JPH1153555A - Method for extracting segment and arc - Google Patents

Abstract

PROBLEM TO BE SOLVED: To perform stable and precise extraction while shortening processing time by introducing a floating polling box on a Hough plane and assuming a shift of an edge dot array, and then extracting segments and arcs. SOLUTION: An image obtained by an image acquiring means is spatially differentiated as a source image to find the differential values du and dv and gradient direction θ of each pixel, and pixels having larger differential values than the differential values du and dv of the pixels on both its sides along the gradient direction θ are regarded as edge points pi . Then it is assumed that an edge dot array consisting of a set of these edge points pi constitute a straight line 1. Then transformation of the straight line to one point on the Hough plane defined by two parameters, the length ρ1 and the angle θ1 of the perpendicular dropped to the straight line 1 from the image origin O, is performed. A floating polling box 1 corresponding to a certain permissible value is introduced around the one point and the edge points are transformed to the Hough plane. When the edge point is in the floating polling box 1, it is judged that the edge point pi constitutes the straight line.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a method for extracting a line segment / arc. In particular, a floating ballot box is introduced on a so-called Hough plane to reduce noise.

[0002]

2. Description of the Related Art In object recognition on a production line, an object which can be represented by a line segment and an arc is often handled. It is effective to apply model-based matching in which a model shape of an object is represented by features such as line segments and arcs and fitting is performed using features extracted from an image.

[0003] Model-based matching is a technique having a feature that is relatively insensitive to deformation and disturbance of a target object. In order to realize stable and accurate object recognition in this method, it is important to accurately extract features such as line segments and arcs.

[0004] On the other hand, it is also necessary to suppress an increase in processing time due to higher precision of feature extraction. As a conventional object recognition method, a method including acquisition of an edge image and Hough transform will be described.

[Acquisition of Edge Image] An edge image is obtained by differentiating the gray scale image obtained from the CCD camera as an original image. Let the image coordinate system be (u,
When v), spatial differentiation is performed on the original image in the u-axis direction and the v-axis direction, respectively. After differentiating for the entire image, it obtains a differential value d _u, the magnitude of the differential value of each pixel from d _v d and gradient direction theta.

[0006]

(Equation 1)

[0007] For all the pixels, the magnitude of the differential value of the pixel closest to the gradient direction θ of the eight adjacent pixels is compared with the magnitude of the differential value of the pixel in the opposite direction, and the magnitude of the differential value of the pixel is larger than both. At this time, the pixel is set as an edge point. Thereby, an edge image with a gradient direction is obtained.

[0008] [Hough transform] standard Hough transform is a coordinate transformation defined by equation (2), are represented the edge points in the edge image (u _i, v _i) in the parameter (theta, [rho) It is converted into a curve on the Hough plane.

[0009]

(Equation 2)

[0010] Here, ρ is the point from the image origin (u _i, v _i)
Is the length of the perpendicular drawn down to any straight line passing through
Is the angle of the perpendicular (vertical angle). When all the edge points are converted, the same number of curves as the edge points exist on the Hough plane. The points where many such curves intersect represent the linear components in the edge image. Applying a gradient direction theta _i of the edge points, the edge points (u _i, v _i) is Ho
It is converted to one point (θ _i , ρ _i ) on the ugh plane.

[0011]

(Equation 3)

According to equation (3), a line segment whose perpendicular angle is θ _i is converted to a point on the Hough plane. However,
The line segment given by the point sequence {(u, v)} includes a quantization error, and considering that the gradient direction θ _{i of} each point includes an error, the line segment given by the point sequence is on the Hough plane. Is converted to a region instead of a point.

Further, there are many errors that need to be tolerated in practical use, such as blurring of an image due to defocus or insufficient illuminance, a change in the amount of light in an environment, and disturbance of an edge line due to surface roughness of an object.

[0014]

SUMMARY OF THE INVENTION The present invention provides a line extraction method based on a floating ballet box, taking into account the above error, and an arc extraction method using the same method and a neural network. It shows. That is,
The present invention proposes a method in which a floating voting box is introduced on the Hough plane, and realizes stable and highly accurate line segment / arc extraction in consideration of reduction of processing time.

