JPS6232584A - Method for extracting contour shape feature - Google Patents

Abstract

PURPOSE:To upgrade the identification capability and to simplify respective feature extracting mechanism by executing an extraction using plural feature groups. CONSTITUTION:Two points (P1 and P9) are optionally selected. The furest point P9 among the each points (P1-P8) such as the points that the distances Di(i=1-8) form the two points between the straight line connecting the two points are below the fixed value (K1) (D1<=K; ik=1-8) is obtained, and the two points are connected directly. Such consecutive two points as whose circumscribed angles are of the same polarity are taken, and the distance between them is assumed as La, the maximum value among respective distances Di between the straight line connecting the two points and respective points existing the said two points is assumed as maxDi; when (Lad<=La<=Lau) and (Did<=dmaxDi<=Diu), the partial stroke between the two points is made as a curved primary feature. The circumscribed angle Ca as a curved primary feature is defined as the sum of external angles (alpha+beta+gamma+delta) between two straight line primary feature. The circumscribed angle X between the two straight line primary features CD and GH is obtained by X=D+Ca+G. Geometrical feature, primary feature group, higher primary feature group are arranged to accommodate the request by the object of identification, and then outputted.

Description

Detailed Description of the Invention (Technical Field to which the Invention Pertains) The present invention uses various methods to analyze the contours of an object in order to identify it as a two-dimensional image, thereby extracting the characteristics of the object. The present invention relates to a method for extracting contour shape features.

(Prior art) References regarding the prior art include ``Hidehiko Takano's Recognition Technology for Pa-shape Patterns'' by the Information Research Group.

Examples include 19g4.10J and ``Special Publication Showa 59-6420J.''

Object identification features in the conventional object identification method using a circular shape can be broadly classified into the following items.

Geometric feature group: Regarding the contour shape of an object converted into an x, y coordinate series and figures inscribed or circumscribed to this shape,
It is an amount obtained by measurement or an amount calculated from its various aspects.

For example, vertical and horizontal dimensions, center of gravity coordinates, area, perimeter, inscribed (
These are the various quantities (radius, area, side length, etc.) of a circumscribed (circumscribed) circle and an inscribed (circumscribed) quadrilateral.

Such a group of geometric features has the advantage of being resistant to noise, since it means that the figure is viewed from a macro perspective due to its statistical properties. However, there is a problem in the discrimination ability that it is weak in detailed discrimination.

2. A feature that focuses on the uneven state of the rim of the rim, and the rim is characterized by approximating the arrangement of elementary features such as straight lines, curves, and circular arcs.

This makes it possible to describe shapes in detail, and is expected to improve identification ability.

Next, we will explain the conventional extraction method of the contour line feature group (
(See Special Publication No. 59-6420).

Any part of the contour line is called a stroke, and three coordinate points are taken out by sequentially shifting one point at a time from the coordinate point series of one stroke, and the direction change of the two vectors given by each three points, and one vector A process is provided to detect the product of and the length of as the amount of change.

Then, the stroke is divided at a coordinate point corresponding to an amount of change larger than a predetermined value, and if there is no such coordinate point and the stroke is not a circular feature.

A process is provided to divide the stroke into predetermined lengths depending on the total length of the stroke to create partial strokes.

Thereafter, for each partial stroke, the ratio between the total amount of change and the total length of vectors is detected.

Then, by comparing the ratio with a value corresponding to the average length of the vector and detecting a magnitude relationship, it is determined whether it is a straight line or a single curve.

Then, each partial stroke is made to approximately correspond to a predetermined elementary characteristic shape of a circle, a curved line, and a linear shape based on information including the comparison result and the direction of the starting and ending points of the partial stroke.

Next, the detection of elementary features and the procedure for detecting elementary features will be explained using the drawings.

FIG. 1 is an explanatory diagram of partial strokes of a contour line, and is a diagram showing parts P0 to P5 (partial strokes) of a contour line.

When connecting P0 to P as shown in the figure, it is determined whether the line segment P0 to PS is a straight line or a curve using the following method.

FIG. 2 is an explanatory diagram of quantization coefficients.

Vector P. The direction change of the vector p1p with respect to P□ is divided by 16 as shown in Figure 2, and the quantization coefficient (
0-7. -1 to -8), and the length of P1P2 is L, the amount of change d in PiP2 is, dx = alX L
, (al is a quantization coefficient), and this is defined as each vector P, P□, P, P4 . Find P and P.

