JPH0822544A - Picture recognition method - Google Patents

Abstract

PURPOSE:To obtain the accurate position of a structure model. CONSTITUTION:The operation is started from a given initial position in a step S301, and a position of energy minimizing is obtained by search of the vicinity of one point in the energy minimizing processing for search of the vicinity of one point in a step S302. The model is moved so as to reduce the energy in a step S303 if it is simultaneously moved to vicinities of two points. In the convergence discrimination of a step S304, the operation is returned to the step S302 to try to minimize the energy if the combination to reduce the energy is found in the step S303; but convergence is discriminated and the position of the model at this time is defined as the position of energy minimizing in a step S305 if it is not found in the step S303.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an image recognition method for recognizing a structure such as a contour of an object in an input image.

[0002]

2. Description of the Related Art Conventionally, techniques in such a field include:
For example, some documents were described in the following documents. Literature; First International Conference on Computer Vision (First Internat
ional Conference on Computer Vision) (1987) I
EEE (US) Michael Kass, Andrew Witkin, and Demetri
Terzopoulos "Snakes: Active contour models" P. Two
59-268 Conventionally, as an image recognition method, for example, a contour extraction method includes a method of extracting a contour of an input image by an active contour model as described in the above-mentioned document. The dynamic contour model is a contour extraction method that obtains a closed curve satisfying a constraint condition attracted to image features such as lines and edges and a constraint condition regarding a shape by energy minimization. Hereinafter, a conventional contour extraction method using an active contour model will be described.

FIG. 2 is a diagram showing the processing contents of a conventional contour extraction method using an active contour model. FIG. 3 is a conceptual diagram of an active contour model showing how a closed curve is attracted to image features as energy converges to a minimum value in FIG. In the following description, "v" is added before the symbol to represent the vector symbol. It is shown in bold type in the figure. As shown in FIG. 3, from the original image data 10 in FIG. 2, the position of the closed curve (hereinafter, referred to as snake) that is a model of the contour line to be extracted is determined by the coordinate system x,
In y, n nodes vv _i = (x _i , y _i ), i =
It is represented by 1, ..., N. However, since it is a closed curve, vv _{i + n} =
vv _i . Energizing function E ^* _{sn ake} definitions 40, such as: (1) the model snake of the contour.

[Equation 1] E _{int in the} equation (1) is an internal energy that represents a constraint condition regarding the shape of the snake, and can be expressed as the following equation (2).

[Equation 2] Each term in the equation (2) is a term related to the continuity of the first and second orders, and the first term is energy that becomes smaller as the snake becomes shorter and the second term becomes smaller as the snake becomes smoother (however,
If α _i > 0, β _i > 0). E _ext of the equation (1) is sn
ake is an external energy that represents a constraint condition that is attracted to the image feature, and is defined by the following equation (3), for example. E _ext (i) = w _edge (− | ▽ I (x _i , y _i ) | ² ) (3) where w _edge ; positive coefficient I (x _i , y _i ); image density external If the energy E _ext is defined as in equation (3),
The closer the snake is to the edge of the image, the smaller E _ext .

As described above, the energy function that minimizes the energy when the constraint condition regarding the shape and the constraint condition regarding the image feature are satisfied is defined, and the position of the snake that minimizes this energy function is determined. The active contour model is used to extract the contours. For example, the constraint condition that the shape of the snake is “smooth” as expressed by the expression (2) and the image that the snake as expressed by the expression (3) is “on the edge of the image”. When the feature constraint condition is given, a smooth contour line that matches the contour of the object in the image is extracted. The position of the _snake that minimizes the energy function E ^* _snake of the equation (1) is obtained by solving the equation of the following equation (4) by the variation method in the energy minimization process 20 of FIG.

(Equation 3) The equation (4) can be expressed in matrix form as the following equations (5) and (6). In the following description, “V” is added before the code to represent the matrix symbol. It is shown in bold type in the figure.

