JP2002049922A - Method for recognizing image - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an image recognizing method for accurately recognizing a recognition object without calculating the initial setting condition of a closed curve to be arranged to the recognition object in a dynamic contour extracting method for extracting the contour of the specific recognition object included in an image. SOLUTION: When a recognition object 110 included in an image 100 is recognized, external energy is dynamically changed in accordance with the image value of the image 100, respective control points on the closed curve 120 are shifted in the direction where the closed curve 120 is reduced/expanded according to the external energy to recognize the contour of the recognition object 110 included in the image 100.

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, and more particularly to an energy control method in a dynamic contour extraction method.

[0002]

2. Description of the Related Art As a method for extracting a contour of a specific recognition target included in an image, a dynamic contour extraction method (Snakes:
“Active contour model”, International Journal of
Computer Vision, Vol.1, No.4, pp.321-331, 1988)
It has been known.

In this method, a deformable closed curve is set for a specific recognition target included in an image, and the energy of the closed curve is defined to quantitatively evaluate the state of the closed curve. . The energy of the closed curve is defined so as to be minimized when the closed curve matches the contour of the object to be recognized. From this, the contour of the object to be recognized is deformed by deforming the closed curve so that the energy becomes minimum. Can be extracted. The closed curve is obtained by connecting discretely arranged control points forming the closed curve. Therefore, the above-described process of deforming the closed curve means moving each control point on the closed curve so that the energy of the closed curve is minimized. Specifically, the control point v (on the closed curve is defined so that the sum of the three energies of the internal energy (internal spline energy), the image energy, and the external energy, which are defined corresponding to the state of the closed curve, is minimized. By moving s), the contour of the recognition target is extracted. The energy of the closed curve at this time is expressed by an equation as shown in [Equation 1].

(Equation 1)

[0004] The internal spline energy is composed of a first derivative and a second derivative of a control point, and is an energy for applying a force for contracting and smoothing a closed curve.
It can be represented by the equation shown in [Equation 2].

(Equation 2)

[0005] The image energy is given by the differentiation of the image at the control point coordinates, and is the energy that exerts a force to stick the closed curve to the edge of the image.
It can be represented by the following equation.

(Equation 3)

[0006] The external energy is energy that gives energy to a closed curve from the outside and exerts a force to contract or expand the closed curve, and is expressed by the following equation (4).

(Equation 4)

[0007]

In the conventional active contour extraction method, the closed curve can be deformed in only one direction, that is, in the direction of reduction or in the direction of expansion. Therefore, when performing control to deform the closed curve in the reduction direction when starting the image recognition process, it is necessary to arrange the closed curve so as to surround the recognition target. Further, when performing control to deform the closed curve in the expansion direction, it is necessary to initialize the closed curve so that all line segments are included in the recognition target. If the initial setting of the closed curve is set so as to intersect with the recognition target area, the closed curve will not stick to the edges of the entire circumference of the recognition target and recognition will fail. That is, as shown in FIG. 3, when the initial closed curve 320 formed by connecting the control points 330 to 343 is set to intersect with the recognition target 310, the initial closed curve 320 is changed to the closed curve 330 by performing the reduction control. Because of the deformation, the closed curve 330 does not stick to the edges of the entire circumference of the recognition target 310, and the recognition target 310 cannot be accurately recognized. That is, the accuracy of recognizing the recognition target in the image in the conventional active contour extraction method largely depends on the initial coordinates of the closed curve, and the recognition target may be recognized depending on the state of the initial setting of the closed curve. There is a possibility that it cannot be done.

[0008] In the image recognition processing by a computer system to which the conventional active contour extraction method is applied, a problem that occurs when the processing speed is increased will be described below. FIG. 4 shows a data flow when image processing is performed by a general computer system. When performing various types of image processing in the computer system 400, data necessary for the image processing is stored on the hard disk 440.
The data is read from a sensor such as a CCD camera or the like into the main memory 430, the data is transferred to the CPU 420, and then stored in the cache memory 410. If necessary data has already been transferred to the cache memory 410,
The CPU 420 reads the data from the cache memory 410. At this time, generally, the cache memory 41
The access speed to 0 is high, whereas the access speed to the main memory 430 is low. However, since the cache memory 410 is expensive, only a small size is mounted in a computer system, whereas the inexpensive main memory 430 is generally mounted in a large capacity.

