JP2000231602A - Feature extracting device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To securely extract feature points representing features of a character from a character image. SOLUTION: An image which is segmented into characters by a character extraction part 12 is stored in an image storage part 21 in the feature extraction device 20. An outline point extraction part 22 extracts an array of outline points forming the outline of a character line. A feature point extraction part 23 scans the outline point array to detect the outline point which is at a distance of dots more than a 1st threshold from a point of interest in the outline point array and at the shortest coordinate distance to the point of interest. Further, the feature point extraction part 23 shifts the point of interest and detects the outline point which is at a distance of dots more than the 1st threshold from the shifted point of interest in the outline point array and at the shortest coordinate distance to the point of interest, and then decides the shifted point of interest is a feature point when the distance between the two outline points in the outline point array is larger than a specific threshold.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a feature extracting device for extracting features of characters and the like including the symbols in order to recognize the characters and symbols.

[0002]

2. Description of the Related Art Character recognition methods are roughly classified into two types, that is, pattern matching and structure analysis.
The pattern matching is suitable for recognizing a print character or the like with little deformation, but it is difficult to identify a similar character or the like. Therefore, when particularly high recognition accuracy is required, a structural analysis method is used. In the structural analysis method, a feature of a positional relationship of a feature point of a reference character is registered in a dictionary in advance, and a feature part is extracted as a feature point from an input character pattern, and the extracted feature point is extracted. Characters are recognized using a dictionary for the positional relationship of.

Conventionally, as a technique relating to such feature point extraction, there is a technique described in the following document, for example. Reference: JP-A-8-77293 In the character recognition device described in the above reference, a direction code sequence of character outline points is created, and a difference between the direction code sequences is determined. It is extracted as character feature points. In other words, the feature point is where the amount of change in the contour point is large. As a feature point of handwritten characters,
It suffices if an end point, a branch point, an intersection, a vertex in a steep arc, etc. can be extracted.

[0004]

However, the feature extraction performed by the character recognition device and the like described in the above document has the following problems. FIG. 2 is a diagram showing a problem in the conventional feature extraction. In the character recognition device of the document,
A feature point can be stably extracted in a portion where the direction of a contour point changes abruptly, such as end points 1, 2, 3, 4 and intersections 5, 6 of a segment constituting a character. In some cases, a portion where the direction change of a contour point is small is not extracted as a feature point. In order to extract the intersections 7 and 8 as feature points, it is necessary to lower the threshold value so that the direction change amount of the contour point can be extracted as a feature point to some extent. However, if the threshold value is lowered, for example, a portion that is not suitable as a feature point such as a vertex that draws a gentle arc will also be a feature point.
There is also a problem that it is difficult to set the optimum threshold. Furthermore, since it is necessary to create a direction code string for all contour points, there is a problem that the processing time becomes enormous when the resolution is sufficiently high and the number of contour points increases.

[0005]

[MEANS FOR SOLVING THE PROBLEMS]
According to a first aspect of the present invention, a character line is drawn.
Input an image of a character and represent the characteristics of the character in the image
In the feature extraction device that extracts feature points, the following image
Image storage unit, contour point extraction unit and feature point extraction unit
You. The image storage unit stores the image extracted in units of one character.
The image is input and stored. The contour point extraction unit
Forming an outline of the character line in the stored image
This is for extracting a contour point sequence. The feature point extracting unit is
The outline point sequence extracted by the outline point extraction unit is scanned, and the outline point
Attention point C which is 1_nFrom the number of dots on the contour point sequence
Point of interest C_nAt coordinate distance from
The closest contour point NR_nIs detected, and
Attention point C_nFrom the outline point sequence by one dot
Attention point C_{n + 1}And then extracted by the contour point extraction unit.
Scans the contour point sequence_{n + 1}From the contour point sequence
Are separated by more than the first threshold value TH1 with the number of dots of
Point C_{n + 1}Contour point NR at the nearest coordinate distance from
_{n + 1}And the contour point NR_nAnd contour point NR_{n + 1}Between
Is larger than the second threshold value TH2
The point of interest C_{n + 1}Is the feature point.
You.

According to the first aspect of the present invention, the outline point sequence forming the outline of the character line is extracted by the outline point extracting unit. Contour points NR _n , NR _{n + 1 at} positions closest to the points of interest C _n , C _{n + 1} apart from the points of interest C _n , C _{n + 1} adjacent to each other in the line of contour points by the first threshold value, respectively. If the distance between these contour points on the contour point is equal to or greater than the second threshold, the point of interest C
_{n + 1} is extracted as a feature point. That is, the feature point is obtained from the distance, not the change in the direction of the outline point sequence.