VA · VX + VF _x (VX, VY) = 0 (5) VA · VY + VF _y (VX, VY) = 0 (6) where VX = (x ₁ ... x _n ). ^T VY = (y ₁ ... Y _n ) ^T VF _x = (f _x (1) ・ ・ ・ f _x (n)) ^T VF _y = (f _y (1) ・ ・ ・_fy (n)) ^T VA; n × n coefficient matrix determined only by α _i and β _i The equations (5) and (6) are the initial position data 5 of FIG.
The initial positions VX ₀ and VY ₀ given as ₀ can be solved by iterative calculation of the following equations (7) and (8), which are output as the contour extraction result 30. VX _t = (VA + γ · VI) ⁻¹ · (γ · VX _t−1 −VF _x (VX _t−1 , VY _t−1 )) (7) VY _t = (VA + γ · VI) ⁻¹ · (Γ · VY _t-1 −VF _y (VX _t-1 , VY _t-1 )) (8) where VI: n × n identity matrix t; number of iterations γ; update step size FIG. 4 shows a flowchart of the energy minimization processing 20 of FIG. 2 in the contour extraction method using the conventional active contour model as described above. In the energy minimization process 20, first, in the initial position process 21, the initial values VX ₀ and VY ₀ of the snake are given from the initial position data 50 of FIG. 2, and the process proceeds to the external energy gradient calculation process 22. In the external energy gradient calculation process 22, the external energy gradients VF _x and VF _y at all nodes on the snake are calculated from the input original image 10 based on the definition 40 of the energy function in FIG. , And proceeds to coordinate value update processing 23. Coordinate value update process 2
3 uses the formulas (7) and (8) to update the updated snake.
The positions VX _t and VY _t are calculated and the process proceeds to the convergence determination process 24. In the convergence determination processing 24, the updated amount | V of each node
_{X t -VX t-1 |,} | VY t -VY t-1 | if the long close enough 0, it is determined that the calculation has converged, V at that time
X _t, and energy minimization value processing 25 VY _t, and outputs the contour extraction result 30 in FIG. On the other hand, if the convergence determination processing 24 does not determine that the convergence has occurred, the process returns to the external energy gradient calculation processing 22 and the above processing is repeated. By the above processing, the position of the snake that minimizes the energy function E ^* _snake of the equation (1) is obtained. The snake at this time is "smooth and is on the edge of the image."
As described above, the contour line satisfies the constraint condition of the shape and the constraint condition of the image feature, and thus the contour line matches the contour of the object in the image as shown in FIG.

[0006]

However, the conventional contour extraction method has the following problems and it is difficult to solve them. (A) In the conventional contour extraction method, only the contour line is the recognition target, and the model expressing the recognition target is limited to the closed curve. Therefore, it cannot be applied to the recognition of objects with more complicated structures that cannot be represented by closed curves alone. (B) When applied to recognition from a plurality of images such as a compound-eye image or a moving image, another mechanism for integrating the results of the conventional contour extraction method for each image is necessary. It was (C) In order to solve the above problems (a) and (b), more complicated processing is required, but it is difficult to process this efficiently, and sufficient processing speed is required. I can't get it. In order to solve the above problems (a) to (c), the present applicants have previously described in Japanese Patent Application No. 5-24604 (hereinafter, referred to as the above-mentioned proposal) not only a closed curve but also a more complicated one. We proposed an image recognition method that can process input images with various structures, recognizes multiple images uniformly, and can efficiently process such an extended recognition method. . That is, in the image recognition method for recognizing the input image by energy minimization, the model of the structure to be recognized is defined as the structure model represented by the arbitrary number of points on the arbitrary number of images, and the structural model is the image. The input image is obtained by defining a constraint condition attracted to the feature and an energy function that minimizes energy when the constraint condition regarding the structure of the structural model is satisfied, and determining the position of the structural model that minimizes this energy function. Any structure is extracted from. However,
In the previous proposal, when the position of the structural model that minimizes the energy function is moved to a point close to one point, only the search processing of the vicinity of one point that reduces the energy is performed, and therefore the accurate structural model There was a problem that the position was not requested.

