JPH06290268A - Edge extraction processing method for image - Google Patents

Abstract

PURPOSE:To extract an edge image which is reduced in position error and has noises removed, from an input light and shade image. CONSTITUTION:An edge extraction part 1 filters the input light and shade image with scales sigmai (i=1...n) to extract (n) kind of edge images. A Gaussian filter ring part 2 filters the (n) kind of edge images with corresponding scales taui (i=1...n) to perform a defocusing process. A linear addition part 3 weights the (n) kind of defocused edge images as desired and linearly adds them. An edge determination part 4 performs a threshold processing for the output image of the linear addition part 3 to obtain an output edge image 20 as the final result. The respective edges obtained with plural scales are defocused by the Gaussian filter and the linear sum is utilized to enable edge extraction which is tolerant to noises and small in position error.

Description

Detailed Description of the Invention [Industrial applications]

[0001] The present invention relates to an edge extraction processing method for an image in which information of filters of different scales (resolutions) is integrated to extract edges when extracting edges in image processing.

[0002]

2. Description of the Related Art Conventionally, in image processing, it has been used to apply a filter to extract an edge portion. This is characterized in that the scale size can be changed freely and is relatively stable against noise.

However, when a filter is applied to perform edge extraction, if the scale size is large (filter with low resolution), a rough feature can be grasped, but a position error occurs in the obtained edge (for example, Lu, " Behavoir o
f Edges in Scale Space, "IEEE Trans PAM
I-11 pp337-356, 1989, etc.), conversely,
When it is small (filter with high resolution), the edge is calculated from local information, so that the position error is reduced, but there is a problem that it becomes sensitive to noise.

Therefore, in recent years, a multi-resolution edge extraction method for obtaining one edge expression from filter outputs of a plurality of scale sizes has been studied (for example, "Edge Conto").
ursUsing Multiple Scales, “Donna J. William
s, Mubarak Shah, Computer Vision, Graphics, a
nd Image Processing, 51, 1990, pp.256-
274). However, a clear and efficient multi-resolution edge extraction method has not been established yet.

[0005]

In the prior art, one of the reasons that makes the clear and efficient method of multi-resolution edge extraction difficult is that the alignment error between edges caused by the same edge portion is caused by the position error at each resolution. There was a problem of difficulty.

In order to solve the above-mentioned conventional problems, the present invention reduces the position error by using a linear sum obtained by blurring each edge obtained on a plurality of scales with a Gaussian filter, The purpose is to efficiently extract edges excluding noise.

The present invention further provides a certain resolution (scale).
The purpose of the present invention is to enable edge processing that emphasizes and soft edge processing for the local shape of an image.

[0008]

An image edge extraction processing method according to the present invention comprises a first step of extracting edges from an input gray-scale image using a plurality of filters of different scales, and a first step. A second step of applying a Gaussian filter having a scale corresponding to each edge extraction result obtained in the step, and a third step of linearly adding the processing results obtained in the second step, And a fourth step of detecting an edge point from the processing result obtained in the third step.

Further, in the image edge extraction processing method of the present invention, when the processing results obtained in the second step are linearly added together in the third step, one of a specific scale and other scales are used. It is characterized in that different weights are given to and those of linearly added.

Further, in the image edge extraction processing method of the present invention, when the processing results obtained in the second step are linearly added together in the third step, it differs depending on the local area on the image. It is characterized by giving weights and adding them linearly.

[0011]

According to the present invention, edges are extracted by a filter of a plurality of scales, a Cous filter of a scale corresponding to each of them is applied to blur, and the outputs thereof are linearly added to extract by various scales. It is possible to obtain one edge by integrating the information including the position error. Also, by blurring with a Gaussian filter and adding them linearly, it is possible to perform integration without matching the edges between the scales.

Further, when a Gaussian filter having a corresponding scale is applied to the edges obtained by each scale filter to add them linearly, a desired weight is given to emphasize the detailed information or the global information. It is possible to perform multi-resolution edge extraction while emphasizing the image or emphasizing the local shape of the image.

[0013]

Embodiments of the present invention will now be described in detail with reference to the drawings.

FIG. 1 shows the overall configuration of the present invention. In FIG. 1, 1 is an edge extraction unit, 2 is a Gaussian filtering unit, 3 is a linear addition unit, 4 is an edge determination unit, 10 is an input grayscale image, and 20 is an output edge image. here,
The edge extraction unit 1 is composed of a plurality of edge extraction units 1-1 to 1-n by filters of scale σ ₁ , ..., σ _n ,
The Gaussian filtering unit 2 is composed of a plurality of Gaussian filtering units 2-1 to 2-n having scales τ ₁ , ..., τ _n .

The processing is performed in the order of the edge extraction unit 1, the Gaussian filtering unit 2, the linear addition unit 3, and the edge determination unit 4. The processing of each unit will be described below in order. The input image 10 is a digital grayscale image.

