JP4525526B2 - Pattern matching method and apparatus - Google Patents

Description

  The present invention relates to a pattern matching method for detecting a recognition target object by comparing a graphic template representing the shape of a recognition target object with a verification target graphic displayed in a verification target image.

  There is pattern matching as a technique for previously registering a set of graphic data representing the shape of a two-dimensional or three-dimensional recognition target object as a graphic template and detecting a graphic similar to the graphic template from the image space.

  This pattern matching can recognize the presence / absence of figures, figure position / posture, figure discrimination, etc., for character recognition, character inspection, alignment by pattern (figure shape) recognition, object recognition by robot vision, etc. Can be applied.

  The pattern matching function detects a known pattern from an image and obtains its position, pattern rotation angle, scale, etc. In positioning applications for inspection and assembly using image processing, edge detection and blob It is a function that is frequently used in industrial image processing because it is more versatile than methods such as analysis, and is intuitive and easy to use for users who do not have basic knowledge and program creation techniques related to image processing. Of the techniques known as pattern matching algorithms, the normalized correlation method and the generalized Hough transform are the most famous.

  The normalized correlation method is a method in which the pattern image itself is superimposed on the recognition target image, and the position where the cross-correlation value between the pixel values of the pattern image and the recognition target image is maximized is detected. Since the normalized correlation method detects a position having a high degree of similarity using the image itself as a template, there is an advantage that there are few matching errors, but there is a problem that the processing amount is large and the execution speed is slow. In particular, when rotation or scale variation occurs, a large number of deformed templates subjected to rotation or scale variation must be prepared and sequentially verified, and a practical speed cannot be obtained. In addition, since the cross-correlation value between the pixel value of the template image and the pixel value of the recognition target image when the images are overlaid is used, no matching error occurs even if the overall brightness of the image fluctuates, but noise and shading If the lightness of the pattern partially changes or the background changes drastically due to the above, etc., there is a problem that the cross-correlation value is greatly lowered and a matching error occurs.

  Also, the generalized Hough transform formulates the desired shape, and votes the trajectory that maps each edge point to the same voting space (Hough space) as the number of elements constituting the feature vector of the mathematical formula. Detecting the shape you want from the image by detecting the position where the vote value is high, it is strong against partial omission of patterns due to rotation, scale fluctuation, concealment by other objects, etc. However, since the vote value increases if the shapes match, there is a problem that erroneous recognition is likely to occur in the case of an image including a partially similar pattern. In addition, the more complicated the pattern to be detected, the higher the dimension of the Hough space has to be, so there is a disadvantage that both the memory consumption and the processing time become enormous.

  As a means for solving the problems of these existing methods, there is a method disclosed in Patent Document 1. In the method of Patent Document 1, first, the contour line of the target pattern is extracted, the length r of the line segment connecting the pattern center point and the point for each point on the contour line, and the angle θ between the normal direction at that point. A template is generated by obtaining a shift vector having and as elements. In addition, a normal vector is obtained for each edge point of the target image at the time of matching, and the values of r and θ when the edge points of the template are applied to the edge points are voted on the voting plane to obtain the highest voting value. Is obtained as a pattern detection position.

  Further, as another means for solving the problems of the existing method, there is a method disclosed in Patent Document 2. In the method of Patent Document 2, the contour line of the target pattern is analyzed and decomposed into a portion effective for rough positioning, a portion effective for angle detection, and a portion effective for detection of magnification, and a pattern effective for rough positioning is used. After positioning the pattern at high speed, the angle and magnification are finely adjusted.

  Furthermore, as another means for solving the problems of the existing method, there is a method disclosed in Patent Document 3. In the method of Patent Document 3, an amount of deviation that occurs when a template placed in alignment with a rotated pattern is collated, and after the template is collated with the pattern, the pattern detection position is calibrated using the amount of deviation. I have to.

  Further, as another means for solving the problems of the existing method, there is a method disclosed in Patent Document 4 as a means relating to pattern matching of a three-dimensional shape in particular. In the method of Patent Document 4, the feature points of the object to be recognized and the dimensions of the circumscribed cuboid of the object are stored as a database. First, the position where the object exists is detected using the circumscribed cuboid, and then the small object defined in the object is detected. Matching is performed for each block.

Here, in the graphic matching process in the prior art disclosed in the above-mentioned patent document, the process of aligning the measurement point group with the graphic template to obtain the similarity is performed with respect to all possible positions / postures. The process of performing “crushing” and obtaining the position where the highest similarity is obtained is performed. Pattern matching using Hough transform or neural network is also a process for obtaining similarities for all possible positions and orientations, with the only difference being a mathematical method for quickly terminating the process when “no matching”. The point is the same. When dealing with simple graphics such as symbol recognition and character recognition, the Hough transform and neural network can often shorten the total processing time. In general pattern matching processing that handles complex patterns, template graphics Since the amount of data increases, processing time is often required. Therefore, in such a general-purpose pattern matching process, it is more effective to perform the position alignment in a collapsing manner by a collation process as fast as possible for a single position alignment.
Japanese Patent No. 3437141 JP 2004-219284 A JP 2002-319028 A JP 2004-226197 A

  However, in order to perform similarity calculation, a process for associating a measurement point group with a graphic template (a process for obtaining a distance and a corresponding point from each point to the closest line among the component graphic elements of the graphic template) is required. However, in this process, the distance and corresponding points from each coordinate point to all the component graphic elements of the graphic template are obtained, and the complicated process of selecting the one with the smallest distance is selected, so that the processing time is increased. There was a problem.

