JPH0921610A - Image-processing apparatus and image-processing method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an apparatus and a method for processing images whereby an image of an optional object is extracted from images and position of the object is obtained stably. SOLUTION: A local model reflecting a change due to photographic environment as a parameter is preliminarily formed for a local part of an object to be extracted and stored in a model storage part 12. Each partial image of an image input through an image input part 11 is subjected to matching with each local model at a matching part 13 with the user of the stored models. In accordance with result of the matching at the matching part 13 and from the viewpoint of how much each partial image agrees with the local model, a local data integration part 21 displays probabilities of a position of the object to be extracted in a parameter space including positional data of the image, and integrates the probabilities. An object position determination part 17 extracts a part of the highest probability in the parameter space, thereby determining and outputting the position of the object in the input image.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an image processing apparatus and an image processing method for extracting an arbitrary object from an input image and discriminating its position.

[0002]

2. Description of the Related Art Conventional techniques for extracting (detecting) or recognizing an object by integrating local information include a technique of sequentially combining hypotheses with the obtained local information and Hou.
Method based on voting from local information as represented by gh transformation (Toshikazu Wada, Takashi Matsuyama "New development of Zuken detection method based on Hough transformation", Journal of Information Processing Society, Vol.3)
6, No. 3, 1995) is known.

[0003]

In the method based on hypothesis verification, a large number of candidates for local information of an object obtained from an image are successively combined and verified. For this reason, an explosion of combinations (enlargement of the number of combinations) occurs and the amount of calculation becomes enormous. The voting method is effective when there is a large amount of information indicating the true value. However, when only a result indicated by a small amount of information can be a true value, a good result cannot always be obtained. In addition, a part of the object to be photographed is hidden, and the object changes variously in the image due to the difference in the photographing environment such as size, brightness, and viewing direction.
Therefore, it is difficult for the conventional object extracting apparatus and method to accurately extract the extraction target object from the captured image and accurately determine the position thereof.

The present invention has been made in view of the above situation, and provides an image processing apparatus and an image processing method capable of extracting an image of an arbitrary object from an arbitrary image and stably obtaining the position thereof. The purpose is to provide.

[0005]

In order to achieve the above object, an image processing apparatus of the present invention comprises a model storage means for storing models of a plurality of local parts of an extraction target, and an image input means for inputting an image. A matching unit that calculates the degree of matching (matching degree) between each model stored in the model storage unit and each part of the input image of the image input unit, and based on each matching degree calculated by the matching unit ,
Local information integration means for obtaining the probability distribution of the position of the extraction object, and output means for outputting the position with the highest probability in the probability distribution calculated by the local information integration means as the position of the extraction object It is characterized by

According to the above configuration, if there is an image of a local portion that shows a true value with a high probability even with a small number, a high probability distribution can be obtained at the accurate position of the position information of the extraction object based on that information. Therefore, even when a part of the extraction target object in the input image is hidden, the extraction target object and its position can be detected.

The model storage means stores a model of an image of each local portion reflecting a change due to the imaging environment as a parameter, and the matching means sequentially cuts out partial images of the same size as the local portion from the input image. , The degree of coincidence with the model of each local part is calculated, and the local information integration means uses the degree of coincidence calculated by the matching means to determine the position of the extraction target in the parameter space including the positional information of the image. The probability distribution of may be obtained.

According to this structure, since the model is created in consideration of the change due to the photographing environment, the position of the extraction object can be stably detected regardless of the change of the photographing environment.

For example, the local information integrating means uses the degree of coincidence calculated for each model by the matching means for each partial image to determine the position of the extraction target object in the input image for each partial image. A probability calculating means for obtaining a probability distribution; a first integrating means for integrating the probability distributions calculated by the probability calculating means for each local part; and a first integrating means for integrating the probability distributions integrated by the first integrating means. It may be configured by two integrating means.

Further, the model storage means stores the change in the image of each local portion for different rotation angles θ of the extraction target,
A first model represented as a curve in an eigenspace represented by a predetermined eigenvector and a second model represented by a curve representing the position of each local portion with respect to different rotation angles θ of the extraction target are stored, and When the matching means projects an image of each part of the image input by the image input means onto the eigenspace, the degree of coincidence with the first model is calculated from the distance between the projection point and the curve of the first model. Then, the integrating means sets the first probability distribution for the image of each part of the input image based on the degree of matching calculated for the first model by the matching means, and sets the corresponding second model to the corresponding second model. A second probability distribution may be set based on this, the first probability distribution and the second probability distribution may be multiplied, and the probability distributions of the partial images obtained by the multiplying means may be integrated.

