JPS6228888A - Method for discriminating object - Google Patents

Abstract

PURPOSE:To discriminate an object even if a shape difference between patterns to be discriminated is not enough large by executing the 2nd collating processing for collating the segment information of all segments of a proposed reference object with that of a unknown object. CONSTITUTION:A processor 17 traces the outline of an object, patterns the outline and searches a feature point on the linear pattern. Segmenting processing for dividing the linear pattern into plural segments while using the feature point as a starting point is executed and segment information including data inherent in each segment and inter-segment connection/array data is extracted in each segment. In the 1st collating stage, objects having the same feature point are searched out of the object to be inspected and plural reference objects and the similarity for the prescribed number of segments is found out on the basis of the feature point to select a proposed reference object. In the 2nd collating stage, all segments of only the proposed reference objects are collated and the reference object having the highest similarity is decided as the reference object for the unknown object.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to an identification method for identifying (inspecting) an object by converting it into a line diagram using an image processing device.

[Conventional technology]

Recently, the need for object identification has increased in various industrial fields. For example, in order to perform automatic assembly of a product, the component parts are identified and/or processed. Conventionally, as such an object identification method, the so-called 5FtI (Stanford Re5e
Arch In5tmutute) method, matching method, etc. are known.

The matching method is a method of extracting all data related to the entire or part of an identification object pattern and comparing it with a standard pattern for identification.

However, this method has b) It is difficult to apply this technique if the posture of the object is not good.

(Example) It is impossible to identify objects unless they are separated from each other.

Not only are there problems and constraints, but the amount of data becomes enormous because the object pattern itself is handled, and the larger the posture variation of the object, the more complex and difficult the position normalization and inquiry processing become.

On the other hand, the SRI method uses the following method.

FIG. 7 is a diagnostic diagram for explaining the 5FtI method.

This involves first encoding a binary image of the object using run lengths (segments), and then performing connectivity analysis on the run length data to extract various feature quantities. Then, many useful parameters are calculated based on the basic % trace amount and its combination. There are many basic features, such as the total area S and perimeter L1 of the object 10 as shown in FIG. 7, the number of holes H in the object 10, the sum of the areas of each hole H, and the 10 center of gravity P. (xo, yc), the maximum value RmaX% of the radius from the center of gravity to the contour, and the minimum value Rmi, ,
Maximum value Xmax and minimum value Xm1ns projected onto the X axis
There is a maximum value YmaX and a minimum value Ymin projected onto the Y axis. Further, as combination parameters of basic feature amounts, for example, Pe (-=4πS/L2), Rr ('Rma
x Rmin), Xa (=Xmax Xm1n
).

Yd (=Ymax Ymin), etc. Note that the identification process is performed in advance by showing each object.
After registering it by showing it to an image processing device),
This is done by comparing and inquiring the parameters of the object to be identified.

[Problem that the invention seeks to solve]

As described above, the SRI method is a method that focuses on the geometric features of the object, but it has the following problems.

a) Identification cannot be made unless the difference in shape between the patterns to be identified is large.

b) The objects to be identified cannot be identified unless they are separated from each other.

In other words, a) is a constraint on the type of objects that can be identified and the complexity of their shapes, and b) is a constraint on the arrangement and background of each object, which requires that the objects do not overlap or combine with each other. The condition is that no other patterns exist in the background.

Therefore, the present invention has the following features: 1) Even if the difference in shape between patterns to be identified is not too thick, it can be handled.

2) It is possible to handle not only the objects to be identified that are separated from each other but also those that overlap.

3) Improve identification accuracy and shorten identification time (save shape features and compress data).

4) In particular, perform query processing efficiently and improve its accuracy.

The purpose is to

[Means for solving problems]

An image processing device is provided which captures an object as an image, extracts each of the eight features of the object, particularly its outline, from the image information, and performs predetermined processing based on this.

[Effect]

The image processing device performs the following processing.