According to a second aspect, in a feature extraction device,
In addition to the same image storage unit and contour point extraction unit as in the first invention, the following line width detection unit and feature point extraction unit are provided. The line width detection unit detects a line width of the character line of the stored image. The feature point extraction unit is
A first threshold TH1 and a second threshold T corresponding to the line width
Determine the H2, scanning the contour point string extracted by the contour point extracting section, the first threshold value or more from the target point C _n by the number of dots on the contour point row away and coordinates from the noted point C _n The contour point NR _n at the closest distance is detected, and the point of interest is further shifted from the point of interest C _n to the point of interest C _{n + 1} shifted by one dot on the line of contour points. The contour point sequence extracted by the section is scanned, and is separated from the point of interest C _{n + 1 by the} number of dots on the contour point row by the first threshold value TH1 or more and the point of interest C
_The contour point NR _{n + 1} which is the closest coordinate distance from _{n + 1}
And when the number of dots on the contour point sequence between the contour point NR _n and the contour point NR _{n + 1} is larger than the second threshold value TH2, the target point C _{n + 1 is set} to the feature point. It is assumed that.

By adopting such a configuration, the line width is detected by the line width detecting section. The first and second thresholds TH1 and TH2 corresponding to the line width are extracted by the feature point extracting unit.
Is determined, and a feature point is extracted using the first and second threshold values TH1 and TH2.

According to a third aspect, in the feature extracting device,
In addition to the same image storage unit and contour point extraction unit as in the first invention, the following thinning scanning threshold value detection unit and feature point extraction unit are provided. The thinning-out scanning threshold detecting unit determines a first thinning-out scanning threshold MTH1 and a second thinning-out scanning threshold MTH2 according to the amount of contour points extracted by the contour point extracting unit. The feature point extracting unit converts the contour point sequence extracted by the contour point extracting unit into the second thinning scan threshold M.
Scans every TH2, to detect contour points NR _n in away from the target point C _n by the number of dots on the contour point row first threshold TH1 or higher and from the noted point C _n to the closest distance at coordinates distance,
Furthermore, attention point shifted the first thinning scanning threshold MTH1 minutes point of interest from the noted point C _n on the contour point sequence C _{n + MTH1}
And then scans the outline point sequence extracted by the outline point extraction unit for each of the second thinning-out scanning thresholds MTH2, and _{calculates the} number of dots on the outline point sequence from the point of _{interest Cn + MTH1} . Threshold T of 1
A contour point NR _{n + MTH1} which is more than H1 away from the point of _interest C _{n + MTH1} and is closest to the coordinate distance is detected, and the contour point N
When the number of dots on the contour point sequence between R _n and the contour point NR _{n + MTH1} is greater than the second threshold value TH2, the point of _interest C
_{n + MTH1} is the feature point.

By adopting such a configuration, the first and second thinning-out scanning thresholds MTH1 and MTH2 according to the amount of the contour points are determined by the thinning-out scanning threshold detecting unit. The feature point extracting unit detects feature points by thinning scanning using the first and second thinning scanning thresholds MTH1 and MTH2.

According to a fourth aspect, in the feature extracting device,
In addition to the same image storage unit, contour point extraction unit and line width detection unit as in the second invention, the following thinning-scanning threshold detection unit and feature point extraction unit are provided. The thinning-scanning-threshold detector detects a first thinning-scanning threshold M corresponding to the line width.
TH1 and a second threshold value MTH2 are determined.
The feature point extracting unit may determine a first threshold value TH1 corresponding to the line width.
And a second threshold value TH2, and the contour point sequence extracted by the contour point extraction unit is subjected to the second thinning scanning threshold value MTH2.
Scans each, contour from which is one from the target point C _n apart first threshold TH1 or more in the number of dots on the contour point sequence and the noted point C _n of the contour points in the closest coordinate distance Point N
Detecting a R _n, further, in the contour point extracting unit after having transferred target point from target point C _n to the target point C _{n + MTH1} shifted first thinning scanning threshold MTH1 minutes on the contour point sequence The extracted contour point sequence is scanned for each of the second thinning-out scanning thresholds MTH2, and is _separated from the _target point C _{n + MTH1 by the} number of dots on the contour point sequence by the first threshold value TH1 or more and the target point C _The contour point NR _{n + MTH1} which is closest to the coordinate distance from _{n + MTH1} is detected, and the number of dots on the contour point sequence between the contour point NR _n and the contour point NR _{n + MTH1} is determined by the second When the value is larger than the threshold value TH2, the point of _{interest Cn + MTH1} is _set as the feature point.

