JP5468332B2 - Image feature point extraction method - Google Patents

Description

  The present invention relates to an image feature point extraction method, and in particular, when performing a matching process accurately and at high speed when a change that can be indicated by affine transformation such as rotation, scale, and parallel movement occurs on an object. The feature points that can be obtained by using various operators that are suitable for use are excluded, and points that are unnecessarily extracted are excluded. The present invention relates to a method for extracting image feature points that can be narrowed down to points having high invariance that are easily changed and have little shape error with a matching model.

  In a conventional image processing system, a technique of extracting feature points by a Harris operator (Non-patent Document 1, Patent Document 1) or a SUSAN operator (Non-patent Document 2, Patent Document 2) is employed.

  For example, Patent Document 1 relates to a method of aligning a plurality of images, and selects a feature point pair using a feature point obtained by these operators as a reference point for associating a plurality of images. In addition, image alignment processing is performed.

  Further, Patent Literature 2 is a method for adjusting feature points obtained by these operators, and Patent Literature 3 is a feature point detecting method for improving the problems of these operators.

JP 2008-197917 A JP 2006-190201 A Japanese Patent No. 3764364

C. G. Harris and M. Stephens, “A combined corner and edge detector,” In Proc. 4th Alvey Vision Conf., Manchester, pages 147-151, 1988. S.M.Smith and J.M.Brady, “SUSAN-a new approach to low lavel image processing”, Int. J. Compt. Vis., Vol. 23, no. 1, pp. 45-78, May 1997.

  Harris operators, SUSAN operators, and the like are corner detection operators based on local filtering. Since the corner has a density change in two directions, it has a characteristic that a stable position can be detected and correspondence can be easily taken. Therefore, the corner is often used as a feature point of correspondence in matching processing. Although these operators can certainly capture corners, the same output results are not always obtained for images with all the variations that can be shown by affine transformation such as rotation, scale, and translation. Absent. For example, these operators react not only to the global corners but also to the jagged portions of straight lines (stepped jagged portions generated on the bitmap image) as shown in FIG. 1 for local processing. Often, points that are not corners of human eyes are extracted.

  By extracting such unnecessary points, problems such as an increase in processing time and false detection occur in the matching process.

  In addition, these operators use circular windows and consider invariance to rotation, but the reaction points do not necessarily output the same point.

  Furthermore, depending on the inclination of the object, the response of the feature extractor to the corner may become dull, and necessary points may not be extracted as illustrated in FIG. This is a rare case, and as shown in FIG. 2B, for example, if the inclination changes, for example, a necessary point is almost always extracted. However, if a necessary point is not extracted, a detection error occurs in the matching process, which is a problem.

  Further, for example, in the embodiment of Patent Document 1, the feature point is extracted by the Harris operator. However, as described above, depending on the captured image, the “image image” defined as the feature point in Patent Document 1 may be used. It is not always possible to stably extract “a point that is an edge portion and not a straight edge”, and as illustrated in FIG. 1, it may react to jaggy.

  These feature points used as reference points for aligning a plurality of images should be able to obtain a higher effect by narrowing down to points having higher invariance.

  Further, in Patent Document 2, one learning model image and a plurality of learning input images are captured, feature points and feature amounts are extracted from each, and the learning model image and all learning input images are extracted. In order to maintain accuracy, the feature amount calculation increases as the number of extracted feature points increases, so that matching processing is performed and only feature points that are frequently used for positioning are registered. This time increases and processing time takes. Furthermore, it is desirable that an ideal imaging environment be prepared when capturing a model image for learning. However, in this case, one of the problems of Patent Document 2 is “elimination of the troublesomeness of the operator”. Is not realized. In addition, when performing aggregation processing, matching processing is performed between the learning model image and the learning input image. However, there is a risk of erroneous recognition due to the problems of various operators described above, and the necessary feature points are not necessarily required. May not be extracted correctly.

  The present invention has been made to solve the above-described conventional problems, and eliminates unnecessary extracted points from feature points that can be obtained by using various operators, and is used for matching processing matching and the like. Image feature point extraction that can be narrowed down to points with high invariance that are easy to match, with little shape error with the matching model that changes in the same way following changes in the object It is an object to provide a method.