[0015]

According to a first aspect of the present invention, there is provided a line segment extracting method for spatially differentiating an image acquired by an image acquiring means as an original image. The size and the gradient direction are obtained, and the pixel whose differential value is larger than the differential value of the pixels on both sides along the gradient direction is defined as an edge point, and it is assumed that an edge point sequence composed of a set of these edge points forms a straight line, A line that determines that the edge point constitutes the straight line when the difference between the distance between the point closest to the edge point and the edge point on the straight line and the difference in the gradient direction are less than or equal to a certain allowable value. In the minute extraction method, the straight line is converted into a point on the Hough plane defined by two parameters of a length of a perpendicular line vertically dropped from the image origin and an angle of the perpendicular line, and the permissible angle is calculated around the one point. Phase to value To introduce the floating ballot box, the edge points when converted into the Hough plane enters said floating ballot box in the above edge point is characterized by determining to constitute the straight line.

According to a second aspect of the present invention, there is provided a method for extracting a circular arc, wherein an image acquired by an image acquiring means is spatially differentiated as an original image, and a magnitude and a gradient direction of a differential value of each pixel are obtained. When the magnitude of the differential value along the gradient direction is larger than both sides, the pixel is regarded as an edge point, and it is assumed that an edge point sequence composed of a set of these edge points forms an arc. When the distance and the azimuth angle are equal to or smaller than a predetermined allowable value, the edge point is determined to form the arc by an arc extraction method, wherein the arc is defined by a radius and an azimuth angle defined by two parameters. When a floating voting box corresponding to the above tolerance is introduced around a point on the plane, and the edge point is converted into a Hough plane and enters the floating voting box, the Point characterized by determining that constituting the arc.

According to a third aspect of the present invention, there is provided a method for extracting a circular arc according to the present invention, wherein the edge point sequence is input to a neural network previously learned about circular arcs / non-circular arcs. It is applied when it is determined that there is a possibility.

[0018]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The line segment / arc extraction method of the present invention comprises a line segment extraction method and an arc extraction method as described below. Note that, as described in the above-described related art, an image is acquired, and an edge point sequence is acquired by differentiating the image.

[Line Segment Extraction Method] (1) Introduction of Floating Ballot Box Consider a condition relating to a point sequence forming a line segment. FIG. 1 shows a straight line l on an edge image and points p _i constituting the straight line l. Assuming that l is a directed straight line, a direction obtained by adding π / 2 [rad] to the direction of l is defined as a normal direction. θ _l and ρ _l represent the normal direction of the straight line l and the distance from the origin of the image coordinates.

The point closest to the point p _i (u, v) on the straight line 1 is pε (uε, vε), and the distance between both points is δρ. The difference between the gradient directions θ _i and θ _l of the point p _i is defined as δθ. Point p _i is a straight line
The condition for judging that it is one point constituting l is expressed by the following equation.

[0021]

(Equation 4)

.DELTA.d and .DELTA..theta. Are constants given in advance as allowable deviations. Next, a range that satisfies equation (4) on the Hough plane is determined. The points p _i (u,
The Hough transform of v) is given by equation (2).
One point (θ, ρ) on the h plane is obtained. Here, ρ is represented by the following equation.

[0023]

(Equation 5)

[0024] (4), (5) than the conversion of the point p _i when the point p _i is assumed to be a point constituting a line l (theta,
The region of ρ) is shown in FIG. This area will be called a floating ballot box. Hereinafter, the floating ballot box for the straight line l is referred to as a ballot box l, and its center position is (θ _l , ρ _l ).

[0025] (2) Conversion of voting any point p _i to floating ballot box (theta, [rho) is the point where p _i constitutes a l by calculation whether or not to enter into the region of the ballot box l Can be determined. If the equations (4) and (5) are used as they are, the amount of calculation processing is large. Therefore, the area determination is performed using the equations (6) and (7) which are equivalent to the two equations. Ballot box l
Voting is satisfied when the expressions (6) and (7) are satisfied.

[0026]

(Equation 6)

[0027]

(Equation 7)

In the present invention, voting is performed twice. In the first voting, the purpose is to obtain an appropriate position of the ballot box. A ballot box is set according to the following algorithm 1, and the center position of each ballot is obtained based on the vote. In the second voting, a voting box is fixed at the position obtained in the first voting, and voting is performed according to Algorithm 2 described below.