[0007]

In order to solve the above-mentioned problems, the first invention defines a model expressing the structure of a recognition target in an input image as a model represented by arbitrary points on the image. Then, the constraint condition that the model is attracted to the image feature and the energy function that minimizes the energy when the constraint condition regarding the structure of the model are satisfied are defined, and the position of the model that minimizes the energy function is defined. An image recognition method for extracting an arbitrary structure from the input image by obtaining the image is configured as follows.
That is, when the position of the model that minimizes the energy function is moved to a point near one point on the model, the search process for the neighborhood of one point that moves such that the energy decreases and the position on the model It is obtained by performing simultaneous search processing to the neighboring points of a plurality of points that move so that the energy decreases when the plurality of points move to their neighboring points at the same time. The second invention is the same as the first invention,
First, energy minimization processing is performed by search processing in the vicinity of the one point, and then movement is performed so that energy is reduced when moving to two neighboring points on the model at the same time. The energy is minimized by returning to the search processing in the vicinity of the point and repeating the processing. A third aspect of the present invention represents the amount of change in energy as an evaluation function of the amount of movement to the vicinity of a plurality of points to be searched in the simultaneous search processing of the plurality of points into the vicinity of the point, the evaluation function Based on the above, the movement is such that the energy decreases. The fourth invention is the third invention.
In the simultaneous search process of a plurality of points to neighboring points of the invention of
Of the amount of change in energy expressed as a function of the amount of movement to the vicinity of multiple points to be searched, terms that depend only on the amount of movement of one point at each point are calculated and stored in advance before starting the search. In addition, the term that depends only on the stored amount of movement of one point is used in the simultaneous search process for the neighboring points of the plurality of points.

[0008]

According to the first aspect of the invention, since the image recognition method is configured as described above, the energy is reduced when the position of the model that minimizes the energy function is moved to one neighboring point. It is obtained by a search process in the vicinity of one point. At the position of the model obtained by the search processing of the vicinity of this one point, the case where the energy decreases when a plurality of points are simultaneously moved to their respective neighboring points has not been searched, so the accurate position of the model is obtained. Has not been done. Therefore,
An accurate model position is also obtained by moving so that the energy decreases when multiple points are moved to their neighboring points at the same time. According to the fourth invention, in the simultaneous search process of a plurality of points in the vicinity, the amount of change in energy is expressed as a function of the amount of movement of the plurality of points to be searched in the vicinity, and the movement of one point of each point. The term that depends only on the quantity is calculated in advance before starting the search, and is used in the simultaneous search processing. For example, when the number of neighboring points to be searched is m and the simultaneous search processing is performed for two neighboring points, m
It requires xm times of search processing, but m values of terms that depend only on the amount of movement of one point of each point are calculated in advance before starting the search, and are calculated in the simultaneous search processing. Since it is used, the processing efficiency is improved. Therefore, the above problem can be solved.

[0009]

EXAMPLE In this example, first, an outline (I) of the example of the previously proposed image recognition method and its problem (II) will be described, and then the present invention for solving the problem will be described. An example (III) of the image recognition method will be described. (I) Example of Image Recognition Method of Previous Proposal In the previous proposal, a model expressing a recognition target is not limited to a closed curve as in the conventional contour extraction method using an active contour model, and a more complex target is obtained. Consider a dynamic structure model that can express the structure of (hereinafter referred to as a dynamic model). The dynamic model expresses the structure such as the outline of the recognition target in the input image by a model of any number of points on the image, and the constraint condition that each point is attracted to the image feature of the input image and the positional relationship of each point. This is an image recognition method that finds the position of a model that satisfies the constraint condition regarding by energy minimization.
In the dynamic model, the structure such as the outline of the recognition target is represented by a model represented by n points on the image (hereinafter referred to as model). These n points are converted into the image coordinate system x, y.
Where vv _i = (v _1i , v _2i ) = (x _i , y _i ), i =
1, ..., n. The structure is extracted by finding the constraint condition that this model is attracted to the image feature and the position of the model that minimizes the energy when the constraint condition concerning the structure of the model is satisfied.

