, a feature is calculated based on a pixel value, and then a movement vector is obtained by using a potential method and mapping is executed.

First, in steps 21A and 21B,
A source image A and a destination image B are input. These images A and B are divided into minute regions in steps 22A and 22B, and a feature amount is extracted for each minute region (steps 23A and 23B).

With respect to the image B, a potential is formed for each feature in step 24B, and in step 25, a minute area of the image A is moved in the steepest direction of the image B to obtain a movement vector. The movement vectors are coordinated between adjacent minute areas in the image A (step 26). The operations of steps 25 and 26 are alternately repeated (step 27).
A group of vectors representing the mapping relationship between them is formed. Thereafter, mapping is executed based on the formed movement vector group (step 28).

In the third embodiment, as in the second embodiment, the potential is set to the destination image B.
Instead of forming the image, it is conceivable to form the original image A. In this case, in the flowchart of FIG. 8, after extracting a feature amount for each minute region of the image A in step 23A, a potential is formed for this feature amount,
Steps 17 and 18 shown in the flowchart of FIG.
19 to obtain a mapping result.

FIG. 9 is a flowchart showing a fourth embodiment of the present invention. The present embodiment is an embodiment for coping with a case where the non-linearity between the image A and the image B is small (that is, a case where the required movement amount of the minute area is sufficiently small), and is shown in FIG. 1 or FIG. The first of the present invention,
The second embodiment is characterized in that the latter half of the processing is replaced by a conventional method instead of the potential method.

First, similarly to the first and second embodiments of the present invention, in step 30, segmentation is performed on the mapping source image A and the mapping destination image B to obtain the segmented images A 'and B'. And these four images A,
B, A ′, and B ′ are input images.

Next, in the movement vector search step 40, the following operation is performed. First, (1) images A, A ′,
B and B ′ are divided into minute regions, and a feature amount is extracted for each minute region. A pixel value p is extracted from each minute area of the images A and B as a feature amount. From each minute area of the images A ′ and B ′, a discrete value s representing the segmentation result is obtained.

Next, (2) calculation (competition) of the movement vector is performed based on the obtained feature amount. Since the calculation (competition) of the movement vector has already been described with reference to FIG. 6, a duplicate description will not be given here.

When the movement vector is calculated for one micro area in this way, then (3) an operation for coordinating this vector with the vector of the adjacent micro area is performed. Cooperation is performed by a median operation (median process) or an averaging process.

The above operations (2) and (3) are repeated for all of the minute regions to obtain a group of movement vectors for each minute region of the image A. Of the image B is executed.

[0059]

As described in the description of the conventional example, in the conventional mapping method that relies only on pixel values,
Before reaching the original mapping, there were many cases in which local minima caused by noise or the like were captured and 100% movement was not performed (for example, Kuwatani, Sase, Kosugi et al. Smooth Deformation of Image ”, IEICE Transactions D-II, Vol.
6-D-II, no. 2, 296-303 (1993)
reference).

However, in the present invention, since the original image is segmented and then mapped based on the result, noise is sufficiently removed based on a priori information attached to the target image. There is less likelihood of being captured by local minima, and more complete movement of each micro-image is achieved, resulting in sufficient mapping capability.

Further, in the conventional mapping method, it is necessary to perform an exhaustive search over a wide area on the other image where a minute area of one image may correspond to the other image. Although it is necessary, the present invention has an advantage that the calculation time is drastically shortened because the direction of movement is obviously given as the steepest direction of the potential by introducing the potential.

[Brief description of the drawings]

FIG. 1 is a diagram showing a processing flow of a first embodiment of the present invention.

FIG. 2 is a diagram for describing a potential method.

FIG. 3 is a diagram for describing a potential method.

FIG. 4 is a diagram for describing a potential method.

FIG. 5 is a diagram showing a segmentation result.

FIG. 6 is a diagram provided for description of a segmentation method.

FIG. 7 is a diagram showing a flow of processing according to a second embodiment of the present invention.