1) Trace the contour of the object and convert it into a line figure.

2) Search for characteristic points (feature points) on the line figure.

These characteristic points include break points and uneven points where the curvature changes significantly.

3) Perform line segmentation processing to divide the line figure into a plurality of line segments using the feature point as a base point, and extract line segment information including line segment-specific data and line segment connection/arrangement data for each line segment.

4) In the dictionary registration mode, the above-described processing is performed for each standard object, and various information extracted as a result is stored in a predetermined memory.

5) In the inspection mode, the same process is performed for the unknown object, and various extracted information is stored in a predetermined memory.

6) In the wJ1 stage of matching, the inspection object and a plurality of standard objects are found to have the same feature points, and then a predetermined number of line segments (sampled line segments) are calculated based on this point.
The degree of similarity is determined for each object, and the candidates for the standard object are narrowed down based on this.

7) In the second stage of matching, matching is performed across all line segments only for the above standard candidates, and as a result, the standard object that gives the greatest degree of similarity is determined to be the corresponding object of the unknown object.

〔Example〕

Figure 1 is a reference diagram for explaining the invention in detail;
Figure 3 is a block diagram showing an image processing apparatus in which the present invention is implemented, Figure 3 is a reference diagram showing an example of storing line segment information obtained by the present invention in memory, and Figure 4 is a contour line consisting of a series of pixels. FIG. 5 is a reference diagram for explaining the direction code, and FIG. 6 is a reference diagram showing the storage fIJ of contour information in the memory.

First, explanation will be given starting from FIG. In the figure, 10 is an object to be identified, 11 is an illuminator, 12 is an imaging device such as a television camera, 16 is an analog/digital (A/D') converter, 14 is a preprocessing circuit, 15 is a feature extraction circuit, 16 is an image memory, and 17 is a processing device such as a microprocessor.

The optical image of the object to be identified 10 illuminated by the illuminator 11 is converted into an image (video) signal by the television camera 12. This video signal is converted into a digital quantity for each pixel by an A/D converter 15, and further processed by a preprocessing circuit 14.
Noise removal, smoothing, etc. are performed on the image, and the image is extracted as a digital image. The feature extraction circuit 15 extracts various feature quantities from this digital image and stores them in the image memory 16. The processing device 17 performs connectivity analysis by 4-connection or 8-connection of 4 or 8 pixels adjacent to each pixel of interest, and extracts each prop (B job ). A prop represents an independent pattern and is separated from the background.

Next, the contour of each prop is extracted and contour tracing is performed. Information regarding this contour line is stored in the image memory 16.

Now, assuming that a part of the contour line is represented as a series of pixels as shown in FIG. 4, this contour line 1 corresponds to the coordinate position P1 of each pixel. It can be expressed by P2...P9 and direction codes Do to D7 at each coordinate point.

Here, the direction code is 0, which tracks the contour line pixel by pixel.
It represents the direction (vector) of the
There are eight directions as shown by o=D7. Note that the angle between each direction vector is 45°. Further, these contour data are stored in the image memory 16 as shown in FIG.

This direction code can be used in various ways; for example, it can be used to calculate the perimeter of a contour line. In other words, the perimeter from point 21 to point P9 in Figure 4 is 0', 90°, 180°, with direction code Do as Oo.
°, 270”, each has a length t for one pixel,
Directional codes are 45° and 135°.

In the case of 225° and 15°, it is considered that the length t of one pixel is multiplied by V2, and it can be determined as the phase of each part. Therefore, the perimeter from point 21 to point P9 in FIG. 4 is c 3X i+sxv'i > x t.

In this way, the image pattern of the object is converted into a line figure using contour lines. Description (representation) of the object using this contour line
allows for significant data compression while containing sufficient shape information and feature information of the object. In addition to the method of using a television camera as described above, methods of using a flying spot scanner or laser light may also be considered for tracing the contour line.