[Algorithm 1] Search for ballot box It is determined whether or not there is a ballot box l _k (k = 1 to _j ) satisfying the equations (6) and (7) for the edge point p _i (θ _i , ρ _i ). (Step 1-1).

[0030] If there is no box that can vote to create p _i of the ballot box, ballot box l
_{j + 1} (θ _{lj + 1} , ρ _{lj + 1} ) is created (step 1-
2). At this time, θ _{lj + 1} = θ _i , ρ _{lj + 1} = ρ _i , and the number of votes is set to 0.

[0031] to vote vote p _i (step 1-3). First, the number of votes in the corresponding ballot box is increased by one. Then, the center position of the ballot box is recalculated.

Step 1-1 to Step 1 for all edge points
-3 is repeated (step 1-4).

[Algorithm 2] Searching for Ballot Box For the edge p _i , a ballot box satisfying the equations (6) and (7) is searched (step 2-1).

[0034] as soon as they are discovered ballot box l _j to satisfy the vote two equations, to vote p _i. The number of votes for l _j is increased by one, and the center of gravity of the votes is calculated (step 2-2).

Cancellation of Search If the voting to the ballot box is continuously performed, the voting to other ballot boxes is not performed (step 2-3).

Repeat Steps 2-1 to 2-3 are repeated for all edge points (step 2-4).

(3) Extraction of line segments by counting votes After two votings are completed, detection of straight line components and extraction of line segments are performed. Hereinafter, this operation is referred to as counting.

[0038]

(Equation 8)

[Arc Extraction] In order to properly detect an arc from a sequence of edge points, it is necessary to divide the point sequence into a portion to be recognized as an arc and a portion to be a line segment. Therefore,
Apply neural network and floating ballot box.

(1) Judgment of arc by neural network Neural network to be applied (hereinafter abbreviated as NN)
Has a simple structure as shown in FIG. 3 and inputs a normalized arc length η and a central angle ψ, and outputs an arc / non-arc judgment expressed by 1/0. A human observes an arc having a random length and a central angle, and creates data for judging an arc / non-arc.

The NN learns this data as a teaching rule by back vacation. The detection of the arc portion will be described below. First, a point sequence is traced starting from one end point of the edge point sequence. The point that follows is called a search point. The length from the starting point to the search point and the difference in the gradient direction between the starting point and the search point are input to the NN. When the output value is close to 1, it indicates that the point sequence portion from the starting point to the search point may be arcuate.

Next, arc fitting is performed to determine whether the point sequence portion is a circular arc. The center position and radius of the arc are estimated by least square approximation using the position information of each point.

A floating ballot box for judging the arc is introduced to check the deviation of the point sequence from the estimated arc, and it is judged from the voting result whether it is appropriate to recognize the point sequence as an arc. The installation and voting of the floating ballot box are performed in the following procedure. Estimated the center of the arc (a, b) the distance from to the points constituting the point sequence (u _i, v _i) and r _i, the azimuth angle from the center and phi _i. Next, an arc plane having axes of φ and r as shown in FIG. 4 is introduced, and a floating voting box whose center is (φ _i , r ₀ ) is installed.

The ballot box has (-Δθ, Δ
θ), (−Δr, Δr). (Φ _i , r _i ) is voted for each point in this ballot box.
If the ballot is made within the ballot box area, the ballot becomes a valid vote, that is, it is determined to be a point on the arc.

Based on the number of votes for the total number of edge points constituting the point sequence, it is determined whether it is appropriate to recognize the point sequence as a circular arc. The vote percentage is used as a criterion,
The reference value can be set arbitrarily. Since only the point sequence that may be an arc is detected by the NN, and the least square approximation is performed only once for the point sequence, the calculation processing cost required for the estimation is kept low.

(2) Arc Detection Algorithm FIG. 5 shows the flow of the arc detection algorithm. First, for one edge point sequence, the difference between the length from the starting point to the search point and the gradient direction is input to the NN according to the method described above.
When the output value is close to 0, if the search point has not reached the end point of the point sequence, the search is repeated.

When the end point is reached, it means that the point sequence does not include the arc component, and the search is terminated. When the output value of NN is close to 1, an arc is estimated by least squares approximation. Based on the estimated arc, an arc is determined using a floating ballot box. If it is recognized as a non-circular arc, the search ends.