FIGS. 5A to 5C are diagrams showing examples of the dynamic model proposed above. For example, as shown in FIG. 5C, there are arbitrary points, and a constraint condition that these points are drawn to the image feature and a constraint condition regarding the structure as shown in the figure are given between these points. Consider a model that has Here, when an input image as shown in FIG. 5A is given, a structure that matches the features of the input image as shown in FIG. 5B can be extracted. Energy function E for this model
_{The model} is defined as the following expression (9). E _model = E _int + E _ext = E _int + E _image + E _con (9) E _int in the equation (9) is an internal energy function expressing the constraint condition regarding the structure of the model, and is represented by the following equation (10). To be defined as

[Equation 4] However, A _{λi μj} , B _λi , and C are coefficients, and
_{Let λi μj} = A _{μj λi} . The subscripts i and j of Σ are 1, ..., N, and λ and μ are 1 and 2. Similarly, in the following description, the range of the subscript of Σ is omitted,
The subscripts of lowercase letters such as i and j take the values 1, ..., N, and the Greek lowercase letters take the values of 1 and 2. For example,

(Equation 5) Is used in the sense of.

The internal energy function E as shown in equation (10)
_{Using int,} we can express the following energy terms.

(Equation 6) This acts as a term that draws the point vv _i to the fixed point vs _i , as shown in FIG.

(Equation 7) As shown in FIG. 7, this acts as a term that brings the vector vv _j −vv _i from the point vv _i toward the point vv _j closer to the constant vector vt _ij . When vt _ij = 0, the first term on the right-hand side of the conventional equation (2) α _i · | vv _i −vv _i−1 | ² /
As in 2, attract vv _i and vv _j . However, in the conventional contour model, such an energy term can be set only between two adjacent nodes.

(Equation 8) This is, as shown in FIG. 8, the amount of change between the vector from the point vv _i to the point vv _j and the vector from the point vv _j to the point vv _k , that is, the bending degree vv of the polygonal line vv _i −vv _j −vv _{k. It} works as a term that brings _i −2vv _j + vv _k close to the constant vector vu _ijk .

[0012]

[Equation 9] As shown in FIG. 9, this is the three points vv _i , vv _j , vv.
_It works as a term that increases the area of the triangle surrounded by _k . By combining these terms, as shown in FIG.
m consisting of arbitrary m points out of n points on model
It is possible to express the constraint condition regarding the area enclosed by the polygon. In addition to the above energy terms, 2n variables V _λi (i
= 1, ..., N, λ = 1, 2), an arbitrary energy term represented by a polynomial of second order or less can be expressed by the equation (10). E _ext = E _image + E _con in the equation (9) is mode
is an external energy function for l. E _con is an energy function for external control. E
_image is a function that represents a constraint condition in which model is attracted to the image feature, and is defined by the following equation (11).

[Equation 10] For example, if E _i (v _1i , v _2i ) is defined by the following expression (12), the point vv _i is attracted to the image edge. E _i (v _1i , v _2i ) = w _edge (− | ▽ I (v _1i , v _2i ) | ² ) (12) where w _edge is a positive coefficient, I (v _1i , v _2i ). Is the density of the image. By obtaining the position of the model so that the energy function E _{model of the} equation (9) defined as above is minimized, it is possible to extract the structure of the recognition target that satisfies the constraint condition regarding the image feature and the constraint condition regarding the structure of the model. .

FIG. 11 is a flowchart of the energy function minimization process of the previous application. In the energy minimization process of the previous application shown in FIG. 11, each point vv _i on the model
Evaluating the change of E _{model in} the equation (9) in the vicinity of
_The energy function is minimized by repeating the movement to the neighboring points where the _model decreases. That is, the sweep S
Starting from the initial position given in 141, step S
The processing from steps S142 to S145 is repeated until the model converges in 146. Steps S142, S
In 144, and S145, the counter i is incremented from 1 to the number n of the model, and in step S143, the neighborhood search process of the point vv _i described below is performed for each i. That is, the neighborhood search process is performed for all points on the model. In the convergence determination in step S146, it is determined that convergence has occurred when there are no neighboring points that reduce the energy function in the vicinity of all points, and the model at this time is determined.
Is set as the energy minimization position in step 147. Hereinafter, a method of evaluating a change in E _model of the expression (9) in the vicinity of vv _i in the neighborhood search processing of step S143 in FIG. 11 will be described.

In the equation (9), the point vv _i on the model is changed by δvv _i = (δv _1i , δv _2i ), that is, when vv _i is moved to vv _i + δvv _i variation [delta] _i E _model of E _model is expressed by the following equation (13).