Based on dt*dx +d31d4 obtained in this way, o = (dt +d, +d, +d4) / (
L+Lz"La"L4) is calculated and compared with a constant value.

As L + L, +L, ÷L, =L.

When L≧T11°, if +01<TH□, it is a straight line.

TH, TH is a constant value determined by the length.

FIG. 3 is an 8-direction classification diagram.

Regarding straight lines and curves determined in this way.

Connect the starting point and the ending point, classify the starting and ending point directions into eight directions as shown in Figure 3, and give code numbers as shown in Figure 4 to straight lines, curves, and circles that fall in each direction. .

FIG. 4 is an explanatory diagram of code numbers.

Straight lines, curved lines, etc. indicated by these code numbers are called a search feature group.

Next, the procedure for detecting elementary features will be explained below.

1) d0 calculated within one stroke of coordinate points P0 to Pn
Check whether there is a point I]1 within ~dn that satisfies 1d11≧Tll. T11. is a constant value determined by the average value of Ll.

2) If Pl exists, divide the stroke by Pl, and
Elementary features are detected for each partial stroke of −P, , Pl, and ~P, .

3) If there is no P, check whether one stroke from P0 to pH is a circle, and if a curve is detected in 2, check whether it is a circle.

The conditions are TH<T11□≦D, Q≦TH4.

Here, TH□ and TH are constant values determined by the average value of Ll, etc.

FIG. 5 is an explanatory diagram showing the definition of a circular arc.

Q is the line segment length connecting Px and PX+G in Figure 5.
Then, Q"y/(L, +LX, 1+...""Lx'
Calculate in 5).

4) If the curve detected in 2 is not a circle, register it as a curve. If Pl is not detected, L from the starting point
When adding L≧TH0, IDI≧TH
1 to determine whether it is a straight line or a curve, and detect and register the elementary feature code based on the direction of the start and end points.

Although such a group of contour features can describe a detailed shape and can be expected to achieve high discrimination ability, it has the drawback of being susceptible to noise.

FIG. 6 is a diagram illustrating the drawbacks of the conventional method.

As a simple example, as shown in FIGS. 6(a) and 6(b), similar portions caused by noise turn into meaningful features as long as the above conditions are satisfied.

FIG. 6(b) shows the original shape, consisting of parts (A) and (B).

On the other hand, in Fig. 6(a), noise is superimposed on part (A) and it is deformed, so that part (B) satisfies the same conditions as part (B) and appears to be two parts.
This example shows how to extract features such that there are multiple features.

FIGS. 6(C) and 6(d) show an example in which clearly different shapes are considered the same due to quantization error as long as they satisfy the conditions.

As mentioned above, geometric features are strong against noise, but have a problem with their discrimination ability because they are weak against detailed discrimination.Contour feature groups can be described in detail and can be expected to achieve high discrimination ability. , which had the disadvantage of being susceptible to plane noise.

(Object of the Invention) The object of the present invention is to improve the drawbacks of the above-mentioned conventional techniques, and to extract both contour line features and geometric features.

(Structure of the Invention) In the process of extracting contour features, the main feature of the present invention is to simultaneously extract the geometric features thereof, and feature extraction is performed using a plurality of feature groups. The difference is that the performance can be increased and furthermore, each feature extraction mechanism can be simplified.

FIG. 7 is a schematic flow diagram of a method for extracting contour shape features according to the present invention.

1 is a contour shape input unit; 2 is a compression unit that removes noise, compresses the amount of data, and extracts line element features for the input coordinate point sequence of the contour shape; 3 is a circuit that uses the contour shape input and its compression results. 4 is a geometric feature extraction unit for extracting geometric features of the contour line, and 4 is an elemental feature extraction unit for extracting curve element features from the compressed contour figure. Depending on whether the stroke satisfies the condition of being a curved feature, it is determined and extracted whether it is a curved feature or a straight line feature. 5 is a high-order contour line feature extraction unit that extracts a higher-order feature group using the elementary feature group and knowledge about the object; 6 is a high-order contour line feature extraction unit that extracts a higher-order feature group using the elementary feature group and knowledge about the object; 6 is a high-order contour line feature group, a high-order contour line feature group, and geometric features. This is a feature output unit for arranging and outputting groups as their attribute values.

Next, a procedure for detecting contour line features and geometric features will be described using the drawings.