  In addition, the method disclosed in Patent Document 1 described above has a feature that the pattern can be detected with a smaller number of dimensions compared to the generalized Hough transform, so that the memory consumption is small and the processing speed is high, and the rotation is invariant. Can also be recognized. However, when there is a change in the scale, detection becomes impossible, and the phenomenon that the vote value becomes high even in a partially matching pattern occurs in the same manner as in the generalized Hough transform, so there is a lack of pattern. There was a problem that misrecognition was likely to occur.

  Further, the method disclosed in Patent Document 2 has a problem that high-speed processing is possible when the angle variation is not large, but it cannot be used when the angle varies greatly.

  In addition, the method disclosed in Patent Document 3 has an advantage that high-speed processing is possible because a pattern including some rotation can be recognized without using a rotation template, but accuracy is inferior. There was a problem that it could not be used when large rotations were involved.

  Furthermore, in the method disclosed in Patent Document 4, high-speed object recognition is possible when an object exists alone, but the apparent dimension of the object varies when a plurality of objects are in contact with each other. In some cases, there was a problem that it might become unrecognizable.

  An object of the present invention is to provide a pattern matching method that can solve the above-described problems of the prior art and perform high-speed pattern matching processing.

  In order to solve the above-described problem, the pattern matching method of the present invention is a pattern matching method for detecting the recognition target object by comparing a graphic template representing the shape of the recognition target object with a verification target image, and A graphic template generation step of generating an image template by processing an image of an object photographed, and measuring a verification target graphic displayed in the verification target image by performing image processing on the verification target image A measurement point data generation step for generating measurement point data displayed as a set of points, and potential space data generation for generating potential space data in which score values set based on the relative positional relationship from the graphic template are recorded Step, collating the measurement point data with the potential space data, A score value determination step for obtaining a score value at a fixed point; a similarity between the verification target graphic and the graphic template is calculated based on the score value determined in the score value determination step; and the verification is performed based on the similarity And a determination step of determining whether or not the target graphic is the recognition target object.

  In the pattern matching method according to the present invention, potential space data in which a score value is set based on a relative positional relationship from a graphic template is generated, and the score value is determined by collating the potential space data with measurement point data. Since the similarity is calculated based on the obtained score value and it is determined whether or not the verification target graphic is a recognition target object, the corresponding point search process that has been necessary in the past is unnecessary, and potential space data Similarity can be obtained simply by calculating the score value obtained for each measurement point when the measurement points are collated with each other. This makes it possible to realize pattern matching processing much faster than conventional methods. .

  An apparatus configuration for carrying out the pattern matching method according to the present invention and an example of pattern matching using this apparatus will be described below.

  Embodiment 1 of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram illustrating a configuration of a pattern matching apparatus for performing the pattern matching method according to the first embodiment.

  As shown in FIG. 1, the pattern matching device 1 of the present embodiment includes a TV camera 3 as an image input device that inputs an image of a sheet-like object 2, and an analog signal image captured by the TV camera 3. A digital image generating device 4 for converting into a digital signal, a storage device 5 for storing digital data generated by the digital image generating device 4, and an image processing device 6 for performing image processing and performing pattern matching are provided. .

  In the pattern matching apparatus 1 configured as described above, a pattern is printed on the surface of the object 2, and when an image including the pattern is captured by the TV camera 3, an image of the captured analog signal is generated as a digital image. A grayscale image in which the pixel value becomes a density value is generated by being input to the device 4 and converted into a digital signal. The grayscale image is stored in the storage device 5, and the image processing device 6 performs pattern matching by performing image processing described below on the grayscale image stored in the storage device 5.

  Note that the object 2 is not limited to the one in which the planar pattern as described above is printed, and the method of the present invention can be used similarly even if it is a cube or the like.

  Here, the concept of template matching performed by the pattern matching method of the present embodiment will be described with reference to FIG. In template matching, the graphic template 31 in which the recognition target object is recorded is superimposed on every position of the verification target image 32 where the verification target graphic is photographed to obtain the similarity, and the verification target at the position showing the highest similarity This is a method of detecting that a figure is a recognition object. For example, since the figure template is aligned at a position which is not related at the position a in FIG. 2, the similarity is close to zero, and at the position of FIG. 2b, the similarity is aligned with a different pattern. Low value. The similarity is extremely high at the position shown in FIG. 2c, and the position c is determined to be the position where the recognition target is present.

Next, the collation process by the pattern matching method of a present Example is demonstrated based on the flowchart of FIG. As shown in FIG. 3, a graphic template is first generated (S301).

  Here, an example of a graphic template used in the pattern matching method of this embodiment is shown in FIG. The graphic template in FIG. 4A is obtained by recording an arrow shape with seven straight lines 11 to 17. This graphic template may be a mathematical expression that represents seven straight lines in advance, or may be an imported graphic drawn with CAD or the like. Further, as shown in FIG. 4B, a series of coordinate points constituting a figure may be arranged in the order of connection. The coordinate points as shown in FIG. 4B are given as coordinate point sequences from the beginning if they are three-dimensional shape data, but are generated by performing the following processing when extracted by image processing. can do.