The local information integration means obtains a probability distribution on the parameter space with the position and the rotation angle θ as parameters, and the output means outputs the position information and the rotation angle θ from which the highest probability is obtained. You may do it. With such a configuration, the rotation angle θ of the detection target can also be detected.

Further, the model storage means stores models of images of local portions of a plurality of extraction objects, the local information integration means obtains a probability distribution of the positions of the extraction objects, and the output means Alternatively, information indicating the extraction target object having the highest probability and its position may be output.
With such a configuration, not only the position information but also the extracted object or the like can be determined.

In order to achieve the above object, the image processing method of the present invention is performed by an image input step of inputting an image, a model of a plurality of local parts of an extraction target prepared in advance, and the image input step. A matching step for calculating the degree of coincidence with each part of the image, a probability calculating step for obtaining a probability distribution of the position of the extraction target object based on the degree of coincidence for each part of the input image, and a probability distribution calculated by the probability calculating step. And an output step of outputting the position with the highest probability in the probability distributions integrated by the integration step as the position of the extraction target.

Also in this method, if there is an image of a local portion that shows a true value with a high probability even with a small number, a high probability distribution can be obtained at the accurate position of the position information of the extraction object based on that information. Therefore, it is possible to accurately detect the extraction target and its position.

[0015]

DETAILED DESCRIPTION OF THE INVENTION An image processing (extracting) apparatus and an image processing (extracting) method according to an embodiment of the present invention will be described below with reference to FIGS. As shown in FIG. 1, the image processing apparatus of this embodiment includes an image input unit 11
The model storage unit 12, the matching processing unit 13, the probability calculation unit 14, the intra-local information integration unit 15, the inter-local information integration unit 16, and the object position determination unit 17.

The image input unit 11 includes a photographing device such as a CCD camera or an image reader, a converting unit for converting an image acquired by the photographing device into a digital image, and an image memory for temporarily storing the converted image.

The model storage unit 12 is composed of a memory and records a model of an object to be extracted (detected). The apparent image of a three-dimensional object changes depending on the direction, orientation, etc. of the object. In order to appropriately extract such a three-dimensional object, the model storage unit 12 uses the parametric eigenspace representation method of the characteristic portion (local portion) of the extraction target in the model generation image. Model that represents the change in the position of each local part of the
Is stored. The details of the pattern model and the location model will be described later.

The matching processing unit 13 is composed of a data reading circuit and an arithmetic unit, and includes an image (partial image) of each part of the input image input from the image input unit 11 and each pattern model stored in the model storage unit 12. The matching degree (matching degree) is calculated.

The probability calculator 14 is composed of an arithmetic circuit,
A probability distribution (probability field, probability density) of the position of the extraction target is calculated for each partial image of the input image based on the degree of coincidence obtained by the matching processing unit 13. The local information integration unit 15
Probability distributions obtained for each partial image are integrated for each local part. The local information integration unit 16 further integrates the probability distributions obtained for the respective local parts in the Bayesian framework. The probability calculation unit 14, the intra-local information integration unit 15, and the inter-local information integration unit 16 form a local information integration unit 21.

The object position determination unit 17 is composed of an output device or the like, and outputs the position showing the highest probability in the probability distribution integrated by the local information integration unit 21 as the position of the extraction object.

Next, a method of creating the pattern model and the location model stored in the model storage unit 12 will be specifically described. The image of a three-dimensional object changes depending on the direction, orientation, etc. of the object. For example, considering a chair as an extraction target, when the chair is rotated, various images (images for model creation) as shown in FIG. 2 are obtained.

Then, the camera is moved around the extraction object (or the camera is fixed and the extraction object is rotated on the spot), and as shown in FIG. The images for model creation are sequentially acquired while changing the.
Next, of the acquired model creation images, the extraction target,
That is, only the image of the extraction target is acquired by substituting the data “0” into the part other than the chair.

Next, an image of the characteristic portion (local portion) of the extraction target is cut out from each model-forming image in a predetermined size. Letting R be the number of angles θ (the number of images for model creation) and S be the number of characteristic portions, R × S images of local portions (local images) can be obtained. For example, assuming that the local portion of the chair shown in FIG. 2 is the three portions A1 to A3 shown in FIG. 3, the three local images A1 to A3 are extracted from each model creating image.