By adopting such a configuration, the line width of the character in the image is detected by the line width detecting unit, and the first and second thinning-out scanning thresholds MTH1, MTH1 corresponding to the line width are detected.
MTH2 is determined by the thinning-out scanning threshold value detection unit. The feature point extracting unit performs thinning-out scanning using the first and second thinning-out scanning thresholds MTH1 and MTH2 corresponding to these line widths, and the first and second thresholds TH1 and T corresponding to the line widths.
The evaluation of the contour points using H2 is performed, and the feature points are detected.

According to a fifth aspect, in the feature extracting device,
In a feature extraction device for inputting an image of a character drawn by a character line and extracting a feature point representing a feature of the character in the image, an image storage unit, a contour point extraction unit,
A line width detection unit and a thinning-out scanning threshold value detection unit are provided, and a feature point extraction unit similar to that of the fourth aspect is provided. By employing such a configuration, the first and second thinning scanning thresholds MTH1 and MTH2 according to the amount of the contour points are determined by the thinning scanning threshold detecting unit. The feature point extraction unit uses the first and second thinning-out scanning thresholds MTH1 and MTH2 according to the amount of these contour points, and uses the first and second thresholds TH1 and TH2 according to the line width. The contour points are evaluated and feature points are detected.

According to a sixth aspect, in the feature extraction device according to the first or third aspect, a contour point R corresponding to the feature point is provided.
_n , a contour point R _{n + TH3} that has advanced the contour point sequence in the forward direction by a preset third threshold value _TH3 and the third threshold value T
The contour point R _n-TH3 returned in the opposite direction by H3 is obtained, and from the positional relationship between the contour point R _n , the contour point R _{n + TH3} and the contour point R _n-TH3 , the contour point R _n of the character is obtained. An unevenness detecting unit for detecting whether the position is convex or concave is provided. By employing such a configuration, it is possible to detect not only the feature points but also information as to whether the feature points are concave or convex in the character.

According to a seventh aspect, in the feature extracting apparatus according to the second, fourth or fifth aspect, a third threshold value TH3 is determined from the line width detected by the line width detecting section, and the characteristic point is determined. contour point 8 from this contour point R _n by the threshold TH3 of the third of the contour point sequence returned by the opposite side threshold TH3 contour point R _{n + TH3} and the third advanced in the forward direction R _{n- TH3} and the contour points R _n , R _{n + TH3} and R _n-TH3
From the positional relationship between the position of the contour point R _n in the character it is provided the concave-convex detecting part for detecting whether become concave or has a convex. By employing such a configuration, it is possible to detect not only the feature points but also information as to whether the feature points are concave or convex in the character.

[0016]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS First Embodiment FIG. 1 is a configuration diagram of a character recognition device provided with a feature extraction device according to a first embodiment of the present invention. This character recognition device includes, for example, a character cutout unit 12 that receives a binary image from an image storage medium 11 and cuts it out in units of one character, and a character recognition unit 13 that performs character recognition. A feature extracting device 20 is connected between the character recognition unit 12 and the character recognition unit 13.

The feature extracting device 20 includes an image storage 21 for storing the binary image provided from the character extracting unit 12. On the output side of the image storage unit 21, a contour point extraction unit 22 for extracting a contour point sequence representing a contour of a character line, and scans the extracted contour point sequence to set a characteristic contour point of a character as a characteristic point. Feature point extracting unit 23 to be extracted, and feature point extracting unit 23
And a concavo-convex detecting unit 24 for detecting whether the feature point extracted in step (1) is a convex feature point or a concave feature point. The output side of the unevenness detection unit 24 is connected to the input side of the character recognition unit 13.

Next, the operation of the character recognition device will be described. The character extracting unit 12 receives the binary image from the image recording medium 11 storing the binary image such as a document or a form in which the character is described, and cuts out the binary image on a character-by-character basis.
Give to 0. The image storage unit 21 in the feature extraction device 20
The binary image is received from the character extracting unit 12 in units of 1 character and stored. The contour point extraction unit 22 receives a binary image in units of one character from the image storage unit 21 and extracts a contour point of the binary image.