  First, from the grayscale image 51 shown in FIG. 5A recorded in the storage device 5 by the image processing device 6, the edge intensity image 52 shown in FIG. 5B and the edge direction image shown in FIG. 53 is extracted. The edge intensity image 52 is an image in which the magnitude of change in brightness (edge intensity) is recorded, and the edge direction image 53 is an image in which the direction in which the brightness increases is recorded.

  Here, since the edge strength image 52 is an image obtained by simply obtaining the edge strengths for all the pixels, threshold processing is performed on the edge strength image 52 (coordinates having edge strengths greater than or equal to the threshold value are extracted. By performing the processing, the edge point image 54 shown in FIG. 5D can be acquired. This process can be performed by Sobel filtering which is a known technique. Further, if the thinning process is performed so that the line width becomes 1 in the edge point image 54 and the edge point is traced starting from an arbitrary edge point, the measurement point data 55 shown in FIG. can get. What is necessary is just to use this as a figure template as it is.

  Further, in the processing for obtaining the edge point image 54 of FIG. 5D, there are portions where the edge strength is lower than other portions, such as the portions d1 and d2 of the edge strength image 52 of FIG. In some cases, if the threshold is higher than the edge strength of this portion, the edge is interrupted as shown in FIG. In such a case, the direction of the line is recorded at both ends of the line data 61 and 62 obtained by the tracking process, and if another line exists in the opposite direction, the two lines are merged into one. If the process (edge extension process) is performed, the broken edge can be repaired as shown in FIG.

  Further, approximation processing is performed from the measurement point data 55 of FIGS. 4B and 5E and the repaired measurement point data as shown in FIG. 6B, and the mathematical formula as shown in FIG. A graphic template described by the above may be generated. This approximation process may be performed using a known method such as broken line approximation.

  When describing a graphic template using mathematical formulas, the data amount of the graphic template should be expressed with as few lines as possible under the condition that the error between the approximate graphic and the original measurement point data is within the allowable range. This is advantageous because the amount of processing can be reduced and the amount of processing can be reduced when generating potential space data, which will be described later. Therefore, in this embodiment, a graphic template is generated by approximating an image of a recognition target object with a set of equations of order 1 or higher.

  For this purpose, a polygonal line approximation is first performed from the measurement point data as shown in FIG. 7A to generate line segments 111 to 118 as shown in FIG. By integrating a plurality of broken lines l13 to 116 generated in the arc portion and approximating the ellipse again, figures 121 to 125 formed from straight lines and arcs as shown in FIG. 7C can be formed. In addition to this, a method of applying Hough conversion may be used.

  When the graphic template is generated in this way, the image to be collated is subjected to image processing to generate measurement point data (S302). However, the image processing performed here is the same as the processing for generating the graphic template described above, and the generated measurement point data displays the verification target graphic as a set of measurement points as shown in FIGS. Is.

  Next, a score value set based on the relative positional relationship from the graphic template is obtained to generate potential space data (S303).

  As described with reference to FIG. 2, in template matching, it is necessary to sequentially obtain the similarity when the graphic template and the image to be collated are superimposed, and various indexes are used as the similarity depending on the template matching method. ing. A typical example is “a cross-correlation value between a pixel of a superimposed template image and a pixel of an image obtained by photographing a recognition object” in a normalized correlation method which is a known technique.

  In this embodiment, since the matching is based on the shape of the edge, for example, it is necessary to use, as the similarity, an index that increases the numerical value when the sum of the distances between the measurement points and the graphic template is small. .

  However, in order to perform such processing, it is necessary to perform processing for obtaining the closest position for the elements constituting the graphic template. For example, FIG. 8 shows a graphic template composed of straight lines 11 to 17, and when this graphic template is collated on the measurement point data to be collated, the graphic template is collated on the measurement point data Pm. After aligning with (x, y), the relative position Pr (xr, yr) from the verification position Pm (x, y) is obtained for each measurement point p (px, py).

The relative position Pr (xr, yr) is expressed as xr = px−x
yr = py-y
Can be obtained.

  And the process which calculates | requires the nearest position about all the components (straight lines 11-17) of a graphic template, and calculates | requires the position pc from which the distance becomes the shortest value dmin (min is a subscript) as a corresponding point is requested | required. However, for each of a large number of measurement point data, the process of evaluating the distances from all the components of the graphic template is extremely large in processing amount, and the execution speed becomes too slow to be practical.

  Therefore, in this embodiment, the correspondence between the measurement point data and the graphic template is determined depending on the relative positional relationship between the graphic template and the measurement point data. Potential space data to which score values are assigned based on the relative positional relationship with the template is generated. Then, when the similarity is obtained by comparing the graphic template with the measurement point data, the score value of each measurement point is obtained by referring to the potential space data instead of the processing for obtaining the corresponding point described above, and for all measurement points. High-speed template matching is realized by using the sum of score values as the similarity.