A contour shape extracted by appropriate image processing such as binarization processing from an image of the object to be recognized captured by an ITV camera or the like is input as a series of X and Y coordinate points.

The input data is compressed to facilitate subsequent analysis, remove unnecessary portions of contour lines (noise removal), and prepare for extraction of line element features.

FIG. 8 is an explanatory diagram of the linear approximation method.

That is, as shown in Fig. 8, any two points (P work, P9)
, and each point (distance o from px to pj, (i=t to 8) between the straight line connecting these two points, (i=t to 8) is less than or equal to a certain value (K,) (0+≦,; Find the farthest point (P, ) that satisfies (i=1 to 8) and connect these two points directly.

The K1 value is determined by the overall size of the figure and the characteristics of the curved feature that will be cursed thereafter. As a result, the initially input sequence of N coordinate points is compressed into a sequence of M (M≦N) coordinate points.

In this processing process, the vertical and horizontal dimensions, area, perimeter, and various quantities calculated from these are calculated as a group of geometric features from the N-gon or M-gon and the figures inscribed (circumscribed) to these. Extract. The compressed data is input to the elementary feature extraction section.

As elementary features, only a line element (L), a curve element (C), and an angle element (A, B) are used to simplify processing and provide flexibility. Since the preprocessing for straight lines has already been completed during the compression process, here it is sufficient to simply determine whether the line is a curved line or not. That is, partial strokes that could not become curved features are replaced with a plurality of straight line features.

FIG. 9 is a definition of each elemental feature and an explanatory diagram thereof.

FIGS. 9(a) and 9(b) show the straight line feature (1,) and the angular element features (A, B).

Length is added as an attribute to the straight line feature.

An angle elementary feature is a circumscribed angle between two adjacent elementary features,
As shown in Figure 9(a), the rightward angle is defined as positive polarity, and B
, and add the size of the angle as its attribute.

The first condition for defining a characteristic endpoint includes the polarity of the circumscribed angle and the size of the angle.

FIG. 9(c) shows the definition of curved features. That is, between any two consecutive points whose circumscribed angles have the same polarity, the distance between the two points is La, and the maximum value of the distance Di from the straight line connecting the two points to each point between the two points is max D. i, La
, maxI) i simultaneously satisfy the second condition and the third condition, that is, (Lad:iLa≦Lau) & (Did≦d maxD
If i≦Diu), the partial stroke between two points is a curved feature (Lad.

The values of Lau, Did, and Diu are constant values determined by the size of the figure, etc.).

The attributes of this curved feature include the center coordinates between two points as position information and the circumscribed angle Ca as the curved feature.

The circumscribed angle of a curved feature is defined as the sum of the external angles between straight line features that make up the curved feature.

That is, in FIG. 9(c), Ca = α + β + γ + δ
becomes.

By using this Ca, two
It is possible to find the circumscribed angle between linear features.

In FIG. 9(d), the circumscribed angle X with the two-line element features CD and GH is determined as X=D+Ca+G.

The higher-order elementary feature extraction unit extracts new features from the array of extracted elementary feature groups using knowledge about the object as a fourth condition. In this case, not only each elementary feature group but also geometric features are used as necessary.

The feature output unit organizes and outputs the geometric feature group, search feature group, and higher-order elementary feature group obtained so far into a form that meets the requirements of the identification target.

(Example) Examples will be described in detail below with reference to the drawings.

FIG. 10 is a diagram showing an embodiment of the compression section.

This is a method of searching for an arbitrarily determined starting point and the farthest point from the starting point, and using these two points as starting points to perform a straight line approximation using the procedure shown below.

In Fig. 10, input data is shown as a broken line, straight line approximation results are shown as a thick solid line, and lowercase letters a-n
indicates the order in which the vertices are determined, in which (i) with parentheses
~(n) shows only the results.

Steps for straight line approximation ■ Decide on the starting point.

■ Find the farthest point from the starting point and use it as the ending point.

■ The point with the maximum distance from the straight line connecting the starting point and ending point (
Pl) is obtained in the analysis direction up to the end point.

If this distance is an acceptable value, go to ■, otherwise go to ■.

(◇ Since the starting point and ending point can be approximated by a straight line, use the 21st point as the new starting point and go to ■.

■ Record the end point on the stack and set the P1 point as the new end point.

■ The point with the maximum distance from the straight line connecting the starting point and ending point (
Pl) is obtained in the analytical direction up to the end point. If the distance is an acceptable value, go to ■, otherwise go to ■.