Next, line segmentation is performed to divide the contour line.

Although there are many possible methods for line segmentation, an example according to the present invention is shown in FIG.

First, a starting point (for example, P5<X5rY5) is set on the contour line as shown by the symbol IVC in FIG. Next, the tangent line 2 at this point P3 is calculated using the following formula.
It represents the inclination (gradient) of , and is determined by an appropriate method. Then, a point P4 on the contour line where the error θ3 between this 112 and the contour line reaches a predetermined punch capacity value is set as the end point of one line segment. This P4 ((X4 * Y4) point is 1
It ends at the end point of one line segment L3, but at the same time it also marks the start point of the next line segment L4, and in the same way, line segment L 4r L
s... can be obtained.

The starting point for line differentiation is determined as follows.

Trace the outline of a line figure and look for points that are characteristic of it (feature points), such as break points or uneven points. In FIG. 1(a), the break point P11P2 is reached at this point. In other words, if we examine the change in the curvature (reciprocal of the radius) during the process of tracing contour line 1, we will find that the curvature changes significantly at point P 1r P 2 as shown in Figure 1 (b), so we consider this as a feature point. can be extracted. Therefore, by performing line segmentation based on the feature points P1+P2, line segments L11L2 . . . as shown in FIG. 1(a) are obtained. At this time, the interior angle of the break point is also determined, and P1+P2
Make sure the points are distinguishable from each other.

Next, line segments L1pL2... are extracted, for example, L 1 y LS * LS e L
7 (extracted line segments), these are described using line segment information. This refers to line segment information, ii) line segment-specific information, iii) connection and arrangement information for each line segment. More specifically, a) is the label of the line segment, the length of the line segment, the coordinates of both endpoints of the line segment, etc.

, the connection order of each line segment, the angle between two line segments, the distance between the representative points (end points, center points, etc.) of each line segment, etc.

It is.

This line segment information is also stored in a predetermined memory, an example of which is shown in FIG. It should be noted that all of the data in (a) 2) above are not necessary (for example, the connection order in (b) labels and ports) can be omitted by associating them with addresses in the memory. In addition, the coordinates of both end points in (a) are a set of information, ranging from the line segment itself to information (specific information).
In addition, it also serves as information representing the connection relationship with adjacent line segments, and is effective as original data before normalizing the position of the target figure. By using the angle θa between the adjacent line segments of the mouth), the amount of stored data will be reduced and processing will be easier, but this is because the angle θ3 is normalized data that does not depend on the orientation of the target figure. It is. This also applies to the length of the line segment in (a).

By converting the object into a linear figure and dividing it into lines in this way,
It becomes possible to significantly compress target object data.

First, the identification process will be explained.

Before unknown object identification processing (also referred to as inspection mode) K, all standard objects are first learned and registered (also referred to as dictionary registration mode). At this time, each object is imaged in a state separated from each other (isolated) by the width of the background with good contrast, and each object is subjected to the line drawing processing as described above to extract feature points. (including coordinate position, interior angle size, etc.),
The feature point information and line segment information are extracted and stored in a predetermined memory.

Next, bile analysis of the test object and each standard object is performed in two stages as follows.

In the first stage, feature points and sampled line segments (Ll in Figure 1 (a)
p L3 * Ls + Ly ) Focus on K. That is, a standard object having the same feature points as the object to be inspected is first found, and then information on each line segment of the sampled line segment is examined. Here, if the line segment information regarding one line segment is expressed by f1 for the standard object and f' for the inspection object, then the similarity S between the two can be considered, for example, as shown in the following equation.

S-(f-1f-f'l)/f Therefore, if both feature quantities f and f' match (f'
-f), S-1. Similarly, the Knth similarity is Sn, and wn is the weighting coefficient of the nth line segment (percentage of the whole)
Then, the similarity S for all line segments (number of line segments; N)
t can be expressed as sl-Σ snxwn n;1.