[Equation 11] This change amount δ _i E _model becomes negative, that is, δ _i E _model
The movement δvv _i such that _model <0 reduces the energy evaluation function E _model as expressed by the equation (9). In order to evaluate such a change in the energy function E _model of the equation (9) in the vicinity of the point vv _i , the following equation (14)
The neighborhood energy evaluation function e _i (δvv _i ) such as

(Equation 12) However, D _λi is the following equation (15), which is δvv
Since it does not depend on _i , it can be calculated in advance before starting the search for the neighborhood.

[0015]

(Equation 13) Further, E _i (vv _i ) is calculated from the input image and the coordinate value of the point vv _i , but if the input image is determined, this is vv _i.
It is a function of two variables (v _1i , v _2i ) related to the coordinate value of.
Therefore, for all pixels of the input image, E _i (x, y)
Of the energy function image (the pixel value is E _i
It can also be obtained in advance as an image such as (x, y). However, using the above method, E
_{Since i} (vv _i ) is an individual function for each vv _i (i = 1, ..., N), it is necessary to obtain n energy function images in advance, which requires a large storage capacity. To do. Therefore, the number of energy function images can be reduced by doing the following. That is, all E _i (x, y), i = 1, ..., N are n _f type functions f _j (x, y), j = 1, ..., n _f (however, n
_{If f} <n) is represented by the following equation (16), by calculating n _f energy function images in advance, E _i (v
The calculation of v _i ) can be processed by a simple product-sum calculation.

[0016]

[Equation 14] The δ _i E _model of the equation (13) is represented by the following equation (17) using the neighborhood energy evaluation function e _i (δvv _i ) of the equation (14). δ _i E _model = e _i (δvv _i ) −E _i (vv _i ) −E _con = e _i (δvv _i ) −e _i (0) (17) Therefore, e _i (δvv _i ) <e If there is a movement δvv _i to a neighboring point such that _i (0), then vv _i becomes vv _i
By moving to _i + δvv _i , E in Eq. (9)
_model decreases. By repeating such movement for all points on the model, the energy function E _{model of the} equation (9) can be minimized.

FIG. 12 is a flowchart of the neighborhood search processing of step S143 in FIG. FIG. 13 is a conceptual diagram showing a state of neighborhood search. As shown in FIG.
First, the neighborhood energy evaluation function e _i (δv of Eq. (14)
In order to efficiently calculate v _i ), step S201
Then, D _λi in the equation (15) is calculated in advance. In step S202, the neighborhood energy evaluation function e _i (δv
v _i and the minimum value E _{mi n} of) initializes the movement Derutavv _i to neighborhood points at the current position that gives the minimum value E _min. In steps S203, S207, S208, the counter j is incremented from 1 to the neighborhood score m, and the point vv _{i is obtained} for each j.
S20 regarding the j-th neighbor point vv _i + δvv _ij of
4, S205 and S206 are performed. For example, the number of neighbors is m
FIG. 13 shows how the j = 5th neighbor point is searched when = 8. In step S204, the neighborhood energy evaluation value by the movement δvv _ij to the j-th neighbor point is evaluated by the neighborhood energy evaluation function e _i (δvv _ij ) of the equation (14) using D _λi obtained in step S201. . In steps S205 and S206, the minimum value E of these evaluation values
movement Derutavv _i of the _min, the neighboring point which gives the minimum value E _{mi n}
Ask for. In steps S209 and S210, if the point giving the minimum value of the neighborhood energy evaluation function is not the current position, the point vv _i is moved to vv _i + δvv _i .

As described above, by the neighborhood search processing as shown in FIG. 12, the point vv _i can be moved to a point among the neighboring points where the energy function E _{model of the} equation (9) is reduced. . As shown in FIG. 11, this neighborhood search process is repeated for all points on the model, and the process ends when all points have stopped moving.
The model at this time is the energy function E _{model of} equation (9).
Since is the minimum, the structure of the recognition target that satisfies the constraint condition given as the energy function can be extracted. FIG. 14 is a block diagram of an image recognition method based on the dynamic model proposed above. As shown in FIG. 14, an energy function image 130 is calculated from one or a plurality of input image data 110 by an energy function image calculation process 120 based on a model definition 160. In the energy minimization process 140, based on the model definition 160 from the input image data 110, the structure extraction result 1 from the initial position data 170 is processed by the processes of FIGS.
Get 50.