FIG. 3 shows an outline of the present invention. When the object is tilted as illustrated in FIG. 3A, the corner extractor reacts with the jaggy and many unnecessary feature points are extracted. When this is rotated as illustrated in FIG. 3B, the jaggy portion becomes flat and the corner extractor does not react. Feature points extracted from many images, that is, only points that change following image fluctuation and are extracted from multiple images are necessary points, and only other corner points are excluded by eliminating other points. Can be extracted as

The present invention has been made paying attention to such knowledge. In an image feature point extraction method for extracting feature points from image data, a plurality of pieces of image data are acquired by varying the image, and each image is obtained. Extract feature points for each data, determine the position of the extracted feature points in the image data before the change, select only the feature points that follow the image change and are extracted from multiple images. By eliminating the above, the above-mentioned problem is solved.

  Here, a voting space divided into predetermined areas is set around the expected feature points, and each time the image is changed, the score increases as the distance around the true value increases in the voting space of the image before the change. When the voting for necessary fluctuations is completed, a labeling process is performed on points where the number of votes obtained exceeds a predetermined value in the voting space, and the area of each label block and the maximum number of votes are obtained. Number, total number of votes, and label centroid, and perform threshold processing on the maximum number of votes and average number of votes for a label block to determine whether or not it is a feature point. The center of gravity of the label block can be used as the coordinate of the feature point.

  Further, the feature points can be extracted only for a preset feature extraction region.

  According to the present invention, it is possible to follow various fluctuations that can be indicated by affine transformation such as rotation, scale, and parallel movement at a high speed without bothering the operator as compared with the conventional method. Can be extracted as a feature point, and when used as a corresponding point in positioning processing, there is little shape error with the matching model, it is easy to associate, and an improvement in positioning accuracy can be expected.

  Further, since the total number of feature points can be reduced by preventing the extraction of unstable points, the amount of positioning processing can be reduced and the speed can be increased.

  Furthermore, by preventing the extraction of unstable points, the occurrence of erroneous detection during the positioning process can be suppressed, and the process of erroneous detection inspection can be simplified, so that stable positioning can be performed at high speed.

The figure which shows the example of jaggy which is one of the causes of the conventional feature point false extraction A diagram showing an example of missing points that is one of the conventional problems Diagram showing the principle of the present invention The figure which shows the example of the system configuration for implementing this invention The flowchart which shows the process sequence of 1st Embodiment of this invention. (A) The figure which shows the example of the corner extraction result of the reference image data in the said embodiment, and (B) voting space The figure which similarly shows the example of the fluctuation | variation image of a target object posture The figure which shows the state of voting processing similarly The figure which also shows the example of the label block The figure which also shows the example of excluding the neighboring feature point The figure which shows the outline | summary of 2nd Embodiment of this invention. Flowchart showing processing procedure The flowchart which shows the process sequence of 3rd Embodiment. Figure showing an example of the original image The figure which shows the example of a process when a feature point adjoins similarly The figure which shows the setting example of edge segment and starting point similarly The figure which shows the example of inflection point acquisition by boundary tracking similarly

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

  FIG. 4 is a diagram illustrating a configuration example of a system used to implement the present invention. This system includes a suction nozzle 1 for sucking a target object 2 from which a feature point is extracted and setting it to an imaging position, a machine control device 10 for controlling operations of the suction nozzle 1 and the illumination device 3, and the target object. 2 and the high-resolution camera 5, and an image processing apparatus 20 that processes the captured image.

  The machine control device 10 normally selects the camera 4 or 5 to be imaged according to the size of the object 2, adsorbs the object 2 with the adsorption nozzle 1, and sets it to the imaging position of the selected camera 4 or 5. Further, the illumination device 3 is moved and turned on so that the selected camera can capture an image, and the image processing device 20 is instructed to execute processing together with the selected camera channel information via the interface 26. The image processing device 20 controls the designated camera 4 or 5, picks up an image of the object 2, digitizes it with the A / D converter 21, and stores it in the image memory 22 as multivalued image data. The calculation unit 24 of the image processing apparatus 20 performs various processes on the data in the image memory 22 and extracts feature points from the object 2. On the way, the processing data to be generated is stored in the work memory 23. In the figure, reference numeral 25 denotes a control unit.

  Next, a processing procedure in the image processing apparatus 20 will be described in detail along the flowchart of FIG.

(1) Imaging (step s1)
The image processing device 20 controls the designated camera 4 or 5, picks up an image of the target object 2 in the reference posture, digitizes it with the A / D converter 21, and stores it in the image memory 22 as reference image data.