Then, using the rotation angle θ as a parameter, each local image is projected onto an eigenspace created by parametric eigenspace representation, and by connecting these, one pattern model for each local part is obtained as a curve.

Here, the number of θ (the number of images for model creation)
Where R is the R, and θ is different, and the set of each local image acquired is
Express as. Here, Xi represents the i-th local image of each local portion, and each is represented by Formula 2. In Expression 2, Xnm represents the gradation of the m-th pixel of the n-th local image. The set of equation 1 is formed for each local part,
In the example of FIG. 3, one set is formed in each of the local portions A1 to A3.

[0026]

## EQU1 ## {x ₁ , x ₂ , x ₃ , ..., x _R }

[Number 2] _{x 1 = {x 11, x} 12, x 13, ···, x 1N} x 2 = {x 21, x 22, x 23, ···, x 2N} ······ x _R = {x _R1 , x _R2 , x _R3 , ..., x _RN }

Then, the average c of the set of each local image is given by
And subtract the average image from each local image,
The matrix X shown in Expression 4 is generated.

[0028]

## EQU3 ## c = (1 / R) Σx _r r = 1 to R

X ≡ {x ₁ −c, x ₂ −c, x ₃ −c, ..., x _R −c}

The covariance matrix Q of this image set X is expressed by equation 5.

## EQU5 ## Q≡XX ^T T represents an orthogonal function.

The k-dimensional eigenspace solves the eigenequation of equation (6), and has k large eigenvalues λ _{1 to} λ _N (λ ₁ ≧ ... ≧ λ _k ≧ ...
Eigenvectors (e ₁ , e ₂ , ... E _k ) corresponding to ≧ λ _N
Is obtained as the basis vector.

[Equation 6] λ _i · e _i = Q · e _i

When each local image is projected onto the k-dimensional eigenspace, one local image corresponds to one point on the eigenspace. Furthermore, if each local image at each rotation angle θ is projected onto the eigenspace, it becomes a series of points. By connecting these points using the interpolation method, one pattern model is obtained as a curve for one local image.

FIG. 4 shows an example in which the local image A1 of the chair, which is the extraction target, is plotted in a four-dimensional (k = 4) eigenspace according to the change in θ to create a pattern model. This pattern model is similarly generated for the local images A2 and A3. The model storage unit 12 stores the pattern model for each local portion thus obtained.

Next, the location model will be described. As the rotation angle θ is changed, the position of the local image in each model creation image changes. As shown in FIG. 5, the location model is a model in which this positional change is represented by an (x, y) coordinate system with the upper left position of the model creating image as the origin.

Next, a procedure for extracting the position of the extraction object in the input image using the pattern model and the location model created in advance will be described with reference to FIGS. 6 to 9.

The image input unit 11 acquires an arbitrary image as shown in FIG. 6, converts it into a digital image, and stores the acquired image.

As shown in FIG. 7, the matching processing section 13 sequentially cuts out an image (partial image) having the same size as the local image while shifting the pixel by one pixel in the X and Y directions. The matching processing unit 13 projects each partial image in the eigenspace in the same manner as when creating the pattern model. Then, each cut out partial image is represented by one point in the eigenspace. Then, the distance D due to the change of the rotation angle θ between this point and the pattern model of each local image
_{Find Ai} (θ). This distance D _Ai (θ) is the local portion A
It is an index indicating the degree of coincidence in the pattern model of i (i = 1, 2, 3).

The probability calculating section 14 uses the degree of coincidence at each (X, Y) coordinate position obtained as a result of matching, with respect to an arbitrary rotation angle θ, the probability that the partial image at that position coincides with the local image. The P _pattern is calculated according to _equation 7.

[0038]

(Equation 7) Here, σ ₁ is an arbitrary constant.

For example, when the partial image at the (0,0) coordinate shown in FIG. 7 is projected onto the eigenspace, focusing on the pattern model of the local portion A1, the distance D _{A1 is obtained} for this pattern.
(Θ) is obtained, and the probability P _pattern for the local portion A1 is calculated according to Equation 7 using this distance D _A1 (θ). This probability P _pattern is a function of θ.

Next, the probability calculating section 14 sets new coordinates (α, β) for each partial image with respect to the arbitrary rotation angle θ by the formula (8).