(II) Problems of Previous Proposal FIG. 15 is a diagram showing problems in the dynamic model of the previous proposal. Even if the model in the state of FIG. 15A moves as shown in FIG. 15B during the neighborhood search of the node vv _i , it moves as shown in FIG. 15C during the neighborhood search of the node vv _j . However, even if the energy increases, FIG.
As shown in (d), the state where the nodes vv _i and vv _j are moved at the same time may have lower energy than the state shown in FIG. That is, in the method proposed above,
Since the state cannot be moved from the state of FIG. 15 (a), the energy minimization process is terminated assuming that this state has the lowest energy, but in reality, the state of smaller energy as shown in FIG. 15 (d). May exist,
The energy minimization process was not performed correctly.

(III) Embodiment of the Present Invention FIG. 1 is a flow chart of energy minimization processing of an image recognition method showing an embodiment of the present invention. In the energy minimization process of this embodiment, the energy is minimized when the energy minimization process is performed by the one point neighborhood search as in the previous application and then the two points are simultaneously moved. The model is moved by such simultaneous neighborhood search. FIG. 16 is a flowchart showing step S303 of moving a model by a neighborhood search of two points in FIG. The method of energy minimization processing of the image recognition method of the present embodiment will be described with reference to these drawings. As shown in FIG. 1, first, from the initial position given in step S301, in the energy minimization process by the neighborhood search of one point in step S302, the flowcharts as shown in FIGS. Then, the energy minimization position is obtained by the neighborhood search of 1 point.

Next, in step S303, the model is moved by a two-point neighborhood search as shown in FIG. This is because if there are combinations that reduce energy when moving two points at the same time for all combinations of two points,
This is the process of moving. The process of moving the model by the two-point neighborhood search shown in FIG. 16 includes steps S321 and S3.
26 and S327, the process from step S322 to S325 is performed for each i while advancing the counter i from 1 to n-1. Steps S322 and S3
24, and the process of step S323 is performed for each j while advancing the counter j from i + 1 to n in S325. That is, all points vv _i, for the combination of vv _j, performs local search processing vv _i, vv _j points described below, to move the combination of decreasing energy when moving two points simultaneously. In the convergence determination of step S304 in FIG. 1, if a combination of decreasing energy is found in step S303, the process returns to step S302 to try to further minimize energy. If no combination of decreasing energy is found in step S303, it is determined that convergence has occurred, and the position of model at this time is set as the energy minimization position in step S305. Hereinafter, vv _i and vv in step S323 in FIG.
_A method of evaluating the change in E _{model in} the expression (9) in the neighborhood of vv _i and vv _j in the neighborhood search processing of _j will be described.

In equation (9), points vv _i and vv _j on model are expressed as δvv _i = (δv _1i , δv _2i ) and δvv _j
= (Δv _1j , δv _2j ), that is, vv
The change amount δ _ij E of E _{mode l in} the equation (9) when _i , vv _j are moved to vv _i + δvv _i , vv _j + δvv _{j
The model} is as in the following expression (18).

(Equation 15) This change amount δ _ij E _model becomes negative, that is, δ _ij E _model
The movements δvv _i and δvv _j such that _model <0 reduces E _model in equation (9).

In order to evaluate such changes in the energy function E _model of the equation (9) at the points vv _i and vv _j, the neighborhood energy evaluation function d as in the following equations (19) and (20) _i (δvv _i ), g _ij (δvv _i , δ
vv _j ) is defined.

[Equation 16] The amount of change δ _ij E _{model in the} equation (18) is (19), (2
0) neighborhood energy evaluation function d _i (δvv _i ), g
_It is expressed by the following equation (21) using _ij (δv _i , δv _j ). δ _ij E _model = d _i (δvv _i ) + d _j (δvv _j ) + g _ij (δvv _i , δvv _j ) −d _i (0) −d _j (0) −g _ij (0,0) ・ ・ ・ ( 21) Therefore, d _i (δvv _i ) + d _j (δvv _j ) + g
_ij (δvv _i , δvv _j ) <d _i (0) + d _j (0) + g _ij
If there are movements δvv _i and δvv _j to be (0,0), then vv _i and vv _j are vv _i + δvv _i and vv _j
By moving to + δvv _j simultaneously, E in Eq. (9)
_model decreases. That is, the current state (δvv _i =
Among combinations of neighboring points including δvv _j = 0), the neighborhood energy evaluation function d _i (δvv _i ) + d _j (δv
The movement may be such that v _j ) + g _ij (δvv _i , δvv _j ) is minimized.