(Edge extraction unit 1) For the input grayscale image 10, a plurality of scales σ ₁ , ..., σ _n are determined. As a method of determining, for example, Japanese Patent Application No.
No. 50415, “Filter scale selection method” can be used.

The edge extraction units 1-1 to 1-n apply a filter of the scale σ _i (i = 1, ..., N) to the input grayscale image 10 to extract edge images (multiresolution edge extraction). As a filter, for example, the Laplacian Gaussian filter of the following formula

[Equation 1] and so on. The larger the scale σ, the lower the resolution, and the smaller the scale σ, the higher the resolution. Each edge image extracted by the edge extraction unit 1 may be binarized or may be obtained as a grayscale image as an intensity value.

(Gaussian Filtering Unit 2) The Gaussian filtering units 2-1 to 2-n respectively scale n 1 (n resolutions) of edge images extracted by the edge extracting unit 1 into scales τ ₁ , ... , Τ _n Gaussian filter

[Equation 2] And the blurring process is performed. The blur is small when τ is small, and becomes large as τ is increased.

In the Gaussian filtering unit 2, each τ
When selecting _i , it is possible to consider the range in which the edges obtained by the edge extraction unit 1 may be displaced. For example, as shown in FIG. 2A, a point a and a point b in which the range in which the position error may occur are large and small as the output results of the filters of σ _a and σ _b , respectively. Suppose that it was obtained in. At this time, in the Gaussian filtering unit 2, by giving a large τ _a and a small τ _b in consideration of the range of the position error to each a and b, the output of the next linear addition unit 3 is ), A peak can be obtained within the positional error of each edge.

As a concrete value of τ, the larger the σ at the time of edge extraction (the lower the resolution) is, the larger the position error becomes. Therefore, as τ, the above-mentioned σ ₁ , ..., σ _n , or its It is possible to use a constant multiple.

(Linear Adder 3) The n kinds of blurred edge images obtained by the Gaussian filtering unit 2 are linearly added. That is, in the linear addition unit 3, each blurred edge image GE
Determine weights a _i for _i (i = 1, ..., N),

[Equation 3] To calculate.

Here, by changing the value of the weight α _i , it is possible to easily emphasize the edge of a certain specific scale, and it is possible to perform flexible edge extraction according to the image and the purpose. Hereinafter, some of the methods will be described.

(1) Method of giving uniform to all resolutions When uniform weighting is given to all resolutions, the probability that a peak appears near an edge image blurred with the smallest τ becomes high. An example is shown in FIG. Now q ₁ , q ₂ , q ₃
Are edge points obtained by using the scale σ _i in the edge extraction unit 1. However, σ ₁ <σ ₂ <σ ₃ and τ ₁ in the Gaussian filtering unit 2 is also the same. An output example of the Gaussian filtering unit 2 is shown in (1) of FIG. 3, and a normalized output of the linear addition unit 3 in this case is shown in (2) of FIG. It can be seen that the peak appears near q ₁ , which has the highest positional accuracy. Therefore, by tracking this peak, an edge can be obtained in the vicinity of q ₁ with the smallest position error.

(2) Method of giving the image following the visual sensitivity curve of the living body Generally, in the qualification sensitivity curve of the living body, the spatial frequency bands of the middle three channels among the spatial frequency bands of the six channels have high sensitivity. (For example, Tasaki, Oyama, Hiwatari, "Visual Information Processing" Asakura Shoten, 1979). With reference to this, it is conceivable that the intermediate channel is given strong and the low frequency and high frequency channels are given weakly. According to the visual sensitivity of the living body, the ratios are approximately as follows in order from high frequency. α ₁ : α ₂ : α ₃ : α ₄ : α ₅ : α ₆ = 1: 2: 4: 4: 4:
Two

(3) Method of emphasizing certain resolution Further, it is conceivable that certain resolution is emphasized. For example, as shown in FIG. 4A, it is assumed that there are two edge points α and β obtained by σ _α and σ _β , and a Gaussian filter is applied to them with a uniform τ. At this time,
By giving the respective weights a _α and a _β as a _α <a _β , the edge point β can be emphasized. Figure 4
(2) shows a result obtained by linearly adding and normalizing a _α : a _β = 2: 3 with respect to FIG. 4 (1). Further, for example, in the example shown in (2) above, α ₁ : α ₂ : α ₃ : α ₄ : α ₅ : α ₆ = 0: 0: 0: 1: 2:
By giving the method of 1, fine features are eliminated,
It becomes possible to extract edges centered on the outline information.