  Here, the potential space data will be described. The potential space data is data in which a score value obtained by an expression defined based on a relative positional relationship with the graphic template is recorded. As the simplest example, there is potential space data to which a score value is assigned so that the score value decreases as the distance from the graphic template increases. An example of this is shown in FIG.

In the potential space data shown in FIG. 9, the score value s is

When the distance d (xr, yr) is 0 or more and less than 10, the equation (1) is the distance d (xr, yr) = 0 and the score value s = 10, and the distance d (xr, yr) As the value increases, the score value decreases linearly, and when the distance d (xr, yr) is 10 or more, the score value s = 0. The potential space data generated thereby becomes a two-dimensional contour figure having a graphic template as a ridge as shown in FIG.

  Further, in the above-described method of defining the potential space data, it is defined by the equation (1) that depends only on the distance, but the method for defining the score value of the potential space data is not limited to the equation (1), Other functions may be used as long as the performance can be realized.

Therefore, as another method for defining the score value, for example, a function represented by the following equation (2) may be used.

  In this equation (2), the score value s (xr, yr) decreases in proportion to the square of the distance, and is steep in a region closer to the graphic template than the function of equation (1). Since the score value increases, it is suitable for matching when there are many similar patterns.

Further, the following equation (3) using the normal vector θ (xr, yr) (FIG. 8) on the surface of the graphic template at the corresponding point and the edge direction δ (px, py) at the measurement point p (px, py) together. There may be.

  In this equation (3), the value at the measurement point p (px, py) of the edge direction image 53 in FIG. 5C is used. In equation (1), the score value decreases linearly with respect to the distance. However, in Expression (3), in addition to the distance, the difference between the normal vector θ (xr, yr) on the surface of the graphic template at the corresponding point and the normal direction δ (px, py) in the measurement point data is within 90 degrees. In this case, the cosine value of this direction difference is multiplied, so that “a high score value is given when the distance is close and the normal direction of the surface is the same”. Compared to equation (1), the collation reliability can be increased.

  In the case of Expression (3), the score value s (xr, yr, δ) depends on the normal direction δ of the measurement point data together with the relative position Pr (xr, yr), so the potential space data is as shown in FIG. In addition, three-dimensional data in which score values of a one-dimensional array of δ are assigned to each point around the graphic template. In order to reduce the data amount, the score value s (xr, yr, δ) is not directly recorded as potential space data, but only the normal direction θ (xr, yr) is recorded. The weight cos (θ (xr, yr) −δ (px, py)) may be calculated when executing the matching process.

  The potential space data generated in this way can be uniquely determined depending only on the relative positional relationship between the graphic template and the measurement points around it. Therefore, “A celestial body has a gravitational field around it, and an object approaching the celestial body is subject to gravity (potential energy) according to the distance between the mass and the celestial body.” “A magnetic field is created around the magnet. The magnetic material close to the magnet receives the reaction force or attractive force according to the magnetic force of the magnetic material itself, the magnetization direction of the magnetic material, and the mutual distance. Set a “potential field” that gives a potential value according to the “shortness of distance” and “the degree of coincidence of normal vectors on the surface”, and for each of the measurement points that entered the potential field, If the potential values uniquely obtained based on the relative positional relationship between the two are larger, the larger the total potential value, the greater the number of measurement points that have entered the potential field. It will be located and the normal vector of the plate is similar "can easily detect the position degree of similarity in the verification process.

  Since the potential value given to the measurement point that has entered the potential field can be obtained only depending on the relative positional relationship between the measurement point and the graphic template as described above, it has been described with reference to FIG. There is no need for such a complicated association process, and it is only necessary to refer to “potential space data” that records the “potential value at that position” in the potential space determined in advance. Even if the alignment matching process is performed in a brilliant manner, not only the conventional pattern matching method that requires the matching process but also the mathematical solution method using Hough transform, neural network, etc. Much faster pattern matching can be achieved.

  When the potential space data is generated in this way, the recognition target object is detected by collating the potential space data with the collation target image. Here, an example in which the image to be collated shown in FIG. 2 is collated using the potential space data shown in FIG. 9 will be described based on FIG.

  Following the processing of step S303 in FIG. 3, first, a collation position Pm (x, y) for collating the potential space data with the collation target image is set as a search start point (S304). Usually, the search start point is set at the upper left end of the image to be collated, and the raster scanning is performed in which the search is made from there to the right end, then moved to the left end one pixel below, and the search is made to the right end again.

  In this way, raster scanning is performed. For example, when the potential space data is collated at the position shown in FIG. 11b, the measurement point number is set as the first measurement point among the measurement points on the star-shaped verification target graphic. Collation (1) The value of similarity S (x, y) in Pm (x, y) is reset to 0 (S305). Here, it is assumed that the measurement point numbers assigned to the measurement points p101 to p120 overlapping with the potential space data are assigned in raster order.

