JP3188899B2 - Image processing apparatus and method - Google Patents

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 method for evaluating image data using a predetermined evaluation function and extracting an outline from the image data by obtaining an optimum solution.

[0002]

2. Description of the Related Art Conventionally, image processing has been performed for object recognition and the like. For example, a computer processes an image signal obtained by a television camera to extract an outline. This contour extraction is based on extracting a point sequence whose luminance change in the image data exceeds a certain level.

As a method of extracting the contour, there is a method of evaluating image data with a predetermined evaluation function to find an optimal solution, and a well-known method of finding an optimal solution is a variational method. However, this variational method has a problem that the solution is unstable.

Therefore, it has been proposed to use a dynamic programming method which is easier to formulate than the variational method, is numerically stable, and has excellent convergence. In this dynamic programming, image data in a predetermined range around a contour model given as an initial value is sequentially evaluated by an evaluation function, and the contour model is gradually converged to an optimal solution (contour). The evaluation value of the point is set to be smaller as the luminance change of the target point is larger, and the contour model is updated in a direction in which the evaluation value is smaller. Image processing using such a dynamic programming method is described in, for example, Ami et al.'S magazine "IEEE Tran. On PAMI, vol.
no. 9, pages 855-867 ".

[0005]

As described above, stable contour extraction can be performed by using the dynamic programming method. However, when contour extraction is performed by the dynamic programming, there is a problem that the local minimum of the evaluation function becomes an optimal solution and correct contour extraction may not be performed. That is, in the above-described dynamic programming, the contour is updated in the direction of the smaller evaluation value for the pixel data in the predetermined range. If there is a place where the evaluation value is locally smaller than the surroundings, this place is regarded as the contour. It is determined that a correct contour cannot be extracted.

In addition, since image data often has noise, various filters for removing noise are often used. However, when such a filter is used, there is a problem that the contour itself in the image data is blurred and accurate contour extraction cannot be performed.

SUMMARY OF THE INVENTION The present invention has been made in consideration of the above problems, and has as its object to provide an image processing apparatus and an image processing method capable of performing high-speed and accurate contour extraction using a dynamic programming method.

[0008]

SUMMARY OF THE INVENTION An apparatus according to the present invention processes image data composed of a plurality of pixel data and processes the pixel data to be processed in consideration of the pixel data around the pixel. A Gaussian filter whose degree of consideration of pixel data can be changed, and a predetermined number of pixel data around each of the contour candidate pixels (contour models) to be processed from image data processed by the Gaussian filter are determined. Sampling at intervals, evaluating the contents of the sampled data by a predetermined evaluation function, and repeatedly moving the contour model to the optimum point in the sampled pixels by dynamic programming to obtain the optimum solution, An optimal solution calculating means for extracting an outline of an object and changing a pixel data sampling interval; Sa of the pixel data around the contour model with changing the degree of the peripheral pixel data considering the filter
Control means for changing the sampling interval in accordance with the degree of consideration of the peripheral pixel data.

Further, the apparatus according to the present invention processes pixel data to be processed for image data composed of a plurality of pixel data in consideration of pixel data in the vicinity of the pixel, and considers the peripheral pixel data. And a contour candidate pixel (limb model) to be processed is determined from image data processed by the Gaussian filter so as to have a predetermined pixel interval. A predetermined number of pixel data around the candidate pixel are sampled at predetermined intervals, the contents of the sampled data are evaluated using a predetermined evaluation function, and the contour model is moved to the optimum point in the pixel sampled by the dynamic programming method. To obtain the optimal solution, extract the contour of the object in the image, and set the pixel data sampling interval. And the optimal solution calculating section is variable, the difference in pixel data around the contour model with changing the degree of the peripheral pixel data considering the Gaussian filter
Control means for changing the sampling interval and the pixel interval of the contour candidate pixel in accordance with the degree of consideration of the peripheral pixel data.