Due to the above computer structure, in order to increase the processing speed of image processing, it is important to effectively use the cache memory 410 which can access data at high speed. The ratio indicating that the data required by the CPU 420 exists in the cache memory 410 is collectively referred to as a hit ratio. Improving the hit ratio is directly linked to an increase in speed. As a method of improving the hit ratio in a computer system to which the active contour extraction method is applied, the image data is divided into small regions, a closed curve is determined, and a target is recognized for each of the regions. Can be reduced to a small amount. However, as described above, in the conventional active contour extraction method, when arranging a closed curve, it is necessary to initialize the closed curve so as not to intersect with the recognition target area. However, when the recognition target 520 is arranged on the image 500 as shown in FIG. 5, the dividing line 510 for dividing the image 500 has a complicated arrangement, and a problem arises that the closed curve cannot be easily determined. Further, even when the image is divided by the dividing line 510 and a closed curve can be determined, the processing time for solving the dividing line arrangement problem becomes an overhead.

Therefore, according to the present invention, in a dynamic contour extraction method for extracting a contour of a specific recognition target included in an image, the recognition target is determined without obtaining an initial setting condition of a closed curve arranged with respect to the recognition target. An object is to provide an image recognition method capable of accurately recognizing an object.

[0011]

According to a first aspect of the present invention, there is provided an image recognition method comprising: arranging a closed curve on an image; an internal energy indicating a smoothness of the curve; an image energy indicating a density gradient of the image; Using external energy from
In the dynamic contour extraction method of performing a reduction process or an expansion process of the closed curve and extracting a contour of a recognition target included in the image, dynamically changing the external energy according to a density value of the image. Features.

According to a second aspect of the present invention, in the image recognition method according to the first aspect, the image is a binary image including a recognition target and a background portion, and the external image in the recognition target is used. The energy is changed in a direction in which the closed curve expands, and the external energy in the background portion is changed in a direction in which the closed curve contracts.

The image recognition method according to claim 3 is
2. The image recognition method according to claim 1, wherein the closed curves are arranged in a plurality of tiles.

[0014]

(Embodiment 1) Hereinafter, an image recognition method according to Embodiment 1 will be described with reference to FIG. In FIG. 1, a binary image 100 includes a recognition target 110 to be recognized and an exterior (background portion) 16 of the recognition target.
0, and the density value (pixel value) of the recognition target object 110 is 1
, And the pixel value of the background portion 160 is set to 0.

Hereinafter, a method for recognizing the recognition target 110 in the binary image 100 shown in FIG. First, as shown in FIG.
0 is set to intersect with the recognition target 110. Here, the closed curve indicates a connection of control points (130 to 137, 140, 141) discretely arranged on the closed curve. In the process of connecting these control points,
It is assumed to have directionality, and the counterclockwise direction is defined as positive. The basic control of the closed curve 120 is performed by moving and controlling each control point such that the energy E _SNAKE (v) of the closed curve shown in [Equation 1] is minimized.

Next, it is determined whether each control point on the closed curve 120 is located inside or outside the recognition target 110. This determination is made at 2 in each control point coordinate.
This is performed based on the pixel values of the value image 100. When the pixel value at the control point coordinates is 1, it is determined that the pixel is located inside the recognition target 110. On the other hand, when the pixel value at the control point coordinates is 0, the recognition target is determined. 110, that is, the background portion 160
It is determined to be located at. In FIG. 1, control points outside the recognition target object 110 are indicated by black circles (130 to 13).
7), and the inside control points are indicated by white circles (140, 14).
1).

Next, the external energy is dynamically changed,
The moving direction of each control point on the closed curve 120 is determined. For a control point outside the recognition target 110, a vector obtained by connecting control points before and after the target control point and rotated by 90 degrees is used as external energy. For the control points inside the recognition target 110, a vector obtained by connecting the control points before and after the target control point and rotated by minus 90 degrees is used as external energy. Note that a normalized value may be used as the absolute value of the vector.

Specifically, for example, when the control point to be controlled is the control point 131 shown in FIG.
Exists outside the recognition target 110, so that the control point 13
The vector 145 obtained by connecting the control points 130 and 132 located before and after 1 is rotated by 90 degrees to generate the external energy 146. The generated external energy 146 is transmitted to the control
Acts from outside to inside. And the control point 131 is
According to the external energy 146, the recognition target object 110 moves from outside to inside, that is, in a direction in which the closed curve 120 is reduced. As described above, external energy is applied to the control points (130 to 137) located outside the recognition target object 110 from the outside to the inside of the recognition target object 110, thereby controlling points (130 to 137). Is moved in a direction in which the closed curve 120 is reduced. In addition, external energy is applied to the control points (140, 141) located inside the recognition target object 110 from the inside to the outside of the recognition target object 110, thereby controlling points (140, 141).
In the direction in which the closed curve 120 expands. As a result, the initial closed curve 120 is transformed into a closed curve 150 and attached to the edges of the entire circumference of the recognition object 310, and the outline of the recognition object 110 is recognized.