(2) Feature extraction (here, corner extraction) processing (step s2)
Corner extraction processing is performed on the captured reference image. Corner extraction may be performed using a corner extractor such as Harris or SUSAN. The corner extraction result illustrated in FIG. 6A is stored in the work memory 23.

(3) Voting space creation (step s3)
The working memory 23 is provided with a voting space as illustrated in FIG. The cell size of the voting space is set in consideration of an allowable error (margin) such as a change in imaging time and a shape error. The size of the voting space is set by obtaining an envelope (effective size) of the corner coordinates extracted from the reference image in the X and Y directions and adding a margin. Here, all the valid voting space traps are cleared to zero.

(4) Change the object posture (step s4)
By controlling the suction nozzle 1 with the machine control device 10, the object 2 is given a variation that can be represented by affine transformation, such as rotation, scale, and parallel movement, as illustrated in FIG.

  Specifically, rotation and parallel movement fluctuations can be given to the object 2 by rotating and translating the suction nozzle 1. The rotation angle is desirably finely rotated, for example, by 1 degree, but may be greatly rotated by 10 degrees. In addition, by attaching a zoom lens to the cameras 4 and / or 5 and changing the magnification, scale fluctuations such as enlargement / reduction can be given to the object 2. When a change is given to the object 2 by such a method, it is clear what kind of change is given, and it is necessary to move the object 2 accurately.

  If there is no mechanism for this purpose, the image data may be artificially generated and used by affine transformation of the image data. At this time, an interpolation error is generated by the image interpolation method, and the image quality deteriorates. Therefore, it is desirable to add to the allowable error to be considered in step s3.

(5) Imaging (step s5)
When the magnification of the object 2 and / or the zoom lens is changed in step s4, the image is captured again and stored in the image memory 22 as multi-value image data. Note that when an image is created by affine transformation by image processing, imaging processing is not necessary.

(6) Corner extraction process (step s6)
A corner extraction process is performed on the image data captured in step s5 or the image data generated in step s4. The process is equivalent to step s2.

(7) Corner point coordinate inverse transformation (step s7)
With respect to the coordinates of the corner points extracted in step s6, inverse transformation of the fluctuation given to the object in step s4 is performed to calculate the coordinates in the voting space.

(8) Voting (Step s8)
As illustrated in FIG. 8, based on the voting coordinates obtained in step s7, a vote is voted on the voting space corresponding to the corner coordinates set in step s3.

  In this voting process, weighted voting values are set such that the score decreases with increasing distance from the center of the square corresponding to the corner coordinates extracted in step s2, centering on the true value. It can be set as the process which adds. By adding not only the corner size of the voting space but also the voting value weighting to the corner extractor error and the object shape error, the ability to deal with errors is improved and feature point positions can be extracted with high accuracy. Become.

(9) Aggregation (step s9)
When voting for necessary fluctuations is completed, a labeling process is performed for a point where the number of votes is a predetermined value, for example, 1 or more, in the voting space, and the area of each label block, the maximum number of votes, Calculate the number of votes and the center of gravity of the label. In FIG. 9, the density of the squares indicates the number of votes. The white part has a larger number of votes as the number of votes is 0 and the density is higher. The one with the largest maximum number of votes indicates that it follows the fluctuation with an error within the square of the voting space, and indicates that the point is suitable as a feature point. In FIG. 9, the outline of the label block is shown by a bold line. The area of the label block indicates the magnitude of detection error at the time of fluctuation. When the area of the label block is within the allowable value, the total number of votes / area (average number of votes) is also evaluated. If the area of the label block is larger than the allowable value, it is regarded as an unsuitable feature point and is excluded in advance.

  Thereafter, threshold processing is performed for the maximum number of votes and the average number of votes of the label block, and it is determined whether or not it is a feature point. The coordinates of the feature points are the center of gravity of the label block.

  In addition, the presence of other feature points in the vicinity increases the possibility of erroneous detection, so that it is also assumed that the distance is more than a specified distance. As illustrated in FIG. 10, neighboring feature points within a specified distance of feature points are excluded in order of the number of votes.

  The number of votes of feature points is normalized to a value of 0 to 1 and output together as additional information of feature points. This data can be used as a weighting factor for feature points during matching processing.

  Further, if the preferential association is performed from the feature point of interest having a high weighting coefficient, the possibility of erroneous detection is reduced and the positioning can be performed efficiently.

  Next, as a second embodiment of the present invention, a technique for extracting necessary points in a shorter time than the first embodiment will be described.