[0010] A method according to the present invention is an image processing method for extracting a contour from image data composed of a plurality of pixel data, comprising an initial data inputting step using predetermined initial contour data as a contour model; A filtering process for filtering input image data in consideration of pixel data around a pixel to be processed; and a contour candidate pixel (contour) to be processed from the filtered image data. Model), a predetermined number of surrounding pixel data are sampled at predetermined intervals, and the contents of the sampled data are evaluated by a predetermined evaluation function. An optimal solution calculating step of obtaining contour data which is an optimal solution by repeating moving;
A contour data updating step of updating the contour data with the contour data obtained in the optimum solution calculating step, and changing the degree of consideration of peripheral pixels in the filtering processing step to be small, and reducing the data sampling interval in the optimum solution calculating step. And a condition changing step of changing the conditions. The conditions in the filtering processing step and the optimum solution calculating step are sequentially changed to perform a filtering process, calculate contour data, and finally extract a contour.

[0011]

As described above, according to the present invention, the degree of blur in the Gaussian filter is changed, and the moving range in the dynamic programming is changed in accordance with the degree. Therefore, it is possible to prevent the optimal solution (contour model) from being settled at the local minimum of the evaluation function by a dynamic programming method based on greatly blurred data, and to increase the moving range at this time to obtain the optimal solution. Time can be reduced. Since this is repeated one after another, efficient and accurate contour extraction can be performed. Further, in changing the pixel interval of the contour candidate pixel, by applying the number of pixels corresponding to the Gaussian filtering process, unnecessary pixels in the image blurred by the Gaussian filter are not calculated, and the processing is performed. Efficiency can be improved.

[0012]

【Example】

Embodiment 1 FIG. Hereinafter, an embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing the overall configuration of an image processing apparatus. As input means, an image input means 1 for inputting luminance (or color) data for each pixel, and an initial contour for giving an initial value of contour data It has an input means 2 and a display means 3 composed of a CRT or the like as an output means. In addition, as arithmetic means,
It has a Gaussian filtering operation unit 4 for performing a filtering process on input image data, and an optimum solution calculation unit 5 for calculating an optimum solution by dynamic programming. Further, as storage means, an image storage means 6 for storing input image data, and a Gaussian image storage means 7 for storing image data after Gaussian filtering.
And an outline data storage unit 8 for storing the outline data calculated by the optimum solution calculation unit 5. The control means 9 is for controlling the exchange of data between the above means.

Next, the overall operation of the image processing apparatus will be described with reference to the flowchart of FIG. To start the process, first, image data is input from the image input means 1 (S1).
This image data is digital data for each pixel obtained by AD-converting an analog signal obtained by a normal television camera or the like (for example, the luminance of the pixel is represented by 8 bits). Then, this data is stored in the image storage unit 6 via the control unit 9. This image storage means 6
Is usually composed of a RAM, and stores image data for at least one frame (one screen). Next, an initial contour and an initial σ are input (S2). This is normally in accordance with an instruction from input means such as a keyboard and a mouse. However, the control means 9 stores a predetermined initial contour (for example, a circle or an ellipse slightly smaller than the screen) in advance, and displays a menu on the display means 3. It is preferable to display which contour is to be selected in a format and to allow the operator to select it, or to input an initial contour with a mouse while displaying an image on the display unit 3.

Σ is a standard deviation for determining the degree of blur of the Gaussian filter expressed by the following equation.

The output F (x, y) of the Gaussian filter for the input image f (x, y) is given by F (x, y) = {G (x + u, y + v) · f (u, u)
v) dudv where G (x, y) = (1 / (2πσ ² )) exp (− (x ²
+ Y ² ) / (2σ ² )) Then, the standard deviation σ is changed from 8.0 → 4.0 → 2.0 → 1.0
In this case, the initial σ = 8.0 is input to the control means 9. Since the value of σ does not change much, it is preferable that the value of σ be stored in the control means 9 and read out from the internal storage section when necessary.

When one frame of image data and data for processing the image data are thus prepared, the Gaussian filtering operation means 4 performs a filtering process (S3). In this Gaussian filtering, in order to determine data on a pixel to be processed, it is determined from the data of the pixel and surrounding pixels. And this filtering is based on the above equation.
Weights for surrounding pixels corresponding to this equation are stored in advance, and data on pixels to be processed are sequentially determined by weighted averaging using the weights. Further, such a filtering process is performed in advance on one frame of image data, and the processed data is stored in the Gaussian image storage unit 7.