(Α, β) = (X, Y) − (x _Ai (θ), y _Ai (θ))
(I = 1,2,3)

As shown in FIG. 8, (X, Y) are the coordinates of the partial image in the input image, and (x _Ai (θ), y
_Ai (θ)) is the position of the local image in the model creation image described in the location model at arbitrary θ, and (α, β) is the coordinates of the reference point of the partial image in the input image. .

If the matching is performed correctly, the places having a high degree of coincidence with respect to each local portion show the same coordinates (α, β). Next, according to equation 9, the probability P that the new coordinate (α, β) is the reference position of the extraction target
_Ask for _location .

[0043]

[Equation 9] Where σ ₂ is an arbitrary constant.

This probability P _location is such that when the partial image at the position of coordinates (X, Y) exactly matches any of the local images, the reference point of the extraction target is the coordinates (α,
β) and represents the probability that the angle of the extraction target is θ.

Then, the probability calculating section 14 multiplies the corresponding P _pattern and P _location by P _pattern.
-P _location is calculated and parameter space (α, β, θ)
Generate a probability distribution at. That is, based on the partial image at the (X, Y) position, the distribution of the probabilities that the reference point of the extraction target object is located at the coordinates (α, β) and the rotation angle is θ is calculated. This probability distribution is, for example, a distribution schematically shown in FIG. In FIG. 9, reference numeral P1
Each of P4 indicates a probability.

The probability calculator 14 calculates a probability distribution for partial images at all positions of the input image. Therefore, if the number of partial images in the input image is m, m sets of probability distributions are obtained for each local portion (A1 to A3).

The local information integration unit 15 takes the sum of all the probability distributions obtained for the same local part and newly generates a probability distribution in the parameter space. As a result, one probability distribution is obtained in the parameter space for each local part. In the above chair example, probability distributions are obtained for the local parts A1 to A3, respectively.

The inter-local-information integrating section 16 integrates the probability distributions of the local parts obtained by the intra-local-information integrating section 15 by the Bayesian framework. The probability density function p _k (α, β, θ) = f
^{If k} (α, β, θ) (where k = 1, 2, ..., N is a number indicating a local part), the probability distribution after integration is expressed by
It is represented by 0.

[0049]

(Equation 10)

From Equation 10, the probability distribution p _n (α, β, θ) finally obtained by integrating the sequential probabilities becomes a probability distribution integrating all information. Note that p ₀ is an initial prior probability, and in the case of no knowledge, an equal probability is used.

Finally, the object position determining unit 17 determines the place (α _high , β _high , θ _high ) showing the highest probability in the finally obtained probability distribution of the parameter space,
Output this. (Α, β) corresponds to the origin of the coordinate system (x, y) of the model creating image. Therefore, the position of the coordinates (α _high , β _high ) of the input image becomes the reference point of the extraction target. Further, θ _high represents the rotation angle of the extraction target object extracted from the input image.

As described above, in this embodiment, a local model is created in advance for each local portion of the extraction target object, in which the change due to the imaging environment is reflected as a parameter. Then, for each partial image of the input image, matching with each local model is performed, and from the viewpoint of how much each partial image matches the local model, the position object in the parameter space including the position information of the image is detected. Probabilistically display position. Finally, we integrate these probabilities,
The position and rotation angle of the extraction target object in the input image are extracted by extracting the part with the highest probability in the parameter space.

By adopting such a method, a good result can be obtained as long as there is information showing a true value with a high probability even with a small number, without successively verifying the combination of local information. Further, the position of the extraction target can be stably extracted in consideration of the change due to the photographing environment.

Therefore, for example, as shown in FIG. 10, since a part of the object to be extracted is hidden by another object, even if the information of the local portion A2 is lost, the images of the local portions A1 and A3 are displayed. Based on this, the position and rotation angle of the extraction target object can be obtained with high probability. Therefore, the position of the extraction target can be stably extracted without being affected by the loss of information or the like.

In the above embodiment, in order to increase the calculation speed, an example in which the input image is sequentially calculated by each processing unit in a pipeline is shown. However, the image processing having the configuration shown in FIG. 11 is performed. The processing may be sequentially performed by the device.