FIG. 17 is a flowchart of the vv _i , vv _j neighborhood search process of step S323 in FIG. FIG. 18 is a conceptual diagram showing a situation of a two-point neighborhood search. Hereinafter, with reference to FIG. 17, step S3 in FIG.
The processing of 23 will be described. First, the neighborhood energy evaluation function d _i (δvv _i ) + d _j (δv) of the equation (21)
In order to efficiently calculate v _j ) + g _ij (δvvi, δvv _j ), d _i (δvv _i ), d _j (δvv _j ) is
Since the term depends only on the movement amounts δvv _i and δvv _{j of} each point, the values at each of the m neighboring points are calculated in advance in step S331 using the equation (19). At this time, if D _λi and D _λj are calculated in advance and this process is performed, the efficiency is further improved. Step S33
In 2, the minimum value E _min and the minimum value E _min of the energy evaluation value
Initialize the movements δvv _i , δvv _j to the neighboring points that give In steps 333 and S337, the counters h and k are respectively incremented from 1 to the neighborhood number m, and <
For all m × m combinations of h, k>, point v
v _i and vv _j are respectively the h-th neighbor point vv _i + δv
v evaluation process of energy when moving _h and k th neighboring point of vv _j + δvv _k simultaneously to S334, S335,
S336 is performed.

For example, if the number of neighboring points is m = 8, <
FIG. 18 shows how the h, k> = <7,5> th neighbor point is searched. In step 334, the evaluation value d _i (δvv _h ) + d _j (δvv _k ) + g _ij of the neighborhood energy by the movements δvv _h and δvv _k to the <h, k> -th neighborhood point.
(Δvv _h , δvv _k ) is calculated using d _i (δvv _i ) and d _j (δvv _j ) obtained in step 331. Step 334 is performed m × m times, but d
_{Since i} (δvv _i ) and d _j (δvv _j ) each take only m values, it is useless to calculate m × m times in step 334, and in step 331 each m value is calculated in advance. Is calculated and stored, the calculation in step 334 can be performed efficiently. In steps S335 and S336, the minimum value E _min of these evaluation values
And moving to a neighboring point that gives the minimum value E _min δvv _i , δ
Find vv _j . If the point giving the minimum value of the neighborhood energy evaluation function is not the current position in steps S338 and S339, the points vv _i and vv _j are set to vv _i + δvv _i and vv.
Move to _j + δvv _j . As described above, the points vv _i and vv _j can be moved to the point of decreasing the energy function E _{model of the} equation (9) among the neighboring points by the processing shown in FIG. As shown in Figure 1, mo
By performing this process for all combinations of two points on del, the model whose energy does not decrease in the one-point neighborhood search is moved to the position where the energy decreases further by the two-point neighborhood search. be able to.

As shown in FIG. 1, if the model moves by the two-point neighborhood search, the process returns to the one-point neighborhood search and the process is repeated, and the process ends when the points do not move even after the two-point neighborhood search. To do. Model at this time is
Since the energy function E _{mode l of the} equation (9) is the minimum,
The structure of the recognition target that satisfies the constraint condition given as the energy function can be extracted. As described above, in the present embodiment, the processing is ended even though the energy is not minimum because the dynamic model proposed above searches only the vicinity of one point, and the recognition target It is possible to extract the structure of the recognition target that minimizes the energy function accurately by solving the problem that the extraction of the structure such as the contour of the object may fail and moving the two points at the same time and repeating the search. There is. The present invention is not limited to the above embodiment, and various modifications can be made. The following are examples of such modifications. (1) In this embodiment, the method of performing a search when two points are moved at the same time has been described, but a search when three or more points are moved at the same time is also possible. That is,
Energy minimization processing is performed by calculating the amount of change in the energy function when moving a plurality of points of three or more points at the same time and repeating the movement of decreasing the energy. (2) In the present embodiment, the method of performing the two-point neighborhood search after the one-point neighborhood search has been described. However, the one-point neighborhood search process and the two-point neighborhood search process may be combined differently from the present embodiment. Various modifications such as processing are possible.

