JPH0525354B2 - - Google Patents

Description

[Detailed description of the invention]

[Industrial Application Field] The present invention relates to a method of capturing an image of the shape of a target object using an iTV camera or the like, and recognizing and classifying the shape. [Prior art] As a typical shape recognition device of this type,
The Vision Module developed by SRi (Stanford Research Institute) in the United States is widely known. this is,
It was mainly developed as an eye for robots.
The shape of the target object is imaged with a television camera,
After converting into a binary or multivalued pattern using A/D conversion, the outline of the target shape is extracted using connectivity analysis, and the feature values of the shape (area, perimeter, shape coefficient, etc.) are measured. By comparing this feature amount with the feature amount of a standard pattern that has been trained in advance, shape identification and classification (i.e.,
Which of the pre-memorized shapes does the shape belong to, which glass does it belong to, etc. ), and the position and direction of its shape within the field of view, and provides information for robot control, etc. FIG. 7 is a schematic diagram showing a general shape recognition device. In the figure, 11 is an input device including a television camera, 12 is a preprocessing circuit, 13 is a shape extraction circuit, 14 is a connectivity analysis section, 15 is a feature calculation section, 16 is a shape recognition section, and 17 is a standard The pattern storage section 18 is an output device. An image (video) signal of the object shape captured by a television camera or the like is sent to an A/D in a preprocessing circuit 12.
After being subjected to processing such as conversion and noise removal and converted into a digital image signal, the shape cutting circuit 13 converts it into a binary or multivalued pattern using a predetermined threshold value. In the following explanation, a case where the pattern is converted into a binary pattern will be described, but it is not necessarily limited to binary patterns. Next, the binary pattern of the target shape is input to an arithmetic processing device such as a microcomputer, the connectivity analysis section 14 extracts and separates the outline of the shape, and the feature amount calculation section 15 calculates the area, perimeter, etc. of the shape. Shape identification parameters such as the number of holes, hole area, shape factor, etc. are calculated, as well as position/direction identification parameters such as the center of gravity position, moment of inertia, etc. Generally, the target shape is identified by the shape identification unit 16 using a standard pattern of shape identification parameters that is created in advance by a shape teaching operation called showing (teaching) and stored in the storage unit 17. In this shooting, a standard pattern of shape identification parameters is created by repeatedly imaging the shape to be identified in advance, calculating the shape identification parameters, and performing statistical processing, etc., and stored in the memory 17. . In addition, the shape identification unit 16 uses a decision tree (Binary
By comparing the shape identification parameters of an unknown shape with the standard pattern of shape identification parameters memorized by shooting, it is possible to determine which of the shapes memorized by shooting is using the decision tree. In addition, in FIG. 8, the feature amounts are up to . [Problems to be Solved by the Invention] Now, in such a shape identification device, discrimination of a target shape is performed based on the position of the shape within the imaging field of view,
A so-called space-independent feature quantity, that is, a scalar quantity, is used as a shape identification parameter to perform the process regardless of direction. The commonly used shape identification parameters are shown in the table below.

【table】

[Means for solving problems]