Next, an optimal solution by a dynamic programming method described later is calculated by the optimal solution calculating means 5 according to the Gaussian image subjected to the Gaussian filtering and the initial contour (S4). Since the optimal solution obtained by the optimal solution calculating means 5 is contour data, it is stored in the contour data storing means 8. Here, in the dynamic programming, data of each contour pixel obtained in advance and eight pixels around the contour pixel are extracted, and a predetermined evaluation function is evaluated for them, and the evaluation value of the entire contour model is minimized. Select a pixel. Then, in the present embodiment, the sampling scale (interval) of the pixel to be processed is made variable (near variable) as shown in FIG. That is,
As shown in FIG. 3, a pixel-to-pixel distance (scale), which is a movable range of each contour point, is defined as σ pixel. Therefore, if σ in the Gaussian filter is 8.0, data of pixels separated by 8 pixels is sampled and set as a movable range.

Then, it is determined whether or not the obtained optimal solution is the final solution (S5). If the optimal solution is not the final solution, the process returns to S4 to perform a search based on the obtained contour data. Do it again. This is repeated until the final solution is obtained. Note that whether or not the solution is final is determined based on whether or not the obtained optimum solution matches the previous solution.
Next, it is determined whether or not σ is final (S6). If it is not final, σ is updated (S7) and S3
Return to Then, Gaussian filtering is performed on the image data of the same frame stored in the image storage unit 6 using different σ, and the optimal solution is calculated by the dynamic programming method based on the data. This is repeated until σ ends (σ = 1 in this example) (σ = 8 → 4 → 2 → 1).

When the filtering process of σ = 8 is performed by such a process, the moving range by the dynamic programming becomes a wide range separated by 8 pixels. Therefore, for the Gaussian-processed image data, the time required for obtaining the optimal solution can be shortened. Further, a large value of σ of the Gaussian filter means that the degree of blur is large, as schematically shown in FIG. 4, and accurate contour extraction cannot be performed originally. Therefore, the rough outline extraction with a large moving range as described above is suitable. According to such processing, the solution stays in a place where the local evaluation value is low (local minimum) for two reasons, that a small change is absorbed by the filtering processing and that the moving range is large. Can be prevented. When σ is changed, the next dynamic programming is performed using the contour data obtained last time as an initial value, so that the time required to obtain an optimal solution can be shortened.

Description of Dynamic Programming The image processing in this embodiment is the extraction of a dynamic contour from a moving image, and this dynamic contour is expressed as a parameter (closed).
It is a model of a curve. v (s, t) means the position of the contour model at time t, and v (s, t) = (x (s, t), y (s, t)). Here, s is a parameter corresponding to the distance along the contour, and corresponds to a number when pixels extracted as the contour are counted in a certain direction. Note that I (x,
y, t) is the pixel intensity of the position at time t (x, y), v _s is once by the variable s in the variable v partial derivatives,
v _ss is the second partial differentiation of variable v with variable s, I _x ,
I _y, I _t, I _yy, I _xy, I _xx means a partial differential by subscript variable of I, respectively.

As the evaluation function, the following energy function is used. The energy used here is based on the sum of the internal energy defining the tendency of the model itself and the image energy defining the influence from the image. Next, these energies will be described individually. (A) to (d) are internal energies, and (e) to (e).
(G) is the image energy.

(A) "Term of sum of primary and quadratic splines" This term is a term indicating the smoothness of the contour, and is a variable v (s,
This is for the magnitude of the first and second partial differentiation of t). E _spline (s, t) = {v _s (s, t)} ² + Δv _ss
(S, t) { ² } / 2

(B) "Area term" This term is for dealing with a concave shape in the contour. _{E area (t) = 1/} 2 {x (s, t) y s (s, t) -x s (s, t) y (s, t)}

(C) "Term for distance averaging between contour points" The contour data is data for each pixel, and this term is used to approximate this difference approximation to a theoretical value. In the following equation, the average value of the distance between contour points is d. E _dist (s, t) = | d−‖v (s, t) −v (s−
1, t) ‖ |

(D) "Term of continuity of movement / deformation" The contour should not change so much over time. Thus, this term reduces the energy as the contour movement and deformation between frames are smaller. _{E continuity (s, t) =} {x s (s, t) -x s (s, t-1)} 2 + {y s (s, t) -y s (s, t-1)} 2

(E) "Term of edge potential" This term is a term of the edge potential defined from the magnitude of the spatiotemporal gradient of a moving image, and the energy decreases as the spatiotemporal luminance change increases. To do. E _edge (s, t) = − {I (x (s, t), y (s, t), t)} ² = (I _x ² + I _y ² + I _t ² )