When the point is recognized as an arc, if a point after the search point and a point before the search start point exist along the point sequence,
Voting is also performed on them, and it is determined whether they are points on the arc. If it is determined that the point is on an arc,
The arc including that point is estimated again. By such a procedure, an arc for the point sequence is detected. By executing this process for all point sequences, all arc components are detected.

[0049]

EXAMPLE FIGS. 6 and 7 show experimental results according to an example of the present invention.
Shown in FIG. 6 is an original image obtained by the CCD camera. The objects are an iron piece having a plurality of holes and two automobile parts. FIG. 7 shows the result of extracting line segments and arcs. The arc is indicated by a circle including the arc.

Neither a line segment nor an arc is recognized as a line segment or an arc if the length is not sufficient. In the case shown in the figure, 32 line segments and 20 arcs are extracted. The processing processor uses microSPARC II-110MHz for about 2
It took seconds.

[0051]

As described above, the present invention has proposed a method in which a floating ballot box is introduced on the Hough plane for extracting line segments and arcs which are important for improving the accuracy of object recognition using images. In particular, in the present invention, line segments and arcs are extracted on the assumption that the edge point sequence is displaced, and the amount of calculation processing and storage capacity are reduced. In the line segment extraction, we introduced a floating ballot box to reduce the effect of the position of the point sequence and the deviation of the gradient direction within a foreseeable range. In the arc extraction, the line segment and the arc were properly separated by applying a neural network, and the arc judgment based on the vote rate was performed by applying a floating ballot box. Extracts line segments and arcs stably without being greatly affected by the deviation of the edge sequence due to the ridge line disturbance due to the surface roughness of the target object or the image sequence of the original image I was able to.

[Brief description of the drawings]

FIG. 1 is a graph showing a relationship between an approximate straight line and an edge point.

FIG. 2 is an explanatory view showing a floating ballot box.

FIG. 3 is an explanatory diagram of a neural network for arc extraction.

FIG. 4 is a graph showing a floating ballot box on an arc plane.

FIG. 5 is an arc detection algorithm.

FIG. 6 is an explanatory diagram showing an original image acquired by a CCD camera.

FIG. 7 is an explanatory diagram of extracted line segments / arcs.

Claims (3)

[Claims] 1. An image acquired by an image acquiring means is spatially differentiated as an original image, a magnitude of a differential value of each pixel and a gradient direction are obtained, and a magnitude of the differential value is calculated from differential values of pixels on both sides along the gradient direction. Is assumed to be a pixel having a large edge point, and it is assumed that an edge point sequence consisting of a set of these edge points constitutes a straight line, and the difference in the distance between the point closest to the edge point and the edge point on the straight line and the gradient direction When the difference is less than or equal to a certain allowable value, in the extraction method of the line segment that determines that the edge point constitutes the straight line, the length of the perpendicular line that vertically descends the straight line from the image origin and the angle of the perpendicular line Is converted to a point on the Hough plane defined by the two parameters, and a floating voting box corresponding to the permissible value is introduced around the point, the edge point is converted to the Hough plane, and the floating voting is performed. Inside the box Wherein the edge point is determined to constitute the straight line when entering. 2. An image obtained by the image obtaining means is spatially differentiated as an original image, and the magnitude of the differential value of each pixel and the gradient direction are obtained. When the magnitude of the differential value along the gradient direction is larger than both sides, With the pixel as an edge point, assuming that an edge point sequence consisting of a set of these edge points forms an arc, when the distance and azimuth from the center of the arc to the edge point are equal to or less than a certain allowable value, In the method of extracting an arc, wherein the edge point determines that the arc is formed,
The arc is converted to a point on a Hough plane defined by two parameters of a radius and an azimuth, and a floating ballot box corresponding to the permissible value is introduced around the point, and the edge point is converted to a Hough plane. Wherein the edge point is determined to constitute the arc when the ball point is converted into the floating ballot box. 3. The method according to claim 2, wherein the sequence of edge points is input to a neural network previously learned about arcs / non-arcs, and is applied when it is determined that there is a possibility that the arc is an arc. Extraction method of circular arc.