(4) Method of locally changing the weight As described above, in addition to making the integrated weight α _i uniform over the entire screen, this weight a _i is a function a of the position coordinates (x, y).
By giving _i (x, y), it becomes possible to perform soft edge processing on the local shape. For example, an output obtained by applying a uniform τ Gaussian filter to the output obtained by the edge extraction unit 1 by the filters of σ _α and σ _β is GE _α ,
Let GE _β . At this time, on a point that belongs to the partial area I ₀ of the image I, increasing the weight a _alpha of GE _alpha, the weight of GE _beta a _beta
It is possible to enhance GE _α only in the partial region I ₀ by making the value smaller. In this case, there is a possibility that the edge is interrupted at the boundary of I _0, by varying the slowly a _alpha, a _beta at the boundary of I _0, it is possible to obtain a smooth output. This is shown in FIG. That is, as shown in FIG. 5A, I ₀ , r ₀ ≤ a region where the distance r from the center of the image is r <r _0.
The r <r ₁ and a region of the area to be I _1, r ≧ r ₁ and I _2. For these three regions, t that changes depending on the value of r
_Is used to weight a _α : a _β = 1−t: t. However, t = 0 (r <r ₀ ), t = (r−r ₀ ) / (r ₁ −
r ₀ ), t = 1 (r ≧ r ₁ ). The value of t is shown in FIG. By giving the weight a _i (x, y) in this way, GE _α is emphasized in the central portion I ₀ and the peripheral portion I ₂
In GE, the GE _β is emphasized, and in the boundary portion I ₁ , the central portion I ₀
And the output of the peripheral portion I ₂ can be smoothly connected.

(Edge determination unit 4) An edge is determined from the output obtained by the linear addition unit 3, and finally the output edge image 2
Get 0. As an edge determination method, (1) a method in which an output intensity value having a certain threshold value or more is obtained and regarded as an edge (2) neighborhood of a point having an output intensity value having a certain threshold value th1 or more At a point, a method of tracking a thing having a threshold value th2 or more and regarding it as an edge is performed. Method of considering as edge (4) A method of giving the output intensity value to the energy function of the contour model and converging the contour model can be considered. As the contour model in the method (4), Snakes (M. Kass, A. Witkin,
D. Terzopoulos, “Snakes: Active contourmodel
s ", Proc. ICCV-87, pp.259-268, 1
987)).

6 to 23 show specific examples of the processing. FIG. 6 is an example of an input grayscale image. Here, a digital grayscale image is shown, and the number of pixels is 512 × 512.

FIGS. 7 to 12 show an example in which the input grayscale image of FIG. 6 is filtered with different scales σ _i (i = 1, ..., 6) to perform edge extraction with 6 types of resolutions. is there. Here, a Laplacian-Gaussian filter is used as a filter, and a method of extracting a zero-cross point is adopted. The filter scales are σ ₁ = 1.74 and σ ₂ =
3.34, σ ₃ = 6.52, σ ₄ = 9.24, σ ₅ = 1
5.1 and σ ₆ = 32.0. As shown in FIGS. 7 to 12, when the scale σ is small (high resolution), complicated edges are extracted, and conversely, when the scale σ is large (low resolution), it is rough, but a positional shift occurs.

FIGS. 13 to 18 show the same scales (ie, τ ₁ = 1.74, τ ₂ = 3.34, σ _{1 to} σ _{6) for} the edge images of FIGS. τ ₃ =
6.52, τ ₄ = 9.24, τ ₅ = 15.1, τ ₆ = 3
It is a blurred edge image obtained by applying the Gaussian filter of (2.0). It is to be noted that, in order to better understand the state of blurring, a partial enlargement of each of the images in FIGS. 13 to 18 is collectively shown in FIG.

FIG. 20 is an output image in which the six types of blurred edge images of FIGS. 13 to 18 are linearly added with the weights a _i (i = 1, ..., 6) being uniform.

In FIG. 21, the output image of FIG. 20 is subjected to threshold processing, and this is the edge extraction image of the final result. However, the maximum value of the linear addition result is 2 here.
A final edge was obtained by normalizing to 55 and starting a point having a threshold value th1 = 100 or more as a starting point and tracing a ridge point having a maximum value in the vicinity and a threshold value th2 = 60 or more. Figure 21
As shown in FIG. 7, according to the method of the present invention, it is possible to reduce the position error and obtain an edge image from which noise is removed.

FIG. 22 shows weights a _i (i = 1, ..., 6) for the six types of blurred edge images shown in FIGS.
As α ₁ : α ₂ : α ₃ : α ₄ : α ₅ : α ₆ = 0: 0: 0: 1: 2:
It is an edge extraction image obtained by linearly adding by giving 1 and similarly performing threshold processing. In this case, fine features are excluded and an edge image centered on the outline information is obtained.