  Next, the process (S306 to S310) of calculating the similarity S (x, y) by obtaining the score value at each measurement point by referring to the potential space data is performed. First, the coordinate value of the first measurement point p101 is converted into relative coordinates with the template reference position as the origin. In this process, since the template reference position is aligned with the collation position Pm (x, y), as shown in S306 of FIG. 3, the collation position Pm (x, y) of the measurement point coordinates p101 (px, py) is used. The relative position Pr (xr, yr) may be obtained. Since the potential space data holds the score value s (xr, yr) at the position of the relative position Pr (xr, yr), the score value s (xr, yr) is referred to (S307) and the matching position Pm ( Add to the similarity S (x, y) in x, y) (S308). For example, in the case of the measurement point p101, the score value s (xr, yr) is 8, as can be seen from FIG. Further, if there are other measurement points, the process moves to the next measurement point (S309 to S310), and the same processing is performed for all the measurement points p101 to p120 that overlap with the potential space data, so that the collation position Pm (x, The similarity S (x, y) in y) is obtained.

  Here, the value of the similarity S (x, y) is compared with the maximum similarity Smax (max is a subscript) (S311). If the similarity S (x, y) at the collation position Pm (x, y) is larger than Smax, the current collation position Pm (x, y) is substituted into the detection position m (x, y) (S312). After that, when the raster scanning of the collation position Pm (x, y) is not completed (there is a next candidate), the collation position Pm (x, y) is set as the next candidate and the similarity is further similar to the above. The process for obtaining S (x, y) is repeated (S313 to S314). In step S313, when raster scanning is completed and there is no next candidate, the detection position m (x, y) at the next point is the result of pattern matching.

  Here, the processing result by the pattern matching mentioned above is demonstrated using FIG. First, the similarity S (x, y) obtained when raster scanning is performed and the collation position Pm (x, y) reaches the position shown in FIG. 11a is 0 because there is no measurement point.

  Next, the similarity S (x, y) obtained when the collation position Pm (x, y) comes to the position of FIG. 11b where a star-shaped figure different from the figure of the figure template exists is the measurement point p101. ˜120 exist, the score S at each measurement point is summed with reference to the potential space data by the above-described process, and a low value of similarity S (x, y) = 156 is obtained.

  Next, the similarity S (x, y) obtained when the collation position Pm (x, y) comes to the position of FIG. 11c where the same figure as the figure of the figure template exists is measured by the measurement points p201 to 223. Since these exist and overlap with a high score portion of the potential space data, a high value of similarity S (x, y) = 230 is obtained, and this value is the maximum similarity in the image. Finally, the position of FIG. 11c is obtained as the detection position m (x, y) of the recognition object.

  As described above, according to the pattern matching method of the present invention, the similarity is obtained by using the potential space data, so that the measurement point is displayed on the graphic template as described in the description of the potential space data generation processing. Complicated processing such as detection of corresponding points, calculation of distances to measurement points, and calculation of score values based on distances are not required for the process of matching templates and measurement point data. Since the similarity can be obtained by simply referring to the spatial data and simply performing the summing process, extremely fast template matching is possible.

  In the example described above, the similarity is the sum of the score values, but for example, an average value of the similarities may be used. When the average value is used, the similarity at the position of FIG. 11a is 0, the similarity at the position of FIG. 11b is 7.8, the similarity at the position of FIG. 11c is 10, and the position of FIG. It becomes the detection position of the recognition object.

  In the above-described example, the direction of the figure to be collated is the same as the direction of the figure template, that is, it is known that rotation does not occur. It is necessary to perform affine transformation (enlargement / reduction / rotation transformation) on the potential space data. However, the process of performing affine transformation on all measurement points is not practical due to the large amount of processing, and the processing of performing affine transformation on potential space data is also necessary when the range of rotation or scale fluctuation is large. Become more. Therefore, in such a case, as shown in FIG. 12, data that has been subjected to affine transformation in advance is prepared for potential space data, and the potential space data that has been subjected to affine transformation is sequentially collated when performing collation processing. By doing so, it is possible to recognize even when there is rotation or scale fluctuation.

  As described above, in the pattern matching method of the present embodiment, potential space data in which score values are set based on the relative positional relationship from the graphic template is generated, and the potential space data and the measurement point data are collated. Since the score value is obtained and the similarity is calculated based on the obtained score value to determine whether or not the figure to be collated is a recognition target object, the corresponding point search process that was necessary in the past is unnecessary. Therefore, it is possible to obtain a similarity by simply calculating the score value obtained for each measurement point when matching the measurement point with the potential space data, thereby realizing a much faster pattern matching process than before. It becomes possible.

  Further, in the pattern matching method of the present embodiment, since the generated graphic template is repaired when there is a missing portion, the original shape of the recognition target can be correctly represented even when there is a missing graphic template, and there is a lot of noise. This method can be applied even to a pattern. Further, with such a configuration, the coordinate point density of the graphic template can be lowered in the redundant straight line portion, and the capacity of the graphic template can be suppressed.

  Further, in the pattern matching method of the present embodiment, the graphic template is generated by approximating the image of the recognition target object with a set of equations of the first order or higher, so that the capacity of the graphic template can be minimized. High-speed pattern matching processing can be realized.

  Next, a second embodiment of the present invention will be described with reference to the drawings. The pattern matching method of this embodiment is different from the first embodiment in that the minimum measurement points are selected within a range in which the features of the figure to be collated are not lost, and the score value is obtained for these measurement points. Yes. However, since other processes are the same as those in the first embodiment, detailed description thereof is omitted.