(F) "Curvature level term" This term is a curvature level term defined from a luminance curve of an image to deal with an end point. E _term = − | I _yy I _x ² −2 I _xy I _x I _y + I _xx I _y ² |

(G) "Brightness-invariant term" Normally, the brightness of the corresponding point between frames does not change. Therefore, this term is set so that the smaller the change in brightness, the smaller the energy. E _intens = [I (x (s, t), y (s, t), t) -I (x (s, t-1), y (s, t-1), t-1)] ²

The sum of the seven energies (a) to (g) is the energy for the contour model to be processed. Since the moving range of each contour point is 8 pixels around it (9 pixels including the stop state), there are 9 ⁿ possibilities when the number of contour points is n. Are taken as contour points. Then, such processing is performed on the contour points obtained in advance, and the optimal point is calculated by one dynamic programming. Note that simply performing such processing on the contour points of one round leaves the influence of the starting point (the differential value is not the one after the change), so that, for example, one round and four rounds
It is assumed that the optimum solution of this time has been obtained when the calculation for one-half is performed. Then, when one optimal solution is obtained in this way, this is updated as the initial value of the contour data, and further calculation for optimization is performed. In this way, a final optimal solution by one dynamic programming can be obtained.

In this embodiment, since a contour is extracted from a moving image, a time t is included in the variable. Therefore, this t
Means that energy is defined by data between frames. Then, when the above energy is decomposed into energy defined between frames and energy defined within a frame, the following is obtained. [In frame] (a), (b), (c), (e) [space edge], (f) [interframe] (d), (e) [time edge],
(G)

Experimental Results Next, the results of extracting and tracking the contour of an object from a moving image using the apparatus of the present embodiment will be described. In this example, the following sequential procedure was adopted.

(1) With respect to the image of the first frame, starting from an arbitrary initial position surrounding the object, the outline of the object is extracted. Here, the initial position was input using a mouse.

(2) With respect to the images of the second and lower frames, the contour is extracted by using the inter-frame energy described above, with the result of the immediately preceding frame as the initial position.

In this embodiment, the standard deviation of the Gaussian filter and σ for determining the moving range of the dynamic programming are set to 1
(Σ: unchanged), 2 → 1 (σ: two-step change), 4 → 2 → 1
(.Sigma .: three-step change), and contour extraction was performed with changes such as 8.fwdarw.4.fwdarw.2-.fwdarw. (. Sigma .: four-step change). FIG. 5 shows the total number of iterations until an optimal solution is obtained in this case. As described above, the total number of iterations until the optimal solution is obtained decreases as the initial σ increases. Therefore, it is understood that the present embodiment achieves higher speed.

In FIG. 6, the value of σ in Gaussian filtering is changed in the same manner as in FIG. 5, but an optimal solution is obtained when the moving range (scale) by dynamic programming is fixed at 1.0. Shows the total number of iterations before it is done. Thus, in the case of the fixed neighborhood, it is understood that the number of iterations until the optimal solution is obtained is very large.

FIG. 7 shows the result of extracting the outline of the slug by the method of this embodiment (variable neighborhood: σ = 8 → 4 → 2 → 1), and FIG. 1.0
(These are the same) are shown. Thus, in the present embodiment, the correct contour was extracted, while σ
In the case of = 1, there is a place where the contour is erroneously extracted due to the local minimum of the evaluation function, and it is understood that the erroneous contour is extracted.

FIG. 9 is a graph in which the convergence result obtained when the slug outline is extracted by the apparatus and method of the first embodiment is displayed in a superimposed manner, and shows the experimental result more clearly.

Embodiment 2 FIG. FIG. 10 is a flowchart for explaining the operation of another embodiment of the present invention. In the figure, steps from step S1 of inputting image data by the image input unit 1 to Gaussian filtering S3 are the same as those in the first embodiment. The overall configuration is the same as in FIG.