The input target shape is extracted by connectivity analysis to extract individual shapes, and the shape outline and feature points [points where the shape outline changes (convex points, concave points) and the center of gravity of the shape (overall center of gravity, hole center of gravity, etc.) ], select one feature point from among them to serve as a reference point (referred to as a reference point), and create a "shape feature vector" from that reference point to other feature points (convex points, concave points, center of gravity). seek,
By using the feature values obtained by performing vector operations including products and sums between the feature vector groups for shape discrimination, the correlation of shapes is incorporated into shape recognition, and the conventional shape feature values (area, Perimeter,
The micro-absolute amount based on shape coefficients, etc.) and the macro-level correlation amount based on the feature vector can be used together for shape discrimination, thereby stabilizing shape discrimination. [Effect] In other words, the shape feature vector is space-independent when viewed from the entire image, but it is space-dependent in the relative relationship of each feature point of the shape.
Since this vector quantity can express the (internal) positional relationship of the shape itself, this vector quantity is used to compensate for the insufficiency of conventional shape feature quantities. [Embodiment] Before explaining the embodiment, the concept will be explained with reference to FIGS. 2 to 4. Furthermore, Fig. 2 is an explanatory diagram for explaining the concept of the present invention, Fig. 3 is an explanatory diagram for explaining an example in which the small protrusion is sufficiently stable, and Fig. 4 is an explanatory diagram for explaining the case where the small protrusion is sufficiently stable. FIG. First, connected components of a shape are extracted by well-known connectivity analysis, and then contour points of the shape are extracted by contour tracking. the result,
Shape 1, contour 2, hole 3, line element 4 as shown in Figure 2 A
etc. are extracted. At this time, by detecting changes in each direction code of the contour points, the convex point P _i of the shape is
(i=1, 2..., n), concave point N _j (j=1, 2...
..., n) is extracted. The actual digital image is
Since there are noises, quantization errors, illumination variations, etc., averaging or smoothing processing is required to detect end points such as convex points and concave points, but this will be described later. Next, from each line element 4 constituting the shape, the center of gravity G _k (k=0, 1,
... n) is calculated. The center of gravity includes the center of gravity of the entire shape and the center of gravity of a hole. Now, in the present invention, since the relative positional relationship of each feature point in the shape is important, P _i , N _j ,
The numbering of each feature point in G _k must not be arbitrary. At this time, there are several methods for labeling P _j , N _i , and G _k in a space-dependent manner regardless of the _position and orientation of the _shape . The left end of the side with the longest side length is set as P ₁ , and the following is labeled counterclockwise as P ₂ , P ₃ , ...Pn, and when a concave point appears, it is labeled N ₁ , N ₂ , ...Nn in order. It shall be. Regarding the center of gravity _Gk , the overall center of gravity is _G0 , the _Gk point closest to _G0 is hereinafter _G1 , and the labels are sequentially labeled _G2 , _G3 ... _G0 . In this way, each feature point (P _i , N _j , G _k )
is extracted as shown in Figure 2B. Next, from among the feature points (P _i , N _j , G _k ) extracted as described above, the "shape reference point" that should be the starting point of the shape feature vector is determined. As shown, the feature points (P _i , N _j , G _k ) are inside the shape.
Since space-dependent labeling is performed, any point can be selected. That is, if P ₂ , it may be fixed as P ₂ , or it may be user programmable and variable for each target shape. In the following explanation, the shape reference point is assumed to be _P2 . Next, as shown in FIG. 2C, shape feature vectors μi→ and νi→ from point _P2 to each feature point are determined. Since shape feature vectors should also be space-dependent quantities like shape feature points,
The shape feature vector formed by P ₂ and the first feature point (P ₃ in Figure 2 C) when going counterclockwise around the contour of the shape is assumed to be μ ₁ →, and then μ ₂ in turn (counterclockwise) →, μ ₃ →..., μn →. Also,
Regarding the feature vector formed by the feature point and the center of gravity, for example, let the feature vector formed by P ₂ and G ₀ be ν ₁ →, hereafter the feature vector formed by P ₂ and G ₁ be ν ₂ →, and similarly, ν ₃ →, ..., νn→. However, the method of determining the feature vectors shown here is just one example and is not limited to this. The following basic vector operations are defined to determine the mutual relationships between the shape feature vectors obtained as described above, and each vectorized feature quantity (vector product between feature vectors, feature vector sum , the ratio of absolute values between feature vectors, etc.). (1) Vector product between feature vectors Between uneven points mij→=μ ₁ →×μj→ ... Between centers of gravity mij→=νi→×νi→ ... Between uneven points and centers of gravity lij→=μi→×νi→ ... ( 2) Feature vector sum M= 〓 ⁱ μi→+ 〓 ^j νi→ ... (3) Ratio of absolute values between feature vectors Between uneven points s _ij = |μi→ | / | μj → | ... Between centroids t _ij = |νi→|/|νj→| ... u _ij between uneven points and center of gravity = |μi→|/|νi→| ... The geometrical meaning of the above vector operation is as follows. In other words, the vector product between feature vectors is the area and orientation (angle) of a portion made up of any three points among the shape feature points, and the ratio of absolute values between feature vectors is the This is the side length ratio of the part made up of three points. Further, the feature vector sum indicates not only the characteristics of the shape but also the orientation of the shape. In addition, several feature vector operations can be defined depending on the complexity or specificity of the shape, but in the present invention, only the above (1) to (3) are shown as basic ones. . Next, the relationship between the shape feature amount based on such a shape feature vector and the shape feature amount conventionally used will be described below. The shape feature vector shown in this invention is the extraction of shape information about the part that forms the skeleton of the shape, and attempts to extract the essence of the shape as a shape without looking at micro errors in the shape. It is. However, if the microscopic part of the shape is the essence of the shape and the part is stable as a shape, the end points of that part are extracted and added to the shape feature vector group. be able to. Figure 3 is for explaining such a case, and shows the end point from which _Ps is extracted. On the other hand, conventional shape features (area, perimeter, shape coefficient, etc.) express the properties of a shape, but since they are parameters in absolute quantities, information extraction from the shape is difficult. It is therefore easy to be affected by fluctuations in small details. However, in many cases it is possible to extract features that are difficult to extract using the shape feature vectors mentioned above.
In other words, it is necessary to use them in a complementary manner depending on the shape. An example of complementary use is shown in FIG. this is,
This is an example in which the shape is composed of only circular arcs 5 that are perfect circles.In such a case, a method of extracting shape feature vectors by linearly approximating the circular arcs may be considered, but the more perfect circles are, the more linear approximation becomes. Because it becomes difficult,
Rather, the shape feature vector should extract the relative positional relationship between circular arcs, and the size of the circular arc portion should be determined based on the area, perimeter, shape coefficient, etc. FIG. 1 shows an outline of the flow of shape identification according to the present invention as described above. That is, as described above, after capturing the target shape, the shape feature points (P _i , N _j , G _k ) are extracted using this method through connectivity analysis and contour tracking (see ). . Next, the linearity of the shape contour is analyzed based on the direction code data obtained by contour tracking of the target shape, and the presence or absence of circular arc parts is investigated (
reference). Even if it is a circular arc part, for example, if a straight line approximation is possible as shown in Fig. 5, a straight line approximation is performed.
Extract feature points. Also, for things that cannot be approximated by a straight line (such as a perfect circle), only that part is sent to the arc part processing routine (see ).
In this method, a shape that is closed only by circular arc components, such as a perfect circle, cannot be approximated by a straight line, and other circular arc components are approximated by a straight line by extracting representative points of the circular arc components. Next, a shape feature vector is extracted from the extracted feature point group (reference), and the vector operation as described above is performed to calculate the essential feature amount of the shape (reference). Next, the shape is identified (reference) by comparing it with a standard pattern of feature quantities (based on feature vectors) that has been learned in advance by shooting or the like. On the other hand, for circular arc parts such as perfect circles, the usual feature values (area, perimeter,
Shape coefficients, etc.) are calculated (reference) and compared with the standard pattern of circular arc parts learned through shooting, and candidates for the shape containing that part are selected (reference). Finally, the final shape recognition effect is obtained by matching the discrimination based on the feature vector with the shape candidate of the circular arc portion (see). Here, we will discuss specific processing for handling digital images that are actually input. First, regarding detection of end points (convex points, concave points),
Basically, it can be obtained by detecting changes in the direction code, but the contour after digitizing the input image is generally not smooth as shown in Figure 6A, and if left as it is, there will be more data than necessary. End points are detected, and the number of end points that are positionally unstable increases. Therefore, stable end points are detected by smoothing or averaging the direction code during contour tracing. This smoothing process can be achieved, for example, by averaging the direction code of the contour pixel of interest and the direction codes of m pixels before and after it, and then reusing this as the direction code of the pixel of interest, as shown in Figure 6B. Yes (generally m = about 2). Next, after smoothing in this way, the sixth
As shown in FIG.
If the change in the direction code of pixels at both ends of the window continues for a certain value or more, that portion is extracted as an end point, thereby preventing small protruding noise shapes from being detected as end points. Further, even if the end points are detected stably in this way, they are not necessarily detected with accuracy on a pixel-by-pixel basis due to variations in the surrounding conditions, and there will be a certain variation (distribution). Therefore, it is assumed that the degree of dispersion is almost the same at each end point, and a certain margin is added in consideration of the spread of points after calculating each feature vector during shooting. As a result, a relatively stable standard pattern can be obtained without necessarily repeating shooting many times. Regarding feature vector calculations in the present invention, since the object is digital, it is not necessarily necessary to execute all calculations, but to speed up the calculation, a table of vector calculations is stored in memory and the table is searched. By doing this, approximate calculation results can be obtained. [Effects of the Invention] According to the present invention, the change in the outline of the shape (inflection part)
Using a shape feature vector that can express the essence of the shape without being affected by shape errors in unchanged parts because it is detected centered on Since identification is performed based on vectorized feature amounts obtained by calculation, stable and highly accurate identification is possible.