■ Since the starting point and ending point can be approximated by a straight line, the P1 point is set as the new starting point.

■ Check inside the stack, and if it is "empty", go to ■.

If it is not "empty", pop up the contents of the stack (take out one item), use that point as the new end point, and go to ■.

■ The above process is repeated until the starting point goes around the closed curve, and by sequentially connecting the starting points, a straight line approximated curve is obtained.

FIG. 11 is an explanatory diagram of a geometric feature group and its extraction method.

In FIG. 11, the hatched area is the target figure. Here, the aspect ratio, perimeter, area, center of gravity position, etc. are shown as geometric features.

Aspect Ratio> When a circumscribed rectangle of input data is found and the vertical and horizontal side lengths are Y enm n X In&, the aspect ratio is defined as Y, n, /x, n1.

The aspect ratio has a normalizing effect on the size of a figure, and can be applied to a wide range of applications. Of course, Y,...,X. , both become geometric features as they are.

<Perimeter length> This is a contour line feature, and the perimeter of an N-gon is determined by the following formula.

Perimeter = Σ5QrtT((X,,i-X,)2+(Y
,,1-Y,)") However, X, , Y are the X and Y coordinate values, respectively, of the vertices of the polygon, and Xn+, = x
x + Yn*x = Yl (SQRT represents square root).

<Area> This is the area surrounded by the contour line, and the area of the N-gon is determined by the following formula.

Vertex coordinates are (Xi, YL), (X2.Y2), ・・
, (Xn, Y,) throat, area = 1/211 ((x
, -x,)-(Y, +Y,)+(x,-X3) Umbrella (Y
,+Y,)+-...””Cxn-xx)Fist(yn”
yj month However, the rate indicates the multiplication symbol.

Complexity: The complexity of a figure is expressed as the following quantity.

Complexity = (Perimeter) 2 / Area <Position of center of gravity> Center coordinate values of figure (Gx, Gy) 1st
FIG. 2 is an explanatory diagram of extraction of elementary features of contour lines.

As a candidate for the starting point in extracting curve element features, a feature end point (P1
~ pg), and the feature end point adjacent to this in a clockwise direction is the end point (for example, P2 etc. for P construction).

If the distance between the starting point and the ending point does not satisfy the conditions of the curve element feature, the ending point is shifted by one position toward the starting point (in the case of pz, Q
, ) and repeat until the start and end points are adjacent.

If the start point and end point are adjacent, the line element feature (P
□~Q) and repeat the above operation using the end point (Ql) at this time as the new starting point.

The first curve element features in Fig. 12 are 01 to 06.
6. The attributes are calculated for each elemental feature, and in the case of a curved elemental feature, the attributes are the circumscribed angle of the curve normalized by a constant angle value A0 and the center coordinates of the curve.

The attribute for the straight line element feature is a length normalized by a constant length L0.

In the above elementary feature extraction process, there is a 9th point between two elementary features.
Insert the angle feature defined in the diagram.

FIG. 13 shows an example of the elementary feature extraction results.

In FIG. 13, the following numerical values for L, C, L, and B indicate the respective normalized lengths and normalized angles.

The first alphabet (L, C, A, B) and outline of the contour line are cursed, and the detailed shape can be calculated using the attributes indicated by the next numerical value. Furthermore, the positional relationship of the feature points can be determined from the center coordinates of the one-curve element feature.

FIG. 14 shows an example of the configuration of the high-order elemental feature extraction section.

11 is the elementary feature group or the (n-1)th iodine feature group, 12 is the n-th blown feature extraction section, and the dictionary section 13 which is the fourth condition.
and an extraction section 14 that extracts the nth blowing feature from the (n-1)th blowing feature while referring to the dictionary section, and 15 is a high hydrogen feature group.

FIG. 15 shows an example of the operation of the high-order elemental feature extraction section.

In Fig. 15, the input (a) is the elementary feature sequence shown in Fig. 13, and (b) is an example dictionary that is the fourth condition indicating knowledge of the subject, and the input element sequence is the sequence of the input elementary feature sequence. This indicates that a new feature (F, N) is extracted when a certain condition is met. (C) focuses mainly on F and N,
If the interval between F and N is large (larger than a constant value Lw, set W within a certain range (larger than a constant value L, Lw
(smaller than ), this is an example of the first hydrogen feature group extracted by adding each length of L as an attribute. Note that the center coordinates of F and N are added as attributes.