FIGS. 3A and 3B are diagrams showing an outline of outline point extraction performed by the outline point extraction unit 22 in FIG. The contour points can be extracted by a well-known translocation tracking method. In the translocation tracking method, as shown in FIG. 3A, the upper left of a binary image is set as a scanning start point, and the entire image is sequentially scanned. At this time, when a black pixel in contact with a white pixel is detected, as shown in FIG. 3B, the contour is extracted by tracing the black pixel in a counterclockwise direction. A black pixel that is in contact with a white pixel is one in which a white pixel exists at any of the upper, lower, left, and right sides of the black pixel. Specifically, the contour point extracting unit 22 sets the black pixel b0 that is in contact with the white pixel detected first in the scanning process as the start point of the contour tracking, and then sets the lower left pixel of the pixel b0 around the pixel b0 as the start point. Then, a black pixel is detected in the counterclockwise direction. Thereby, the black pixel b1 at the contour point is detected. When the pixel b1 is detected, the pixel b1 is centered on the pixel b1 and the pixel b0 is detected.
A black pixel is detected in the counterclockwise direction starting from. Thereby, the black pixel b2 at the contour point is detected. The process of following the contour is repeated until the process returns to the start point b0,
Find the outline point sequence.

FIG. 4 is a diagram showing a configuration example of the contour table. The contour point extraction unit 22 that tracks the contour points uses the tracked contour points as contour data, and uses them as contour point numbers as shown in FIG. 4, coordinates of the contour points, and whether the contour points are feature points. Is written in a contour table including columns (feature points) indicating
However, since the feature point has not been extracted yet at this time, “0” indicating that the contour point is not a feature point is entered in the column indicating whether or not the feature point.

FIGS. 5A and 5B are diagrams showing an example of a character configuration. Depending on the character, there are a separated character as shown in FIG. 5A and a character including a loop as shown in FIG. In these, a contour table is created for each of the independent segments Sg1 and Sg2, and information on each contour point is written. The contour point extracting unit 22 gives the created contour table to the feature point extracting unit 23. The feature point extracting unit 23 receives the contour table created by the contour point extracting unit 22, and extracts a feature point by the following processing.

FIGS. 6A and 6B show the characteristic point extracting unit 23.
Is a diagram showing a procedure of feature point extraction performed by
The feature point extraction at the end point of the character, and FIG.
Are respectively shown. FIG. 7 (a),
(B) is a diagram showing an outline of feature point extraction. Feature point extraction
The output unit 23 outputs the first contour point R in the contour table.₁Select
Attention point C₁And the point of interest C₁X coordinate of₁, And Y
Coordinate Y₁Contour point NR satisfying a certain condition₁Detect
You. Next, the attention point is shifted by one, and the contour point R is shifted._TwoSelect
Attention point C _TwoSimilarly, the contour point NR_TwoIs detected. this
When contour point NR₁And NR_TwoWhen certain conditions are met
Attention point C_TwoIs determined as a feature point. In addition,
Advance one to contour point R_ThreeThe attention point C_ThreeAnd the same process
Do The point of interest is the last contour point R_eRepeat until
You. Contour point R_nThe attention point C_n6 and FIG.
7 will be described in detail.

The contour table is scanned from beginning to end,
From X-coordinate X _n and the Y coordinate Y _n of the noted point C _n, detects the following conditions contour point NR _n. Condition 1: dot distance on the contour point sequence between the contour point and the target point C _n are separated from the first threshold value TH1. Condition 2; attention point C among contour points satisfying (condition 1)
_The contour point whose coordinate distance is closest to the coordinates ( _Xn , Yn) of _n

[0024] Assuming the case where the threshold value TH1 determined in advance and 6, as shown in FIG. 7 (a), the the target point C _n, distant contour point than the threshold value TH1 dot distance along the contour point row,
Contour points R _n-7 , R _n-8 ,... And contour points R _{n + 7} , R _{n + 8} ,
... These outline points _{R n-7, R n-} 8, ... and contour point _{R n + 7, R n +} 8, ... among the contour points those actual coordinates distance from the target point C _n is closest It is detected as NR _n . For example, since the contour point R _n-7 is shifted two dots in the vertical direction and seven dots in the horizontal direction with respect to the point of interest C _n , the actual coordinate distance is about 7.28 (= √ (2 ² +7 ² )). ).
On the other hand, the contour point R _{n + 7} is shifted two dots in the vertical direction and six dots in the horizontal direction with respect to the point of interest C _n , so that the actual coordinate distance is about 6.32 (= √ (2 ² +6 ^2). )). Thus, the contour point R _{n + 7} is a contour point NR _n.