Further, in the above embodiment, the degree of coincidence D _Ai (θ) of the pattern model of each local portion is obtained, and the probability distribution is obtained for each local portion. However, for example, for each partial image, a pattern model having the highest degree of coincidence may be obtained and subsequent calculations may be performed using this model. For example, when a certain partial image is projected on the eigenspace and the pattern model that is most approximated (the minimum value of the distance D (θ) is obtained) is the pattern model of the local portion A1, D (θ) of pattern model
May be used to perform the subsequent probability calculation. By adopting such a method, the time required for the subsequent calculation of the probability distribution can be shortened, which is effective when the number of local parts is large.

As shown in the figure, this image processing apparatus is C
PU 31, RAM 32, ROM 33, auxiliary storage device 3
4, an imaging device 35, an input / output device 36, and a bus 37 that connects these to each other.

The CPU 31 has a RAM 32 and a ROM 33.
It operates in accordance with the control program stored in and executes arithmetic processing and control processing. The RAM 32 includes a main memory and an extended memory, and stores data to be processed, control programs, and the like. The ROM 33 stores fixed data, a fixed control program, and the like. The auxiliary storage device 34 is composed of a hard disk device, a floppy disk device, a CD-ROM device, etc., stores the operation program, model, etc. of the CPU 31, and transfers these data to the RAM 32 as necessary. The imaging device 35 acquires the image data to be processed and transfers it to the RAM 32. I / O device 36
Is the position of the extracted object while inputting the control information,
Outputs the rotation angle etc.

Next, the operation of the image processing apparatus of FIG. 11 will be described with reference to FIG.
This will be described with reference to the flowchart of FIG. First, the pattern model, the location model and the like are stored in the auxiliary storage device 34 and transferred to the RAM 32 as needed. Further, it is assumed that the input image (image data to be processed) is acquired by the imaging device 35 and stored in the RAM 32.

In this case, the CPU 31 cuts out a partial image from the input image (step S1), performs matching with each pattern model prepared in advance, and obtains the pattern model showing the highest degree of matching (step S2). Then, it is determined whether or not the matching process is completed for all the partial images (step S3), and the matching process is repeated until the completion.

When it is determined in step S3 that the matching process has been completed for all partial images, the probability P _pattern , P _location, and their product P _{p attern} · P _location are obtained for each partial image ( Step S4
S5, S6). Then, it is determined whether or not the probability calculation process has been completed for all partial images (step S7),
The probability calculation process is repeated until the end.

When it is determined in step S7 that the probability calculation process has been completed for all partial images, the probabilities P _pattern and P _location are integrated for each local part, and the probability distribution in the parameter space for each local part is integrated. Is calculated (step S8). Then, the probability distributions of each local part are integrated,
A final probability distribution is obtained (step S9). The coordinates of the position where the probability in the final probability distribution shows the maximum value (α
_high , β _high ) and the rotation angle θ _high are output via the input / output device 36 (step S10).

With such a configuration, the processing program and the model (or the program for creating these models, etc.) can be stored on the floppy disk or the CD-R.
It is stored in the OM or the like, and if necessary, the auxiliary storage device 3
By mounting the image pickup device 4 on a general-purpose computer, the general-purpose computer can be made to function as an image processing device.

In the above embodiment, for the sake of easy understanding, three chairs and three local parts are shown as extraction objects. However, the extraction objects, the size of the input image, the size and number of local parts, etc. Is not limited in any way.

The equation for calculating the probability distribution is not limited to the equations 8-10. The probability distribution may be calculated based on a calculation formula other than the formulas 8 to 10.

Further, in the above embodiment, the rotation angle θ is adopted as one of the parameters for controlling the change of the image of the extraction object, but the size, brightness,
Color change or the like may be adopted as a parameter. In this case, a pattern model and a location model based on the introduced parameters are similarly created and used for matching and probability calculation.

The extraction target does not have to be one kind. Models of local images of a plurality of extraction objects are stored in the model storage unit 12 or the RAM 32, and these are compared with an input image to obtain a probability distribution for each extraction object, and the extraction object with the highest probability is obtained. The object and its position may be output.

[0068]

As described above, according to the image processing apparatus and the image processing method of the present invention, it is possible to accurately extract an extraction target object from an arbitrary image and determine its position.

[Brief description of drawings]

FIG. 1 is a circuit block diagram showing a configuration of an image processing apparatus according to an embodiment of the present invention.

FIG. 2 is a diagram illustrating how the appearance of an image changes according to a rotation angle θ.