[Brief explanation of drawings]

FIG. 1 is a flowchart showing an embodiment of the present invention, FIG. 2 is an explanatory diagram for explaining the concept of the present invention, and FIG. 3 is an example in which the small protrusion is sufficiently stable. Figure 4 is an explanatory diagram to explain the feature extraction method for a perfect circle. Figure 5 is an explanatory diagram to explain a circular arc pattern that can be approximated by a straight line. Figure 6 is a digital diagram. An explanatory diagram for explaining a specific image processing method, FIG. 7 is a schematic diagram showing a general shape recognition device, and FIG.
It is a tree diagram which shows an example of the discrimination method in a figure. Code explanation, 1...Object (shape), 2...Outline,
3...Hole, 4...Line element, 5...Perfect circle, 11...
Input device, 12... Preprocessing circuit, 13... Shape extraction circuit, 14... Connectivity analysis section, 15... Feature value calculation section, 16... Shape identification section, 17... Standard pattern storage section, 18 ... Output device, P ₁ - P _o ... Convex point, N ₁ - N _o ... Concave point, G ₀ - _{G o} ... Center of gravity.

Claims (1)

[Scope of Claims] 1. A shape recognition method for recognizing the shape of an object by capturing an image of the object and processing the image signal, comprising: tracing the outline of the object shape and extracting each feature point; Analyze the linearity between the feature points to determine whether linear approximation is possible, and if direct approximation is possible, select one standard feature point from among the feature points, and Find the shape feature vectors from the reference feature point to each other feature point, perform vector operations including product and sum among the group of shape feature vectors to find vectorized features, and perform recognition based on the vectorized features. However, when it is determined that linear approximation is not possible as a result of the contour tracing, recognition is performed based on a shape feature vector and a shape feature amount based on each feature point.