Contour point NR_nFeature point extraction unit
23 indicates that the contour point R_{n + 1}The attention point C
_{n + 1}Similarly, using the above conditions, the contour point NR_{n + 1}Seeking
Confuse. In this case, as shown in FIG.
_{n + 1}From the threshold TH1 with the dot distance along the contour point sequence
The contour point farther away is the contour point R_n-6, R_n-7, ... and rings
Guo point R_{n + 8}, R_{n + 9}, ... These conditions 1 are satisfied
Among the contour points, the contour point R _n-6Attention point C_{n + 1}Seat from
The target distance is about 6.71 (= √ (3^Two+6^Two)).
Contour point R for this_{n + 8}Attention point C_{n + 1}Coordinate distance from
The separation is about 7.07 (= √ (1^Two+7^Two)). Yo
And contour point R_n-6Is the contour point NR_{n + 1}Becomes Contour point NR
_nAnd NR_{n + 1}The feature point extraction unit 23 that has detected the
Table, the contour point NR_nAnd NR_{n + 1}Contour between
The number of contour points (the number of dots) on the point sequence is determined, and FIG.
As shown in (b), the number of contour points is a predetermined second threshold value.
If it is larger than TH2 (for example, 7), the contour point C_{n + 1}(R
_{n + 1}) Is determined as a feature point.

[0026] Here, the number of contour points on the contour point sequence between contour points NR _n and NR n _{+ 1} is the contour point follows the contour point sequence in the forward direction as a starting point the contour points NR _n NR n ₊ the contour points of the route to _1, 2 contour points of a path leading to the contour point NR _n follows the contour point sequence in the forward direction starting from the contour points NR n _{+ 1}
Although there are types, the one with the smaller number of contour points is used among the two types. Feature point extraction unit 23, further, the R _{n + 2} as the target point C _{n + 2} point of interest one advanced, performs the same processing hereinafter target point is the last contour point R _e of the contour table Repeat until Incidentally, in the case greater than the threshold value TH2 contour points on the contour are determined in advance between the contour points NR ₁ and contour point NR _e is, C ₁ is also a feature point.

For a character having a plurality of segments (a character having a plurality of contour tables), the above processing is performed for each contour table of each segment. Further, the feature point extracting unit 23 sets “0” in a column indicating whether or not the feature point of the contour table is a feature point with respect to the contour point extracted as the feature point,
It is replaced with “1” to indicate that it is a feature point, and the contour table is passed to the unevenness detecting unit 24.

FIG. 8 is a diagram showing scanning of the contour table by the unevenness detecting section 24. In order to indicate the state of the extracted feature point, the unevenness detection unit 24 determines whether the feature point is a convex feature point or a concave feature point as follows. The convex feature point is present at the protruding portion of the character, and corresponds to an end point or an outer vertex in which a sharp arc is drawn. In addition, the concave feature point exists in the concave portion of the character, and corresponds to a branch point, an intersection, or an inner vertex in which a sharp arc is drawn. The concavity and convexity detecting unit 24 scans the given contour table, and when a contour point serving as a feature point is found, as shown in FIG. 8, the contour point sequence is forwardly shifted by a predetermined threshold TH3 from the contour point of the feature point. And the contour point with the number returned in the opposite direction of the contour point sequence by the threshold value TH3 are detected.

[0029] For example, when a contour point is determined to be the feature point and R _n, the concave-convex detecting part 24, the contour point R _n of coordinate (X _n, Y _n) and the threshold from the contour point R _n The coordinates ( _{Xn + TH3} , Y) of the contour point Rn _{+ TH3} , which has advanced in the forward direction by _{TH3
n + TH3} ) and the contour point R advanced in the opposite direction by the threshold value TH3.
_n-TH3 of coordinates and _{(X n-TH3, Y n} -TH3), detects from the contour table. Incidentally, when detecting the contour points contour point advanced from R _n just forward threshold TH3, when that would exceed the last contour table continues to detect back first of the contour table. Further, when detecting the contour points back in the opposite direction only contour points threshold TH3 from the contour point R _n, if that would exceed the first contour table,
The detection is continued from the end of the contour table.

When the coordinates of the three points are determined, the following equations (1) to (5) are calculated based on the coordinates. _{x1 = X n + TH3 -X n} -TH3 ··· (1) y1 = Y n + TH3 -Y n-TH3 ··· (2) x2 = X n -X n-TH3 ··· (3) y2 = Y _n -Y _n-TH3 (4) x1 × y2-y1 × x2 <0 (5) If the equation (5) is satisfied, the feature point R _n is a concave feature point; Judge as a point. If the judgment result is convex feature point, to indicate this, writing "1", for example in the field of the feature points of the contour point R _n, if the concave feature point writing "2".