In addition, FIG. 23 shows that the weights a _i (i = 1, 1) for the six types of blurred edge images shown in FIGS.
, 6) is given as the function a _i (x, y) of the position coordinate (x, y) as described with reference to FIG. 5, linear addition is performed, and the edge extraction image is obtained by similarly performing threshold processing. . However,
r ₀ is 1/6 of the image width and r ₁ is 1/12 of the image width
It was used. From FIG. 23, it can be seen that fine edge extraction is locally performed.

[0035]

According to the present invention, by performing edge extraction with an edge extraction filter of a plurality of scales, it is possible to consider local information and global information at the same time. By multiplying them and adding their outputs linearly, it is possible to combine the different outputs obtained by the edge extraction filters of various scales, so that the position error is small but the noise-sensitive small scale and the noise-resistant scale are strong. By using the information of a large scale with a large number of position errors, edge extraction with a strong noise and a small position error can be realized.

In the integration, firstly,
When blurring with the Kauss filter, by selecting the scale of the Gaussian filter in consideration of the position error range of the extracted edge, it is possible to more accurately extract the edge from which the position error is excluded. Secondly, by adding a Gaussian filter of the corresponding scale to the edges obtained by the edge extraction filter of each scale and adding them linearly, by changing the weight, detailed information can be emphasized or global information can be emphasized. Can be emphasized or local information of the image can be emphasized, so that flexible multi-resolution edge extraction can be performed for each image.

[Brief description of drawings]

FIG. 1 is an overall configuration diagram of the present invention.

FIG. 2 is a diagram illustrating selection of a scale used in Gaussian filtering.

FIG. 3 is a diagram showing an example in which weights used in linear addition are uniformly applied to all resolutions.

FIG. 4 is a diagram showing an example in which weights used in linear addition are given while being changed between resolutions.

FIG. 5 is a diagram illustrating a case where a weight used in linear addition is locally changed.

FIG. 6 is an example of an input grayscale image.

7 is an edge image when σ ₁ = 1.74 with respect to the input grayscale image of FIG.

8 is an edge image in the case of σ ₂ = 3.34 with respect to the input grayscale image of FIG. 6;

9 is an edge image when σ ₃ = 6.52 with respect to the input grayscale image of FIG.

10 shows σ ₄ = 9.54 for the input grayscale image of FIG.
It is an edge image in the case of.

11] σ ₅ = 15.1 for the input grayscale image of FIG.
It is an edge image in the case of.

FIG. 12 shows σ ₆ = 32.0 for the input grayscale image of FIG.
It is an edge image in the case of.

13 is a blurred edge image when τ ₁ = 1.74 with respect to the edge image of FIG. 7.

14 is a blurred edge image when τ ₂ = 3.34 with respect to the edge image of FIG.

15 is a blurred edge image when τ ₃ = 6.52 with respect to the edge image of FIG. 9;

16 is τ ₄ = 9.54 for the edge image of FIG.
It is a blurred edge image in the case of.

FIG. 17 shows τ ₅ = 15.1 for the edge image of FIG.
It is a blurred edge image in the case of.

FIG. 18 shows τ ₆ = 32.0 for the edge image of FIG.
It is a blurred edge image in the case of.

FIG. 19 is a partially enlarged view of FIGS. 13 to 18.

FIG. 20 is an output image obtained by linearly adding the blurred edge images of FIGS. 13 to 18 with uniform weighting.

FIG. 21 is a final edge extraction image obtained by thresholding the output image of FIG. 20.

22 is a final edge extracted image of the blurred edge images of FIGS. 13 to 18 when the weight corresponding to high resolution is set to 0. FIG.

FIG. 23 is a final edge extraction image in the case where the weights are locally changed in the blurred edge images of FIGS. 13 to 18.

[Explanation of symbols]

 1 edge extraction unit 2 Gaussian filtering unit 3 linear addition unit 4 edge determination unit 10 input gray image 11 output edge image

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Isamu Isawa 1-1-6 Uchisaiwaicho, Chiyoda-ku, Tokyo Nihon Telegraph and Telephone Corporation

Claims (3)

[Claims] 1. A processing method for extracting an edge, which is a set of feature points, from an input grayscale image, the first step of extracting each edge from the input grayscale image by using a plurality of filters of different scales, A second step of applying a Gaussian filter with a scale corresponding to each edge extraction result obtained in the first step, and a linear addition of each processing result obtained in the second step. An image edge extraction processing method comprising: a step; and a fourth step of detecting an edge point from the processing result obtained in the third step. 2. The image edge extraction processing method according to claim 1, wherein in the third step, when the respective processing results obtained in the second step are linearly added, those of a specific scale and other An image edge extraction processing method characterized by adding different weights to those of a scale and adding them linearly. 3. The image edge extraction processing method according to claim 1, wherein in the third step, when the processing results obtained in the second step are linearly added, the processing is performed according to a local region on the image. An image edge extraction processing method characterized by adding different weights and adding them linearly.