  FIG. 11 shows an outline of the second embodiment. In the second embodiment, as illustrated in FIG. 11C, the feature point coordinates as illustrated in FIG. 11B extracted from the reference image as illustrated in FIG. A region A from which feature points can be extracted in the variation image is set as illustrated in FIG. When feature extraction processing as illustrated in FIG. 11E is performed on the set area A and a feature point C is obtained as illustrated in FIG. 11F, a point on the reference image is set as the necessary point C. If a feature point is not obtained, a point on the reference image is removed as an unnecessary point C ′.

  In the first embodiment, the entire region of the fluctuation image is set as the feature extraction target. In the second embodiment, the feature extraction target is limited to the region A, so that the time required for feature extraction is shortened. It becomes possible to perform processing.

  However, in the second embodiment, it is assumed that all necessary points are extracted from the reference image. As shown in FIG. 2A, when a necessary point is not extracted from the reference image, a feature extraction region is not set in the variation image, which causes a detection leak. Therefore, when it is not guaranteed that all feature points are extracted from the reference image, the first embodiment is better.

  Next, the processing procedure of the second embodiment will be described in detail along the flowchart of FIG.

(1) Imaging (step s101)
The image processing apparatus 20 shown in FIG. 4 controls the designated camera 4 or 5, picks up an image of the target object 2 in the reference posture, digitizes it with the A / D converter 21, and stores it as reference image data in the image memory 22. It is memorized (FIG. 11A).

(2) Feature extraction (here, corner extraction) processing (step s102)
The corner C is extracted from the captured reference image (FIG. 11B). Corner extraction may be performed using a corner extractor such as Harris or SUSAN. The corner extraction result is stored in the work memory 23.

(3) Change the object posture (step s103)
By controlling the suction nozzle 1 with the machine control device 10, a change that can be indicated by affine transformation such as rotation, scale, and parallel movement is given to the object 2 (FIG. 11C).

  Specifically, rotation and parallel movement fluctuations can be given to the object 2 by rotating and translating the suction nozzle 1. The rotation angle is desirably finely rotated, for example, by 1 degree, but may be greatly rotated by 10 degrees. In addition, by attaching a zoom lens to the cameras 4 and / or 5 and changing the magnification, scale fluctuations such as enlargement / reduction can be given to the object 2. When a change is given to the object 2 by such a method, it is clear what kind of change is given, and it is necessary to move the object 2 accurately.

  If there is no mechanism for this purpose, the image data may be artificially generated and used by affine transformation of the image data.

(4) Imaging (step s104)
When the magnification of the object 2 and / or the zoom lens is changed in step s103, the image is captured again and stored in the image memory 22 as multi-value image data. Note that when an affine-transformed image is created by image processing, the imaging process in step s104 is unnecessary.

(5) Corner point coordinate conversion (step s105)
The coordinates corresponding to the corner point coordinates extracted in step s102 are obtained from the image captured in step s104 or the affine transformation image created in step s103. At this time, conversion is performed using the fluctuation amount given in step s103, and the association is performed.

  Here, when extracting feature points (corners) of the reference image in steps s103 to s105, the same processing as steps s4 to s8 in the first embodiment is performed, and corner extraction processing by voting is performed.

(6) Extraction area setting (step s106)
A feature extraction area A is set by providing a certain range around the coordinates obtained in step s105 (FIG. 11D). Unlike the first embodiment, the processing time can be shortened by narrowing down the area from which corners are extracted.

(7) Corner extraction process (step s107)
For only the region A set in step s106, the same variation as in steps s4 to s8 in the first embodiment is performed, and corner extraction processing by voting is performed (FIG. 11E). When the corner point C is obtained, it is determined that the point extracted from the reference image is a necessary point. Conversely, if a corner point is not obtained, it is determined that the point extracted from the reference image is an unnecessary point C ′, and is eliminated (FIG. 11 (F)).

  In the above description, corner points are used as feature points, but the types of feature points are not limited to these.

  Next, when the feature points extracted as described above are insufficient, the boundary tracking start point and end point are set in accordance with the shape characteristics of the edge segment, and the feature points by boundary tracking are set. In addition, the processing procedure of the third embodiment devised so as to obtain a more stable result will be described in detail along the flowchart of FIG.

(1) Imaging (step s201)
The image processing apparatus 20 shown in FIG. 4 controls the designated camera 4 or 5, picks up an image of the object 2 ′ in the reference posture as exemplified in FIG. 14, digitizes it with the A / D converter 21, It is stored in the image memory 22 as reference image data.