FIG. 9 is a diagram showing detection results of concave feature points and convex feature points. The unevenness detection unit 24 performs the above (1) to
The processing using the equation (5) is performed on all the contour points extracted as the feature points. For a character having a plurality of segments (a character having a plurality of contour tables), the same processing is performed for each contour table. When these processes are performed on the image of FIG. 2, for example, as shown in FIG.
The positions of the concave feature points and the convex feature points are obtained. Irregularity detector 2
4 gives the character recognition unit 13 a contour table to which information of the concave feature points and the convex feature points is added. The character recognition unit 13 performs character recognition while referring to contour points and feature points in the contour table.

As described above, in the first embodiment,
A feature point extraction unit 23 is provided in a feature extraction device 20 including an image storage unit 21 and a contour point extraction unit 22, and an arbitrary point of interest C _n
Dot distance on the contour point sequence between the leaves than the threshold TH1, and detects the coordinate distance is closest contour point NR _n from the noted point C _n, further one of the target point C _n shift dot distance on the contour point sequence between the target point C _{n + 1} leaves than the threshold value TH1, and to detect the contour points NR n _{+ 1} coordinate distance closest from the noted point C _{n + 1} When the number of dots on the contour point sequence between the detected contour points NR _n and NR _{n + 1} is larger than a predetermined threshold TH2, the target point C
It is configured to detect _{n + 1} as a feature point. For this reason, the feature extracting device 20 not only extracts the end points and the steep branches and intersections as the feature points, but also does not extract the gentle arc portion as the feature point. A feature point can be reliably extracted from a branch point or an intersection where there is little change. Further, the feature extracting device 20 is provided with the unevenness detecting unit 24 and detects whether the feature point is a convex feature point or a concave feature point, so that the process suitable for the character recognition performed by the character recognition unit 13 at the subsequent stage can be performed. Will be able to do it.

Second Embodiment FIG. 10 is a block diagram of a character recognition device provided with a feature extraction device according to a second embodiment of the present invention. This character recognition device includes an image storage medium 11, a character cutout unit 12, and a character recognition unit 13 similar to those of the first embodiment, and is connected between the character cutout unit 12 and the character recognition unit 13. A feature extraction device 30 different from that of the first embodiment is provided. As in the first embodiment, the feature extracting device 30 includes an image storage unit 31 that stores the binary image given from the character extracting unit 12.
And an outline point extraction unit 32 connected to the image storage unit 31. The output side of the image storage unit 31 further includes:
A line width detection unit 33 that detects a line width W of the binary image stored in the image storage unit 31 is connected.

A feature point extracting section 34 is connected to the output sides of the contour point extracting section 32 and the line width detecting section 33, and an unevenness detecting section 35 is connected to the output sides of the feature point extracting section 34 and the line width detecting section 33. It is connected. The feature point extracting unit 34 scans the contour points and extracts characteristic portions of the character as feature points, and the unevenness detecting unit 35 determines whether the feature points extracted by the feature point extracting unit 34 are convex feature points. Or a concave feature point. The output side of the unevenness detection unit 35 is connected to the input side of the character recognition unit 13.

Next, the operation of the character recognition device shown in FIG. 10 will be described. The character extracting unit 12 receives a binary image from the image recording medium 11 in which a binary image such as a document or a form in which characters are described is stored, and cuts out the binary image in units of one character. The image storage unit 31 receives and stores a binary image in units of one character from the character extraction unit 12. The contour point extracting unit 32 receives the binary image for each character from the image storage unit 31 and extracts the contour points of the binary image by the same processing as in the first embodiment. On the other hand, the line width detection unit 33 detects the line width W in the binary image by referring to the binary image of one character unit from the image storage unit 31. There are several methods for detecting the line width W. For example, the following method is used.

FIG. 11 is a view showing an observation window for detecting a line width. The line width detection unit 33 includes a0 to a as shown in FIG.
Using an observation window of 2 rows and 2 columns, a raster scan is performed using an observation window of 2 rows and 2 columns while observing all pixels constituting the image data stored in the image storage unit 31 with the observation window a0. Then, the image storage unit 3 seen from the observation window of 2 rows and 2 columns
When all pixels on one memory are pixels (black pixels) constituting a character, the number of four black points Q is sequentially counted so as to increase the count by one. When the pixel on the memory of the image storage unit 31 viewed from the observation window a0 is a pixel constituting a character, the total number of black spots A is sequentially counted so as to increase the count by one. However, the initial values of the number of four black points Q and the total number of black points A are 0.