  In FIG. 11 showing an example of the pattern matching process of the first embodiment, the score values are obtained with reference to the potential space data for all the measurement points, but the measurement points are thinned out within a range that does not impair the feature of the figure. As a result, higher-speed processing can be realized.

  Therefore, in this embodiment, a minimum measurement point is selected within a range in which the feature of the verification target graphic is not lost, and a score value is obtained for this measurement point.

  Here, FIG. 13 shows a state in which the minimum measurement points are selected from the measurement points in the example of FIG. As shown in FIG. 13, the number of measurement points is thinned from 20 to 10 in the star-shaped verification target graphic at the position of FIG. 13b, and the number of measurement points is reduced from 23 to 12 at the position of FIG. 13c. Has been.

  If the similarity is calculated in this case, it is 66 at the position of FIG. 13b and 120 at the position of FIG. 13c, and the similarity is maximum at the position of FIG. In this case, by selecting the thinning interval in the graphic template so as to be, for example, about half the length of the shortest line segment, the measurement points can be thinned out within a range that does not impair the characteristics of the graphic.

  As described above, in the pattern matching method of the present embodiment, the minimum measurement points are selected within a range in which the features of the verification target graphic are not lost, and the score value is obtained for the measurement points. It is possible to reduce the processing load without causing erroneous determination when collating.

  Next, Embodiment 3 of the present invention will be described with reference to the drawings. In the pattern matching method of this embodiment, potential space data is configured by sampling point groups representing score values, and a plurality of sampling points close to the measurement points are selected from the sampling point groups, and the score values of these sampling points are selected. The difference from the first embodiment is that the score value of the measurement point is obtained based on the above. However, since other processes are the same as those in the first embodiment, detailed description thereof is omitted.

  The potential space data shown in FIG. 9 and FIG. 10 described in the first embodiment is drawn so that score values are assigned for each region for convenience of explanation, but the potential space data actually generated by a computer is Since the sampling point is arranged at an appropriate sampling interval around the graphic template and the potential space data is assigned a score value for each sampling point, the form is as shown in FIG.

  In addition, the sampling point interval may be determined based on the shape of the graphic template so as not to impair the shape of the original graphic. Therefore, as shown in FIG. Then, if the sampling interval is set large and the sampling interval is set finely in a complicated part, the amount of potential space data can be kept low.

As described above, since the score values in the potential space data of the present embodiment are discretely arranged with a certain sampling interval, the relative position pr (xr, yr) of the measurement point shown in FIG. 15 is not necessarily the coordinates of the sampling point. It does not always match s (xr ′, yr ′). Therefore, in order to obtain the score value, the score values s (xr ′, yr ′), s (xr ′ + 1, yr ′) of four sampling points close to the relative position pr (xr, yr) of the measurement point, It is necessary to perform the interpolation processing shown in Expression (4) from s (xr ′, yr ′ + 1) and s (xr ′ + 1, yr ′ + 1).

  As described above, in the pattern matching method of this embodiment, the potential space data is configured by the sampling point group representing the score value, and a plurality of sampling points close to the measurement point are selected from the sampling point group, and these sampling points are selected. Since the score value of the measurement point is obtained based on the score value, it is possible to reduce the data amount of the graphic template by roughening the sampling interval when generating the potential space data. In addition, by using the method of this embodiment instead of generating potential space data with a narrow sampling interval in order to perform sub-pixel detection, it becomes possible to increase the accuracy without increasing the amount of data.

  Next, a fourth embodiment of the present invention will be described with reference to the drawings. In the pattern matching method of the present embodiment, a cluster approximated to a matching target graphic is generated, and a matching target graphic that is not a recognition target is detected based on blob information representing the characteristics of the cluster, and a score value is obtained from the target. This is different from the first embodiment. However, since other processes are the same as those in the first embodiment, detailed description thereof is omitted.

  In the example shown in FIG. 11 in the first embodiment, the position shown in FIG. 11a is a position where there is no measurement point, so it is meaningless to perform the collation process. Also. If it can be predicted in advance that the positions in FIG. Therefore, if such a case can be extracted easily, the processing speed can be greatly improved.

  Therefore, in this embodiment, as shown in FIG. 16, first, a process is performed in which a collection of measurement points that are relatively close to each other is used as a cluster, and then, for clusters c1 to c4 obtained by this process, the region Blob information representing pattern features such as width, height, density of measurement points, etc. is extracted, and collation is not performed when the blob information is clearly different.

  For example, in FIG. 16, the c1 and c2 clusters are completely different from the graphic template in height and width information, so that the matching process is not performed in this vicinity. In addition, high-speed matching processing can be executed by focusing on the vicinity of the center of the cluster at the positions c3 and c4.

  As described above, in the pattern matching method of this embodiment, a cluster approximated to the verification target graphic is generated, and the verification target graphic that is not the recognition target is detected based on the blob information representing the feature of the cluster, and the score value is detected. If the amount of processing that recognizes parts from 3D shape data measured at the same time for a large number of parts is very large, the parts to be recognized and other objects to be recognized such as noise Can be quickly identified, and the processing speed can be increased.