(2) Edge detection process (step s202)
Edge detection processing is performed on the captured reference image. The edge is detected by applying an edge detection filter such as a Sobel filter or a Gaussian filter and performing threshold processing. An edge detection result with a mark indicating whether or not each pixel is an edge is stored in the work memory 23.

(3) Corner degree calculation (step s203)
For the edge point detected in step s202, for example, a corner extractor such as Harris or SUSAN is used, and for example, a 180 ° folded (reverse) corner is the highest point, and a straight line without a corner is 0 point, The 90 ° corner is calculated as the intermediate point. The calculation result is stored in the work memory 23 as attribute data of each edge point.

(4) Segmentation by labeling (step s204)
The edge points detected in step s202 are labeled, and adjacent edge points are connected to generate several edge segments. Here, as illustrated in FIG. 15, when a plurality of feature points are extracted on the same segment within the vicinity distance, for example, i) the smaller corner degree is ignored depending on the corner degree obtained in step s203. Ii) Combine the features into one feature point by using a weighted average, etc. to prevent the extraction of unstable points.

(5) Edge segment shape characteristic determination processing (step s205)
As illustrated in FIG. 16, the edge segment obtained in step s204 is composed of only four types: an open segment, a closed segment including a broken line and having a corner point, and a curved line, and a directional closed curve segment and a circular segment. Classify.

  Specifically, the edge segment can be classified into two types: a closed segment that can return to the starting point when boundary tracking is performed, and an open segment that ends at the starting point and the ending point.

  Furthermore, the closed segment can be classified into two types, one having a corner and including a polygonal line, and one having only one corner and no curve. In the case of only the latter curve, it can be further classified into two depending on whether the shape has directionality, that is, whether it is a circle or not.

  If it is determined to be a circle, the origin of boundary tracking that is invariant to rotation cannot be set, and using these feature points will act as a noise component that reduces the degree of coincidence during matching. End up. As for a circle, a better result can be obtained if a feature point is not generated on the edge segment, so the center of gravity (= center) is used as the feature point.

(6) Starting point setting at open segment (step s206)
If it is determined in step s204 that the segment is an open segment, it has at least two end points as highly invariant feature points. Since the end point also shows a high value for the corner degree calculated in step s203, the end point may be determined based on this value, and boundary tracking is performed using this point as a temporary starting point, and the next direction to be advanced is inverted by 180 °. By checking whether it is a point to be performed, the end point can be reliably set as the starting point.

(7) Set start point of closed segment including corner point (step s207)
When there is a corner point in the edge segment, the invariance of the starting point can be increased by starting from the point with the highest corner degree, and thereby the invariance of the feature point extracted by boundary tracking can also be increased. . When there are a plurality of points that can be determined to be corner points with high invariance, the points may be divided into a plurality of open segments and processed.

(8) Setting the starting point of a directional closed curve segment (step s208)
Even a closed curve segment without a corner point has directionality, that is, it is not a circle, the principal axis of inertia can be obtained. The intersection of the main axis and the closed curve segment is a point that is highly invariant to rotation. Thus, this closed curve segment can be split and processed at this point into a plurality of open segments.

(9) Inflection point acquisition by boundary tracking (step s209)
By dividing the edge segment into four types in step s205 and performing the processing in steps s206 to s208 for each type, it is possible to result in a continuous cluster of pixel rows from the start point to the end point.

  In step s209, the one-stroke curve data is traced, and a point having a high curvature (a point farthest from the start point to the end point) is extracted.

  Various methods for extracting a point having a high curvature are conceivable. For example, a method as shown in FIG. First, the farthest point in the pixel row is selected with respect to the straight line connecting the starting point and the ending point. At that point, the pixel column is divided into two, and the same processing is recursively repeated for each pixel column. At this time, the minimum value of the distance between the line segment and the point is determined, and if the value is smaller than that value, the recursive process is terminated.

  If this value is increased, a rough polygonal approximate curve can be obtained and narrowed down to points with higher curvature. Moreover, if it is made smaller, an accurate broken line approximation curve including a point with a smaller curvature is obtained, but the number of output points is increased. This value must be adjusted to the shape of the object.

  Further, since the starting point and ending point of the boundary tracking section are highly invariant points, a simple method that generates feature points at equal intervals may be applied.