Next, the line width W of the binary image is calculated by the following equation (6). W = A / (AQ) (6) By the way, the line width W of the character shown in FIG.
= 2.5.

Next, the feature point extracting section 34 receives the contour table created by the contour point extracting section 32 and the line width W detected by the line width detecting section 33, and determines the characteristic points as follows. Extract. First, the feature point extracting unit 34 uses the line width W to determine the threshold values TH1 and TH2 in the following (7) and (8).
It is calculated by the formula. TH1 = W × K1 (7) TH2 = W × K2 (8) where K1 and K2 are constants. For example, if K1 is 2.5 and K2 is 3.0, the threshold value TH1 is 6. The threshold value TH2 is 7 (rounded down after the decimal point). If the threshold values TH1 and TH2 are calculated, the subsequent processing is the same as that of the feature point extracting unit 23 of the first embodiment, and a detailed description will be omitted. When the feature points are extracted by the feature point extraction unit 34, the contour table in which the feature points are written is sent to the unevenness detection unit 35.

The unevenness detecting section 35 receives the contour table from the feature point extracting section 34, and receives the line width W from the line width detecting section 33.
From the line width W to the threshold value TH3 according to the following (9).
Is calculated. TH3 = W × K3 (9) where K3 is a constant. If K3 is 2.0, the threshold value TH3 in the character shown in FIG.
Is 5 (rounded down after the decimal point). If the threshold value TH3 is calculated, the subsequent processing performs the same processing as that performed by the unevenness detection unit 24 of the first embodiment. The unevenness detecting section 35 gives the contour table to the character recognizing section 13, and the character recognizing section 13 recognizes the character while referring to the contour points and characteristic points of the contour table.

As described above, in the second embodiment,
A line width detection unit 33 is provided, and the feature point extraction unit 3 is provided in accordance with the line width W.
The threshold values TH1 and TH2 used in step No. 4 are calculated, and feature points are extracted. Therefore, it is not affected by the line width of the character of the binary image, and it is possible to always extract a stable feature point even if the resolution of the input pattern changes. further,
Since the threshold value TH3 used by the unevenness detection unit 35 is also calculated according to the line width W, and the unevenness of the feature point is detected, similarly, the stable unevenness information of the feature point can be detected.

Third Embodiment FIG. 12 is a block diagram of a character recognition device provided with a feature extraction device according to a third embodiment of the present invention. This character recognition device includes an image storage medium 11, a character cutout unit 12, and a character recognition unit 13 similar to those of the first embodiment, and is connected between the character cutout unit 12 and the character recognition unit 13. A feature extraction device 40 different from that of the first embodiment is provided. As in the first embodiment, the feature extraction device 40 includes an image storage unit 41 that stores the binary image given from the character cutout unit 12.
And an outline point extraction unit 42 connected to the image storage unit 41.

The contour point extracting section 42 is connected to a thinning-out scanning threshold value detecting section 43, and a feature point extracting section 44 is connected to the output side of the contour point extracting section 42 and the thinning-out scanning threshold value detecting section 43. The thinning-scanning-threshold detecting unit 43 determines a thinning-scanning threshold according to the amount of the contour points extracted by the contour-point extracting unit 42. The feature point extracting unit 44 scans the contour points extracted by the contour point extracting unit 42 using the threshold value detected by the thinning scanning threshold value detecting unit 43, and extracts a characteristic portion of the character as a feature point. . Feature point extraction unit 4
4 is connected to an unevenness detecting unit 45 for detecting whether the feature point extracted by the feature point extracting unit 44 is a convex feature point or a concave feature point, and the output side of the unevenness detecting unit 45 is connected to the character recognition unit 13. Is connected to the input side.

Next, the operation of the character recognition device shown in FIG. 12 will be described. The character extracting unit 12 receives a binary image from the image recording medium 11 in which a binary image such as a document or a form in which characters are described is stored, and cuts out the binary image in units of one character. The image storage unit 41 receives and stores the binary image clipped from the character clipping unit 12. The contour point extracting unit 42 receives the binary image of one character unit from the image storage unit 41, and extracts the contour points of the binary image by the same processing as in the first embodiment. The thinning scanning threshold detecting unit 43 receives the contour table created by the contour point extracting unit 42, and counts all the contour points. For a character having a plurality of segments (a character having a plurality of contour tables), the number of contour points of all the contour tables is counted. Let the count value be Rsum.