  Next, a fifth embodiment of the present invention will be described. The pattern matching method of this embodiment differs from that of the first embodiment in that the absolute similarity is calculated by calculating the cross-correlation value, which is a statistical value indicating the similarity between the matching target graphic and the graphic template. ing. However, since other processes are the same as those in the first embodiment, detailed description thereof is omitted.

  In the first embodiment, the recognition target object is detected by collating the potential space data with the collation target image obtained by photographing or measuring the recognition target object and recognizing the position / orientation of the pattern. The degree is merely outputting a position having a relatively high degree of similarity in the screen, and an absolute degree of similarity to the shape of the figure template is not obtained.

Therefore, in this embodiment, after the collation process described in the first embodiment is performed, the cross-correlation value Sc shown in Expression (5) is calculated again to obtain the absolute similarity.

  This cross-correlation value Sc is a position between the relative position Pr (xr, yr) and the foot of the perpendicular drawn to the element of the graphic template closest to the relative position, that is, the corresponding point Pc (xc, yc) in FIG. This is a statistical value indicating the overall similarity in the relationship. When the relative position Pr and the corresponding point Pc completely match, the cross correlation value Sc = 1 and the graphic template and the verification target graphic are completely overlapped. Show. If they do not match, the cross-correlation value Sc approaches zero. Therefore, the cross-correlation value Sc is obtained by normalizing the absolute similarity to a value of 0 to 1.

  By obtaining absolute similarity in this way, for example, when applying to character recognition, etc., it can be used for processing such as distinguishing which figure is photographed by using a plurality of different figure templates together Is possible. In addition, since the coordinate value of the corresponding point Pc is not recorded in the potential space data described above, the corresponding point Pc is anew after performing the matching process described in the first embodiment in order to carry out the method of the present embodiment. Or the coordinate value of the corresponding point Pc may be recorded in advance for each sampling point of the potential space data.

  As described above, in the pattern matching method of this embodiment, the cross-correlation value, which is a statistical value indicating the degree of similarity between the matching target graphic and the graphic template, is calculated. The absolute similarity to the figure can be obtained, and the reliability can be improved.

  Next, a sixth embodiment of the present invention will be described with reference to the drawings. In the first embodiment, the case where the object is two-dimensional data (image) has been described. In this embodiment, a case where pattern matching is performed on three-dimensional data (three-dimensional shape data) will be described.

  FIG. 17 is a block diagram illustrating a configuration of a pattern matching apparatus for performing the pattern matching method according to the sixth embodiment. As shown in FIG. 17, the pattern matching device 71 of this embodiment includes a laser three-dimensional sensor 73 as an image input device that inputs a three-dimensional shape of a three-dimensional object 72, and an output from the laser three-dimensional sensor 73. A three-dimensional data generation device 74 that converts a signal into three-dimensional shape data, a storage device 75 that stores data generated by the three-dimensional data generation device 74, and an image processing device 76 that performs image processing and performs pattern matching. And.

  In the pattern matching device 71 configured as described above, the laser three-dimensional sensor 73 scans the surface of the three-dimensional object 72 with a laser beam, and an output signal from the laser three-dimensional sensor 73 is sent to the three-dimensional data generation device 74. It is input and processed, and converted into a three-dimensional coordinate point group on the surface of the object 72. The three-dimensional coordinate point group is stored in the storage device 75, and the image processing device 76 performs image processing on the stored three-dimensional coordinate point group to perform pattern matching.

  The matching process performed by the pattern matching apparatus 71 of the present embodiment can be performed in the same manner as the process for the two-dimensional data described in the first embodiment. However, in the process of generating a graphic template, FIG. A graphic template composed of curved surfaces such as triangular patches, NURB, and Bezier as shown in FIG. Also, as shown in FIG. 18A, when there is a division of a figure or a hole (h in FIG. 18), a complementing process as shown in FIG. 18B is performed.

  By using the graphic template generated in this way, three-dimensional potential space data is generated, and verification processing with the three-dimensional measurement point data is performed in the same manner as in the first to fifth embodiments. According to the pattern matching method, pattern matching can be performed not only on two-dimensional data but also on three-dimensional data.

  The pattern matching method of the present invention has been described based on the illustrated embodiment. However, the present invention is not limited to this, and the configuration of each part is replaced with an arbitrary configuration having the same function. Can do.