By the way, although the final target is the similarity St, the first
At this stage, the degree of similarity Sp as shown in the following equation is calculated only for the sampling line segments of the inspection object and the standard object.

Sp7Σ S I X w H 1 sieve j, 3, 5.7 -4Σ
f11!1.3,5.7 Then, a predetermined tolerance value (value close to 1) f: is set for this degree of similarity Sp, and standard objects exceeding this are selected as candidate objects 7. In this way, in the first stage, efficient processing is performed to narrow down the candidates.

Next, in the second stage, only the candidates selected in the first stage are matched against all line segments, that is, the above-mentioned similarity St is calculated, and a standard object that gives the largest similarity with the unknown object is selected. Determine the object as an unknown object.

In this way, the efficiency and precision of verification are improved.

〔Effect of the invention〕

According to this invention, by extracting the contour information of an object and treating the object as a line figure made up of multiple line segments, it is possible to perform significant data compression while sufficiently extracting the shape characteristics of the object. This provides the advantage of reducing capacity and computation time. In addition, we have extracted characteristic breaking points and uneven points (feature points) on line figures and used these as standards for line segmentation, thereby reducing the instability of line segment information due to variations in the reference points. It is possible to obtain stable line segment information that has been lost. Furthermore, in the query process, we narrow down the standard candidates to those that have the same feature points and have similar sampled line segments, and then query only these standard candidates using all line segments, increasing the efficiency of matching. and its accuracy can be improved.

Therefore, this invention b) The shape of the object to be identified is not simple and there are many types.

Not only are the objects to be identified separated individually, but there are also cases where they are superimposed (partial Bakun).

It is also possible to deal with such cases. In other words, needs such as (a) 1) above are increasing and becoming stronger with the reverse layer of various automations (FA, OA, LA), but conventional SRI methods and matching methods cannot meet these needs. Therefore, the present invention has been designed to meet these needs.

[Brief explanation of drawings]

Figure 1 is a reference diagram for explaining the invention in detail;
FIG. 6 is a block diagram showing the configuration of an image processing device in which the present invention is implemented. FIG. 6 is a reference diagram showing an example of storing line segment information in memory obtained by the present invention. FIG. A reference diagram showing an example of a contour line, FIG. 5 is a reference diagram for explaining the direction code, FIG. 6 is a reference diagram showing an example of storing contour information in memory, and FIG. 7 is a reference diagram for explaining the SRI method. FIG. Code explanation 1... Contour line, 2... Tangent line, 10...
...Identification object, 11...Illuminator, 12.
... TV camera, 13 ... A/D converter, 14 ... Preprocessing circuit, 15 ... Feature extraction circuit, 16 ... Image memory, 17...
...processing device. Agent Patent Attorney Akio Namiki Agent Patent Attorney Kiyoshi Matsuzaki II Figure P8 Figure 2 Figure 1m! 3 Figure! 41i11 1F5 Figure θ2 M 6 Figure 7”I

Claims (1)

[Claims] A method of identifying the object by capturing an image of the object, processing the image, converting the object into a line figure, and performing predetermined processing based on the line figure, the method being a method of identifying the object, the feature of which is on the line figure. (feature point) is first extracted, the line figure is divided into multiple divisions using this point as a starting point, and line segment information including data specific to the line segment and line segment connection/arrangement data is extracted for each line segment. After performing the extraction process on each standard object, the same process is performed on the unknown object to extract its line segment information, and among these standard objects, those that have the same feature points as the unknown object are extracted. In addition, a first matching process selects as a standard candidate a predetermined one whose line segment information is similar among all the line segments forming each line figure, and a first matching process that selects as a standard candidate the line segment information of all the line segments for the standard candidate. A method for identifying an object, characterized in that a second matching process is performed to match that of the object, and as a result, a predetermined standard that most closely resembles the object is determined as a corresponding object of the unknown object.