[0027]

As described in detail above, the first to sixth aspects
According to the invention, in the image recognition method, a search is made for a neighborhood of one point that moves so that energy is reduced when the position of the model that minimizes the energy function is moved to a neighborhood point of the one point on the model. Since it is determined by processing and simultaneous search processing of a plurality of points that move so that energy decreases when a plurality of points on the model are simultaneously moved to their respective neighboring points, the energy is accurately calculated. The structure of the recognition target that minimizes can be extracted.

[Brief description of drawings]

FIG. 1 is a flowchart of energy minimization processing of an image recognition method according to an embodiment of the present invention.

FIG. 2 is a diagram showing a conventional contour extraction method using an active contour model.

FIG. 3 is a diagram showing the active contour model of FIG. 2;

FIG. 4 is a flowchart showing energy minimization processing in FIG.

FIG. 5 is a diagram showing the dynamic model proposed above.

FIG. 6 is a diagram showing an energy term attracted to a fixed point.

FIG. 7 is a diagram showing terms related to a vector connecting two points.

FIG. 8 is a diagram showing terms related to a polygonal line connecting three points.

FIG. 9 is a diagram showing terms relating to the area of a triangle having three points as vertices.

FIG. 10 is a diagram showing terms relating to the area of an m-sided polygon having an apex at the point m.

FIG. 11 is a flowchart showing an energy minimization process of the above proposal.

FIG. 12 is a flowchart showing a neighborhood search process in FIG.

FIG. 13 is a diagram showing the neighborhood search proposed above.

FIG. 14 is a diagram showing the structure of the previously proposed dynamic model.

FIG. 15 is a diagram showing a problem in the previously proposed dynamic model.

FIG. 16 is a diagram showing movement of model by neighborhood search in FIG. 1.

FIG. 17 is a diagram showing movement of a model by a neighborhood search of two points in FIG.

FIG. 18 is a diagram showing a two-point neighborhood search process.

[Explanation of symbols]

 110 Original Image Data 120 Energy Image Calculation Processing 130 Energy Function Image 140 Energy Minimization Processing 160 Definition of Structural Model

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Yoshinori Shimosakoda 1-7-12 Toranomon, Minato-ku, Tokyo Oki Electric Industry Co., Ltd.

Claims (4)

[Claims] 1. A model representing a structure of a recognition target in an input image is defined as a model represented by an arbitrary number of points on the image, and a constraint condition for attracting the model to image features and a structure of the model. In an image recognition method for extracting an arbitrary structure from the input image by defining an energy function that minimizes energy when a constraint condition regarding is satisfied, and determining a position of the model that minimizes the energy function. , The position of the model that minimizes the energy function,
Search processing for the neighborhood of one point that moves so that the energy decreases when the point moves to one of the neighboring points on the model, and energy when moving a plurality of points on the model to the respective neighboring points at the same time. An image recognition method characterized in that it is obtained by performing simultaneous search processing of a plurality of points that move such that the number of points decreases in the vicinity of the points. 2. First, energy minimization processing is performed by a search processing in the vicinity of the one point, and then 2
The energy is minimized by performing a movement such that the energy is reduced when the points simultaneously move to a neighboring point, and if there is a movement, returning to the search processing in the neighborhood of the one point and repeating the processing. The image recognition method according to item 1. 3. In the simultaneous search process for the neighboring points of the plurality of points, the amount of change in energy is expressed as an evaluation function of the amount of movement of the plurality of points to be searched in the vicinity, and the energy is calculated based on the evaluation function. The image recognition method according to claim 1, wherein the movement is performed so as to decrease. 4. In the simultaneous search process for neighboring points of the plurality of points, one point of each point is selected from among the amounts of change in energy expressed as a function of the amount of movement of the plurality of points to be searched to the neighboring points. A term that depends only on the movement amount is calculated in advance before the search is started, and the term that depends only on the calculated movement amount of one point is used in the simultaneous search process for the neighboring points of the plurality of points. The image recognition method according to claim 3, characterized in that