The pixels to be processed are determined so that the contour candidate pixels have a predetermined pixel interval according to the degree of blurring of the Gaussian filter (S9). Then, the optimal solution by the above-described dynamic programming is calculated by the optimal solution calculating means 5 in accordance with the contour candidate processing target pixel and the Gaussian image on which the Gaussian filtering has been performed (S4). Since the optimal solution obtained by the optimal solution calculating means 5 is contour data, it is stored in the contour data storing means 8. Here, in the dynamic programming, data of each contour pixel obtained in advance and eight pixels around the contour pixel are extracted, and a predetermined evaluation function is evaluated for them, and the evaluation value of the entire contour model is minimized. Select a pixel. In this embodiment, the pixel interval of the contour candidate pixel is made variable, and the sampling scale (interval) of the pixel to be processed is made variable (variable neighborhood) as shown in FIG. For example, the pixel interval between the contour candidate pixels is twice σ, and the sampling scale is σ pixel, as shown in FIG. . Therefore, if σ in the Gaussian filter is 8.0, the contour candidate processing target pixel is determined so that the pixel interval becomes 16 pixels, and the data of σ pixels, that is, data of pixels separated by 8 pixels in this case, is sampled. It is assumed that each contour point can be moved.

Then, it is determined whether or not the obtained optimal solution is the final solution (S5). If the optimal solution is not the final solution, the process returns to S4 to perform a search based on the obtained contour data. Do it again. This is repeated until the final solution is obtained. Note that whether or not the solution is final is determined based on whether or not the obtained optimum solution matches the previous solution.
Next, it is determined whether or not σ is final (S6). If it is not final, σ is updated (S7) and S3
Return to Then, Gaussian filtering is performed on the image data of the same frame stored in the image storage unit 6 using different σ, and the optimal solution is calculated by the dynamic programming method based on the data. This is repeated until σ ends (σ = 1 in this example) (σ = 8 → 4 → 2 → 1).

According to such processing, when the filtering processing of σ = 8 is performed, the number of pixels to be processed decreases when the pixel interval between the contour candidate pixels is 16 pixels or more on average. And the moving range by dynamic programming is 8
It is a wide range separated by pixels. Therefore, for the Gaussian-processed image data, the time required for obtaining the optimal solution can be reduced, and the calculation of the optimal solution can be accelerated. Further, a large value of σ of the Gaussian filter means that the degree of blur is large, as schematically shown in FIG. 4, and accurate contour extraction cannot be performed originally. Therefore, the rough outline extraction with a large moving range as described above is suitable. According to such processing, the solution stays in a place where the local evaluation value is low (local minimum) for two reasons, that a small change is absorbed by the filtering processing and that the moving range is large. Can be prevented. Then, unnecessary pixels in the image blurred by the Gaussian filter are not calculated, and the processing efficiency can be improved.

Next, the effectiveness when the apparatus and method of this embodiment are applied to a short-axis ultrasonic tomographic image of the heart to extract and track the contour of a non-rigid object in an image containing noise. An example of the result showing is described with reference to FIGS.

The target image in this case is the first one of three moving images of the standard image database SIDBA, Vol. 701 (record numbers 2 to 37, data names NM01 to NM).
36, short axis tomogram near the mitral valve tip)
There are 36 grayscale images with 256 bits and 8 bits each. The procedure for extracting and tracking the contour of the heart is basically the same as that of the first embodiment. However, the first frame employs an expanded contour model in which the signs of the primary term spline and the area energy term are inverted. The energy terms at the curvature level were not used.

As shown in FIG. 11, in the image NM01, a contour is extracted from an initial position set inside by energy minimization. Next, as in the slug experiment shown in FIG. 9 (result of the first embodiment), the contours of the images NM02 and thereafter are traced using the results of the immediately preceding images as initial positions. Image NM01 representing the result of this contour tracking
The odd-numbered images from to NM29 are shown in FIG. From these results, it can be seen that the contour of the non-rigid object can be extracted and tracked almost correctly even in such an image including noise, and it is effective for obtaining quantitative diagnostic data in medical diagnosis.

[0045]

As described above, according to the image processing apparatus and method of the present invention, the degree of blur in the Gaussian filter is changed, and the moving range in the dynamic programming is changed in accordance with the degree. Therefore, it is possible to prevent the optimal solution (contour model) from settling at the local minimum of the evaluation function, and to shorten the time required to obtain the optimal solution.

[Brief description of the drawings]

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

FIG. 2 is a flowchart for explaining the operation of the present embodiment.

FIG. 3 is a diagram illustrating a moving range in dynamic programming.

FIG. 4 is an explanatory diagram of Gaussian filtering (in a one-dimensional case).

FIG. 5 is a diagram illustrating the number of iterations until an optimal solution is obtained in the case of a variable neighborhood.