It is a block diagram which shows the structure of the pattern matching apparatus for enforcing the pattern matching method which concerns on Example 1 of this invention. It is a figure for demonstrating the concept of template matching. It is a flowchart which shows the collation process by the pattern matching method which concerns on Example 1 of this invention. (A), (b) is a figure which shows an example of the figure template produced | generated by the pattern matching method based on Example 1 of this invention. (A)-(e) is a figure for demonstrating the image processing by the pattern matching method which concerns on Example 1 of this invention. (A), (b) is a figure for demonstrating the method to repair the missing part of the figure template by the pattern matching method based on Example 1 of this invention. (A)-(c) is a figure for demonstrating the approximation process of the figure template by the pattern matching method which concerns on Example 1 of this invention. It is a figure for demonstrating the process which calculates | requires the corresponding point in pattern matching. It is a figure which shows an example of the potential space data by the pattern matching method which concerns on Example 1 of this invention. It is a figure which shows an example of the potential space data by the pattern matching method which concerns on Example 1 of this invention. It is a figure which shows an example of the collation process by the pattern matching method which concerns on Example 1 of this invention. It is a figure which shows an example of potential space data in case there exists rotation and a scale fluctuation by the pattern matching method which concerns on Example 1 of this invention. It is a figure which shows an example of the collation process by the pattern matching method which concerns on Example 2 of this invention. (A), (b) is a figure which shows an example of the potential space data by the pattern matching method based on Example 3 of this invention. It is a figure for demonstrating how to obtain the score value by the pattern matching method which concerns on Example 3 of this invention. It is a figure which shows an example of the collation process by the pattern matching method which concerns on Example 4 of this invention. It is a block diagram which shows the structure of the pattern matching apparatus for enforcing the pattern matching method based on Example 6 of this invention. (A), (b) is a figure for demonstrating the three-dimensional figure template by the pattern matching method based on Example 6 of this invention, and its approximation process.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1,71 Pattern matching apparatus 2,72 Object 3 TV camera 4 Digital image generation apparatus 5,75 Storage apparatus 6,76 Image processing apparatus 31 Graphic template 32 Matching target image 51 Grayscale image 52 Edge intensity image 53 Edge direction image 54 Edge Point image 55 Measurement point data 61, 62 Line data 73 Laser three-dimensional sensor 74 Three-dimensional data generators 111-118 Line segments l13-116 Polygonal lines 121-125 Graphic

Claims (9)

A pattern matching method for detecting the recognition target object by matching a graphic template representing the shape of the recognition target object with a verification target image,
A line figure representing the shape of the edge of the image in which the recognition object is photographed by extracting the edge of the image in which the recognition object is photographed from the magnitude of the change in brightness of the image in which the recognition object is photographed A graphic template generation step for generating the graphic template comprising data ;
From the magnitude of change in the brightness of the comparison target image, wherein an edge is extracted in collation symmetrical shape, generates a measurement point data displayed line graphic data representing the shape of the edge of the verification symmetrical shape as a set of measurement points A measurement point data generation step;
A potential space data generation step of generating potential space data in which the score value is recorded so that the score value decreases as the distance from the graphic template increases ;
A score value determining step of collating the measurement point data with the potential space data, and determining a score value corresponding to the position of the measurement point on the potential space data as a score value at the measurement point;
Determination of determining whether or not the verification target graphic is the recognition target based on the similarity by adding the score values at the measurement points to calculate the similarity between the verification target graphic and the graphic template A pattern matching method comprising: and steps. In the template generation step, when there is a missing portion at the edge of the image in which the recognition object is photographed , the direction of the line at both ends of the edge is recorded, and the line at both ends is extended in the direction of the line at both ends, respectively. The pattern matching method according to claim 1, wherein the pattern matching method is repaired by fusing .   3. The pattern matching method according to claim 1, wherein in the graphic template generation step, a graphic template is generated by approximating the image of the recognition target object with a set of equations of the first order or higher. . In the score value determining step, the measurement points are excluded so as to leave the measurement points of the verification target graphic at intervals of half the shortest of the line segments constituting the graphic template. 4. The pattern matching method according to claim 1, wherein a score value is obtained. The score value determining step selects each vertex of the figure to be collated as a measurement point, and obtains a score value for the measurement point. Pattern matching method. The potential space data is composed of a sampling point group representing score values. In the score value determining step, a plurality of sampling points close to the measurement point are selected from the sampling point group, and the score values of these sampling points are used. pattern matching method as claimed in any one of claims 5, wherein the determination of the score value of the measuring points by interpolation based. In the score value determining step, a cluster approximated to the verification target graphic is generated, and a verification target graphic that is not the recognition target object is detected based on the blob information representing the characteristics of the cluster, and a score value is obtained from the target The pattern matching method according to claim 1, wherein the pattern matching method is deleted. In the determining step, it is determined whether or not the verification target graphic is the recognition target object by calculating a cross-correlation value that is a statistical value indicating a similarity between the verification target graphic and the graphic template. The pattern matching method according to any one of claims 1 to 7 . A camera that captures a recognition object; a storage device that stores an image captured by the camera; and a graphic template that represents a shape of the recognition object; A pattern matching device to detect,
  A line figure representing the shape of the edge of the image in which the recognition object is photographed by extracting the edge of the image in which the recognition object is photographed from the magnitude of the change in brightness of the image in which the recognition object is photographed Graphic template generation means for generating the graphic template comprising data;
  The edge of the collation symmetrical figure is extracted from the magnitude of the change in the brightness of the collation target image, and measurement point data is generated in which line figure data representing the shape of the edge of the collation symmetrical figure is displayed as a set of measurement points. Measuring point data generating means;
  Potential space data generating means for generating potential space data in which the score value is recorded so that the score value decreases as the distance from the graphic template increases;
  A score value determining means for collating the measurement point data with the potential space data, and determining a score value corresponding to the position of the measurement point on the potential space data as a score value at the measurement point;
  Determination of determining whether or not the verification target graphic is the recognition target based on the similarity by adding the score values at the measurement points to calculate the similarity between the verification target graphic and the graphic template Means and
  A pattern matching apparatus comprising:

Images