FIG. 8 is a diagram showing a flow of processing according to a third embodiment of the present invention.

FIG. 9 is a diagram showing a flow of processing according to a fourth embodiment of the present invention.

 ──────────────────────────────────────────────────続 き Continuing on the front page (72) Inventor: Yu Hosoya 3-2-1 Sakado, Takatsu-ku, Kawasaki-shi, Kanagawa 211 KSP East Wing Inside Brain Research Institute, Inc. (72) Inventor: Kazuyuki Kudo Midori-ku, Yokohama-shi, Kanagawa 4259 Nagatsutacho Tokyo Institute of Technology (72) Yuri Nojima Inventor 4259 Nagatsutacho, Midori-ku, Yokohama-shi, Kanagawa Prefecture F-term (reference) 4C096 AA20 AB50 AC01 AD14 DC18 DC21 DC33 DC40 5B057 AA07 BA06 CA08 CA12 CA16 CB08 CB12 CB16 CC02 CD11 DA07 DA08 DB02 DB09 DC03 DC08 DC30 DC32 DC36 5L096 BA13 DA01 FA01 FA67 GA19 HA04

Claims (4)

[Claims] 1. An adaptive nonlinear mapping method for mapping an image A to an image B, wherein the image A and the image B are subjected to a segmentation step based on a configuration attribute of the two images; A dividing step of dividing the image A ′ and the image B ′ after these segmentation into minute regions; a feature amount extracting step of extracting a feature amount of the divided minute regions; and the extracted feature amount of the image B. Forming a potential every time; moving the minute area of the image A in the steepest potential direction of the image B; And the moving step and the coordinating step are alternately repeated so that the images A and B are copied. An adaptive nonlinear mapping method, comprising: obtaining a vector group representing an image relationship; and performing a mapping of an image A to an image B based on the obtained vector group. 2. An adaptive non-linear mapping method for mapping an image A to an image B, wherein the image A and the image B are subjected to a segmentation based on a configuration attribute of the two images; A dividing step of dividing the image A ′ and the image B ′ after the segmentation into minute regions; a feature amount extracting step of extracting feature amounts of the divided minute regions; and the extracted feature amount of the image A. Forming a potential every time; moving the minute area of the image B in the steepest potential direction of the image A; and moving the movement vector in the movement of the image B to an adjacent minute area in the image B. And the moving step and the coordinating step are alternately repeated so that the images A and B are copied. An adaptive nonlinear mapping method, comprising: obtaining a vector group representing an image relationship; and performing a mapping of an image A to an image B based on the obtained vector group. 3. An adaptive nonlinear mapping method for mapping an image A onto an image B, comprising: a dividing step of dividing the images A and B into minute regions; and a feature amount extracting a feature amount of the divided minute regions. An extracting step; a step of forming a potential for each of the extracted feature amounts of the image B; a moving step of moving the minute area of the image A in the steepest potential direction of the image B; A coordination step of coordinating a movement vector at the time of movement between adjacent minute areas in the image A; and a step of repeatedly repeating the movement step and the coordination step to obtain a vector group representing a mapping relationship between the image A and the image B; Performing the mapping of the image A to the image B based on the determined vector group. 4. An adaptive nonlinear mapping method for mapping an image A to an image B, wherein the image A and the image B are subjected to a segmentation step based on a configuration attribute of the two images; A dividing step of dividing the image A ′ and the image B ′ after these segmentation into minute regions, a feature amount extracting step of extracting a feature amount of each of the divided minute regions, and A movement vector calculation step of calculating a movement vector of the minute area; a coordination step of collaborating a movement vector obtained in the movement vector calculation step between adjacent minute areas in the image A; and the movement vector calculation step and the cooperation Obtaining a group of vectors representing the mapping relationship between the image A and the image B; Comprising the step of performing a mapping to the image B of the image A on the basis of the obtained vector group, an adaptive nonlinear mapping method.