FIG. 6 is a diagram illustrating the number of iterations until an optimal solution is obtained in the case of a fixed neighborhood.

FIG. 7 is an explanatory diagram illustrating an example of a contour obtained according to the present embodiment.

FIG. 8 is an explanatory diagram showing an example of a contour obtained by σ = 1 (processing on a single scale).

FIG. 9 is a diagram in which a convergence result in a case where slug outline extraction / tracking processing is performed by the apparatus and method of the first embodiment is superimposed and displayed.

FIG. 10 is a flowchart illustrating the operation of the second embodiment.

FIG. 11 is an explanatory diagram showing a result of applying the apparatus and method of the second embodiment to a short-axis ultrasonic tomographic moving image of the heart.

FIG. 12 is an explanatory diagram showing the result of applying the apparatus and method of Example 2 to a short-axis ultrasonic tomographic moving image of the heart.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Image input means 2 Initial contour input means 3 Display means 4 Gaussian filtering calculation means 5 Optimal solution calculation means 6 Image storage means 7 Gaussian image storage means 8 Contour data storage means 9 Control means

Continuing from the front page Examiner Isao Matsuura (56) References JP-A-5-61977 (JP, A) JP-A-3-190469 (JP, A) JP-A-60-243781 (JP, A) JP-A-60 −243780 (JP, A) Fujimura et al., “Motion image tracking of deformable object by energy minimization”, Proceedings of the 22nd Conference on Image Engineering, December 1991, p. 45-48 (58) Field surveyed (Int. Cl. ⁷ , DB name) G06T ^7/ 00-7/60 JICST file (JOIS)

Claims (3)

(57) [Claims] An image data composed of a plurality of pixel data is processed in consideration of pixel data to be processed in consideration of pixel data around the pixel, and the degree of consideration of the peripheral pixel data can be changed. For a certain Gaussian filter, and for each contour candidate pixel (contour model) to be processed from the image data processed by the Gaussian filter, a predetermined number of surrounding pixel data are sampled at a predetermined interval. The content is evaluated by a predetermined evaluation function, the contour model is repeatedly moved to the optimal point in the sampled pixels by the dynamic programming method to obtain an optimal solution, and the contour of the object in the image is extracted. An optimal solution calculating means capable of changing a pixel data sampling interval; While changing the degree of the pixel data around the contour model
Control means for changing the sampling interval in accordance with the degree of consideration of the peripheral pixel data. 2. An image processing method for extracting a contour from image data composed of a plurality of pixel data, comprising: an initial data input step of using predetermined initial contour data as a contour model; A filtering processing step of performing filtering processing in consideration of pixel data around a pixel to be processed; and for each contour candidate pixel (contour model) to be processed from the filtered image data, Sampling a predetermined number of pixel data at predetermined intervals, evaluating the contents of the sampled data by a predetermined evaluation function, and repeatedly moving the contour model to the optimum point among the pixels sampled by the dynamic programming method. Solution calculating step of obtaining contour data that is the optimum solution by using the optimum solution calculating step. A contour data updating step of updating contour data by contour data; and a condition changing step of changing the degree of consideration of peripheral pixels in the filtering processing step to be small and changing the data sampling interval in the optimum solution calculating step to be small. An image processing method comprising: sequentially changing conditions in the filtering processing step and the optimum solution calculating step, performing filtering processing, calculating contour data, and performing final contour extraction. 3. An image data composed of a plurality of pixel data is processed in consideration of pixel data around the pixel in consideration of pixel data to be processed, and the degree of consideration of the peripheral pixel data can be changed. A contour candidate pixel (limb model) to be processed is determined from a certain Gaussian filter and image data processed by the Gaussian filter so as to have a predetermined pixel interval. A predetermined number of pixel data is sampled at predetermined intervals, the contents of the sampled data are evaluated by a predetermined evaluation function, and the contour model is repeatedly moved to the optimum point in the sampled pixels by the dynamic programming method. The optimal solution is obtained, the contour of the object in the image is extracted, and the pixel data sampling interval can be changed. A solution calculating means, thereby changing the degree of the peripheral pixel data considering the Gaussian filter, the pixel data around the contour model
Control means for changing a sampling interval and a pixel interval of a contour candidate pixel in accordance with the degree of consideration of the peripheral pixel data.