As described above, according to the image recognition method of the first embodiment, the recognition target 11 included in the binary image 100
When recognizing the contour of 0, the external energy is dynamically changed according to the image value of the binary image 100, and each control point on the closed curve 120 is changed to the closed curve 12 according to the external energy.
Since 0 is moved in the direction of contraction / expansion, the contour of the recognition target 110 can be accurately recognized even when the closed curve 120 is initially set to intersect with the recognition target 110.

(Embodiment 2) An image recognition method according to Embodiment 2 will be described below with reference to FIG. The image recognition method according to the second embodiment divides an image into small regions in order to increase the processing speed when recognizing a specific recognition target included in the image, and recognizes each region. It is characterized by recognizing an object.

Specifically, first, the image 200 is divided
10 is divided into small areas, and closed curves (231 to 238)
To determine. By performing such processing, as shown in FIG. 2, a state is obtained in which a plurality of closed curves are arranged in a tile shape with respect to the image 200.

Next, the recognition target 220 is recognized for each divided area using the image recognition method of the first embodiment. In the second embodiment, the first embodiment is used for each divided area.
Since the recognition target 220 is recognized using the image recognition method described above, when arranging a plurality of closed curves in the image 200 in a tile shape, the recognition target 220 may be closed curves (231 to 231).
238) does not need to be initialized.

As described above, according to the image recognition method of the second embodiment, a plurality of closed curves (231 to 23)
8) are arranged in a tile shape, and the recognition target 22
Since the image 200 is recognized using the image recognition method of the first embodiment, the image 200 is simply divided into tiles without obtaining the initial setting condition of the closed curve for the recognition target 220, and each area is divided into tiles. The recognition target 220 can be recognized quickly, and the recognition processing speed of the recognition target 220 can be improved.

[0024]

As described above, according to the image recognition method of the first aspect, a closed curve is arranged on an image, and the internal energy indicating the smoothness of the curve and the image energy indicating the density gradient of the image. Using an external energy from the outside, performing a reduction process or an expansion process of the closed curve, and extracting a contour of a recognition target included in the image, a dynamic contour extraction method, wherein the external energy is a density of the image. Since it is dynamically changed according to the value, the reduction and expansion control of the closed curve can be performed according to the pixel value of the image, and the initial setting condition of the closed curve to be arranged with respect to the recognition target is It is possible to accurately recognize the outline of the recognition target included in the image without obtaining the outline.

According to the image recognition method of the second aspect, in the image recognition method of the first aspect, the image is a binary image including a recognition target and a background portion. Since the external energy in the object is changed in the direction in which the closed curve expands, and the external energy in the background portion is changed in the direction in which the closed curve contracts, the control of the reduction and expansion of the closed curve is performed on the pixels of the image. This can be performed according to the value, and the contour of the recognition target included in the binary image can be accurately recognized without obtaining the initial setting condition of the closed curve to be placed on the recognition target.

According to the image recognition method of the third aspect, in the image recognition method of the first aspect, the closed curves are arranged in a plurality of tiles.
The image is simply divided into small areas by dividing lines without finding the initial setting conditions of the closed curve to be placed on the recognition object,
The recognition target can be accurately recognized for each area,
It is possible to improve the recognition processing speed of a recognition target included in an image.

[Brief description of the drawings]

FIG. 1 is a diagram for describing an operation outline of an image recognition method according to a first embodiment;

FIG. 2 is a diagram for explaining an operation outline of an image recognition method according to a second embodiment.

FIG. 3 is a diagram for explaining operation control of a conventional active contour extraction method.

FIG. 4 is a diagram showing a data flow when image processing is performed by a general computer.

FIG. 5 is a diagram for explaining a problem in increasing the processing speed of image recognition processing in the conventional active contour extraction method.

[Explanation of symbols]

100 Binary image 110, 220, 310 Recognition target 120, 150, 320, 330 Closed curve 130, 131, 132, 133, 134, 135, 1
36,137,140,141,330,331,33
2,333,334,335,336,337,33
8,339,340,341,342,343 Control point 145 Vector 146 External energy 160 Background part 200,500 Image 210,510 Dividing line 400 Computer system 410 Cache memory 420 CPU 430 Main memory 440 Hard disk

Claims (3)

[Claims] 1. A closed curve is arranged on an image, and the closed curve is reduced by using internal energy indicating the smoothness of the curve, image energy indicating a density gradient of the image, and external energy from outside. A dynamic contour extraction method for performing an expansion process and extracting a contour of a recognition target included in the image, wherein the external energy is dynamically changed according to a density value of the image. . 2. The image recognition method according to claim 1, wherein the image is a binary image including a recognition target and a background portion, and the external energy in the recognition target is applied in a direction in which the closed curve expands. An image recognition method, wherein the external energy in the background portion is changed in a direction in which the closed curve is reduced. 3. The image recognition method according to claim 1, wherein the closed curve is arranged in a plurality of tiles.