(10) When the processing from step s205 to step s209 is repeated for the number of edge segments, the processing ends.

  In the third embodiment, edge points are extracted from image data, adjacent edge points are connected and segmented, the shape of the edge segment is determined, and the start and end points of boundary tracking are determined according to the shape of the edge segment. By setting, it is possible to select feature points that are stable with respect to fluctuations that can be indicated by affine transformation such as rotation, scale, and parallel movement, and that can be easily associated.

  Here, the section given by the starting point and the ending point is boundary-tracked, and a point having a high curvature change can be detected as a feature point.

  In addition, the center of gravity of the points in the section given by the start point and the end point can be obtained and used as a feature point.

  An operator that measures the corner degree is applied to the edge point, and the output value is stored as an evaluation value of the feature point, so that it can be determined whether or not the segment is a broken line segment having a clear corner.

  Also, the edge segment boundary can be tracked to determine whether it is a closed segment or an open segment.

  In the case of a closed segment that is not a polygonal line segment, it is determined whether or not it is a circle, and if it is not a circle, it can be processed by being divided into two open segments by the inertia main axis.

  In the case of a circle segment, the feature point is only the center of gravity of the segment, and the feature point on the edge segment can be deleted.

  In addition, when a plurality of feature points are extracted on the same segment within a certain neighborhood distance, they can be combined into one feature point by weighting with the evaluation value (for example, the degree of corner) of the feature points.

  According to the third embodiment, the following effects can be obtained.

(1) By using an invariant point as a starting point for boundary tracking, it is possible to extract a point having high followability as a feature point with respect to fluctuations that can be indicated by affine transformation such as rotation, scale, and parallel movement. When used as a corresponding point, an improvement in positioning accuracy can be expected.

(2) In the case of a circle, there is no invariant point on the segment, so the feature point is only the center of gravity of the segment, and the feature point on the unstable segment is deleted, reducing the total number of feature points and speeding up. can do.

(3) By using the invariant point as the starting point for boundary tracking, it is possible to extract a point having high follow-up as a feature point with respect to fluctuations that can be indicated by affine transformation such as rotation, scale, and parallel movement. The occurrence of false detection can be suppressed, and the process of false detection inspection can be simplified, so that stable positioning can be performed at high speed.

  In any of the above-described embodiments, the present invention is applied to feature point extraction of an image of an object sucked by a suction nozzle. However, the application target of the present invention is not limited to this.

DESCRIPTION OF SYMBOLS 1 ... Suction nozzle 2, 2 '... Object 4 ... Standard camera 5 ... High resolution camera 10 ... Machine control apparatus 20 ... Image processing device 22 ... Image memory 23 ... Working memory 24 ... Calculation part 25 ... Control part 26 ... Interface

Claims (7)

In an image feature point extraction method for extracting feature points from image data,
Acquire multiple image data by changing the image,
Extract feature points for each image data,
Find the position in the image data before the extracted feature points change,
An image feature point extraction method characterized by selecting only feature points that fluctuate following image variation and are extracted from a plurality of images, and exclude other points. Set up a voting space divided into predetermined areas around the expected feature points,
Each time the image is changed, a weighted voting value is added to the voting space of the image before the change so that the score decreases as the distance from the true value increases. In the space, the labeling process is performed for points where the number of votes is greater than or equal to a predetermined value.
Calculate the area of each label block, the maximum number of votes, the total number of votes, the label center of gravity,
Threshold processing is performed for the maximum number of votes and the average number of votes for a label block to determine whether or not it is a feature point.
The image feature point extraction method according to claim 1, wherein for those determined to be feature points, the center of gravity of the label block is used as the coordinates of the feature points.   3. The image feature point extraction method according to claim 1, wherein the feature points are extracted only for a preset feature extraction region.   Extract edge points from image data, connect adjacent edge points to segment, determine edge segment shape, set boundary tracking start and end points according to edge segment shape, add feature points The feature point extraction method according to claim 1, wherein the feature point is extracted.   5. The feature point extracting method according to claim 4, wherein a boundary point is traced between the start point and the end point, and a point having a high curvature change is detected as a feature point.   5. The feature point extraction method according to claim 4, wherein a centroid of points in a section given by the start point and the end point is obtained and used as a feature point.   The operator who measures a corner degree is applied to the edge point, the output value is stored as an evaluation value of the feature point, and it is determined whether or not it is a polygonal line segment having a clear corner. Feature point extraction method.