FIG. 16 is an example of a higher-order elemental feature, and is an example in which only F and N are illustrated (also referred to as the second hydrogen feature group).

In the feature output section, geometric features, elementary features of contour lines, and the first
The blowing feature group, . . . , the n-th blowing feature group, etc. are organized and output.

(Effects of the Invention) As explained above, in the present invention, the statistical quantities and structural features related to the object are effectively connected with the knowledge about the object to output the characteristics, so high discrimination ability can be expected. Specially possible.

In addition, each feature group can be simple relative to each other, and compared to conventional methods that aim for complete identification using only one feature group, the feature extraction section is simpler, or the feature extraction section can be operated in parallel. It has the advantage of being able to perform high-speed feature extraction operations.

The field of application of the present invention is not only systems related to objects with relatively fixed shapes such as industrial products, but also systems related to objects that are curved and flexible, such as human body parts (such as hands and faces). , an identification system for objects that move or deform and for which it is generally difficult to set a reference point.

[Brief explanation of drawings]

Figure 1 is an explanatory diagram of partial strokes of contour lines, Figure 2 is an explanatory diagram of quantization coefficients, Figure 3 is an 8-direction classification diagram, Figure 4 is an explanatory diagram of code numbers, and Figure 5 is a definition of circular arcs. An explanatory diagram showing
Fig. 6 is a diagram illustrating the drawbacks of the conventional method, Fig. 7 is a schematic flow diagram of the method for extracting contour shape features according to the present invention, Fig. 8 is an explanatory diagram of the linear approximation method, and Fig. 9 is a diagram of each elementary feature. Definition and its explanatory diagram, Fig. 10 is a diagram showing an example of the compression section, Fig. 11 is an explanatory diagram of a geometric feature group and its extraction method, Fig. 12 is an explanatory diagram of extraction of elementary features of contour lines, FIG. 13 shows an example of elementary feature extraction results, FIG. 14 shows an example of the configuration of the high hydrogen feature extraction section, FIG. 15 shows an example of the operation of the high hydrogen feature extraction section, and FIG. 16 shows an example of high-order elementary features. 1...Ring shape input section, 2...Compression section, 3...
Geometric feature extraction unit, 4...Elementary feature extraction unit, 5...High hydrogen feature extraction unit, 6...Feature output unit, 11...Elementary feature group or (n-1)th iodine feature group, 12
. . . nth blowing feature extraction section, 13. . . dictionary section, 14.
... Extraction unit, 15... Feature output. Patent applicant: Nippon Telegraph and Telephone Corporation Figure 1 Figure 2 Figure 3 Figure 4 Figure 6 (c) (d) Figure 7 Figure 8 Heel≦+ (Le=1 to 8) O ) Kl Figure 9 Figure 10! + Ri[#<Kakuaki Figure 13 (B1) Figure 14

Claims (2)

[Claims] (1) In recognizing an object based on a circular shape, information on the pattern of the circular shape to be recognized is input as a series of N coordinate points, and the N coordinate point series is converted into a series of M (M≦N) coordinate points. , calculates each apex angle of an M-gon whose vertices are M coordinate points, and extracts geometric features from the M-gon and figures inscribed or circumscribed to the M-gon. If the vertex angle satisfies the first condition arbitrarily set by Defining the distance between the feature end points as a straight line feature,
Calculate a predetermined attribute value, and connect the feature end points once at least one vertex exists between any adjacent feature end points and the distance between the feature end points satisfies the arbitrarily set second condition. When the distance from the straight line to each vertex between the feature end points satisfies the arbitrarily set third condition, the distance between the feature end points is defined as a curve element feature,
After calculating the predetermined attributes, if the second condition and the third condition are not satisfied at the same time, the vertex that exists between the two feature end points and is adjacent to one feature end point is set as a new feature end point, and the straight line element feature is Alternatively, the feature is to recursively extract curve element features, go around the M-gon from an arbitrary feature end point as a starting point, and replace each part with a predetermined element feature or a series thereof to form a contour line feature. A method for extracting contour shape features. (2) In the elementary feature extraction process, by inputting the elementary feature or its series extracted at least once, a new elementary feature or its series is generated corresponding to another feature group based on the fourth condition specified elsewhere. A contour shape feature extraction method according to claim (1), characterized in that the contour shape feature extraction method is output.