FIG. 3 is a diagram showing an extraction target object and a local portion.

FIG. 4 is a diagram illustrating a method of creating a pattern model.

FIG. 5 is a diagram illustrating a method of creating a location model.

FIG. 6 is a diagram showing an example of an input image.

FIG. 7 is a diagram illustrating a process of cutting out a partial image from an input image.

FIG. 8 is a diagram showing a relationship between a coordinate system of an input image and a coordinate system of a model creating image.

FIG. 9 is a diagram schematically showing an example of a probability distribution on a parameter space.

FIG. 10 is a diagram showing an example in which a part of the extraction target object included in the input image is hidden by another object.

FIG. 11 is a circuit block diagram showing another configuration example of the image processing apparatus.

12 is a flow chart for explaining the operation of the image processing apparatus shown in FIG.

[Explanation of symbols]

11 ... Image input unit, 12 ... Model storage unit, 13 ... Matching processing unit, 14 ... Probability calculation unit, 15 ... Local information integration unit, 16 ... Local information integration Part, 17 ... Object position determination unit, 21 ... Local information integration unit, 31 ... CPU, 32
... RAM, 33 ... ROM, 34 ... auxiliary storage device, 3
5 ... Imaging device, 36 ... Input / output device, 37 ... Bus

Claims (7)

[Claims] 1. Model storage means for storing models of a plurality of local parts of an extraction target, image input means for inputting an image, each model stored in the model storage means, and input of the image input means. Matching means for calculating the degree of coincidence with each part of the image, local information integrating means for obtaining the probability distribution of the position of the extraction object based on the degree of coincidence calculated by the matching means, and the local information integrating means An image processing device, comprising: an output unit that outputs a position having the highest probability in the calculated probability distribution as a position of the extraction target. 2. The model storage means stores a model of an image of each local portion in which a change due to an imaging environment is reflected as a parameter, and the matching means extracts a partial image of the same size as the local portion from the input image. Sequentially cutting out, obtaining the degree of coincidence with the model of each local portion, the local information integrating means, using the degree of coincidence obtained by the matching means, in the parameter space including the position information of the image, the extraction object The image processing device according to claim 1, wherein a probability distribution of positions is obtained. 3. The local information integration means uses, for each partial image, the degree of matching calculated for each model by the matching means, for each partial image.
Probability calculating means for obtaining a probability distribution of positions of the extraction object in the input image; first integrating means for integrating the probability distributions calculated by the probability calculating means for each local part; The image processing apparatus according to claim 2, further comprising a second integration unit that integrates the probability distributions integrated by the integration unit. 4. The model storing means extracts a first model that represents a change in an image of each local portion with respect to different rotation angles θ of an extraction target as a curve in an eigenspace represented by a predetermined eigenvector, and an extraction. A second model that represents the position of each local portion with respect to different rotation angles θ of the object with a curve, and the matching unit stores the image of each part of the image input by the image input unit as the eigenspace. When projected onto
The degree of coincidence with the first model is calculated from the distance between the projection point and the curve of the first model, and the integration unit calculates the first model by the matching unit for each image of the input image. The first probability distribution is set on the basis of the degree of coincidence, the second probability distribution is set on the basis of the corresponding second model, the first and second probability distributions are multiplied, and the result is obtained by multiplication. The image processing apparatus according to claim 1, 2 or 3, wherein the probability distributions for each of the partial images are integrated. 5. The position and rotation angle θ are set by the local information integration means.
The image processing apparatus according to claim 4, wherein a probability distribution in a parameter space having a parameter is obtained, and the output unit outputs the position information and the rotation angle θ at which the highest probability is obtained. 6. The model storage means stores models of images of local portions of a plurality of extraction objects, the local information integration means obtains a probability distribution of positions of the extraction objects, and the output means The image processing apparatus according to any one of claims 1 to 5, wherein the extraction target object having the highest probability and the information indicating the position thereof are output. 7. An image input step of inputting an image, a matching step of calculating a degree of coincidence between a model of a plurality of local parts of an extraction target prepared in advance and each part of the image input by the image input step. , A probability calculating step of obtaining a probability distribution of the positions of the extraction objects based on the degree of coincidence calculated in the matching step, an integrating step of integrating the probability distributions calculated by the probability calculating step, and an integrating step of the integrating step. An output step of outputting the position with the highest probability in the generated probability distribution as the position of the extraction target object.