The thinning-scanning-threshold detecting unit 43 that has obtained the count value Rsum calculates the first and second thinning-scanning thresholds MTH1 and MTH2 according to the following equations (10) and (11). MTH1 = Rsum / N1 (Truncation below decimal point) (10) MTH2 = Rsum / N2 (Truncation below decimal point) (11) where N1, N2; constant (for example, 100) When MTH1 = 0, MTH1 = 1 When MTH2 = 0, MTH2 = 1

The feature point extractor 44 is provided with the contour point extractor 42
The contour table is input from the
Receiving the thinning-out scanning threshold values MTH1 and MTH2 from
Feature points are extracted by the following processing. First, the contour tape
First contour point R₁Select the point of interest C₁And the contour
The point table is set from the beginning for every thinning-out scanning threshold MTH2.
Draw and scan the point of interest C ₁X coordinate of₁And Y coordinate
Y₁From the contour point NR satisfying the following condition₁Is detected. Condition 1: the contour point and the attention point C₁On the sequence of contour points between
The dot distance is greater than the first threshold TH1. Condition 2; attention point C among contour points satisfying (condition 1)
₁Coordinates (X₁, Y₁) Is, for example, six.

Next, the contour point R _{1 + MTH1} is _set to the attention point C _{1 + MTH1} by advancing the point of interest to the thinning threshold MTH1, and the contour point NR _{1 + MTH1} is similarly obtained using the above conditions. Then, a NR ₁ and NR 1 _{+ MTH1} contour points on the contour points between the (number of dots) with reference to the contour table, if the number of contour points is larger than the threshold TH2 is set to, for example 7, C 1 _{+ MTH1} Is determined as a feature point. Note that the number of contour points on the contour point sequence between the contour points NR ₁ and NR _{1 + MTH1} includes NR ₁ as a starting point and NR ₁
NR ₁ contour points and NR 1 _{+ MTH1} to _{1 + MTH1} as the starting point
, There are two types of contour points, and one of the two types having the smaller number of contour points is adopted.

After _{finding the} feature point C _{1 + MTH1} , the point of interest is further advanced by MTH1 to make R _{1 + MTH1x2 a} point of interest C1 _{+ MTH1x2,} and the same process is performed. Repeat until reaching. However, for a character having a plurality of segments (a character having a plurality of contour tables), the same processing is performed for each contour table. The feature point extracting unit 44 replaces the contour points extracted as feature points with “1” in the feature point column of the contour table to indicate that the feature points are feature points, and replaces the contour table with the unevenness detecting unit 45. Pass to. The unevenness detection unit 45 performs the same processing as in the first embodiment, and gives the contour table to the text recognition unit 13. The character recognizing unit 13 recognizes characters while referring to contour points and feature points in the contour table.

As described above, in the third embodiment,
Since the thinning-scanning threshold detection unit 43 is provided to perform thinning-scanning in accordance with the amount of contour points when extracting feature points, the processing time can be reduced even when the number of contour points increases when resolution is high. And feature points can be extracted. In addition,
Simply performing the thinning-out scanning may slightly shift the feature points. When the resolution is high, a slight shift is not noticeable, but when the resolution is low, a slight shift greatly deviates from a characteristic portion. However, in the third embodiment, since the thinning amount is determined according to the amount of the contour points, the problem can be dealt with.

Fourth Embodiment FIG. 13 is a block diagram of a character recognition device provided with a feature extracting device according to a fourth embodiment of the present invention. This character recognition device includes an image storage medium 11, a character cutout unit 12, and a character recognition unit 13 similar to those of the first embodiment, and is connected between the character cutout unit 12 and the character recognition unit 13. A feature extraction device 50 different from that of the first embodiment is provided. As in the first embodiment, the feature extraction device 50 includes an image storage unit 51 that stores the binary image given from the character extraction unit 12.
And an outline point extraction unit 52 connected to the image storage unit 51. The output side of the image storage unit 51 further includes:
A line width detection unit 53 that detects a line width W of the binary image stored in the image storage unit 51 is connected. Line width detector 53
Is connected to a thinning-scanning-threshold detecting unit 54 that detects a thinning-scanning threshold in accordance with the value of the line width W.

On the output side of the contour point extraction unit 52, the line width detection unit 53, and the thinning scanning threshold value detection unit 54, a feature point extraction unit 55
Is connected to an output side of the line width detection unit 53 and the feature point extraction unit 55. The feature point extraction unit 55 uses the thinning threshold value provided from the thinning scanning threshold value detection unit 54 to scan contour points and extract characteristic portions of characters as feature points.
Numeral 6 detects whether the feature point extracted by the feature point extraction unit 55 is a convex feature point or a concave feature point. The output side of the unevenness detection unit 56 is connected to the input side of the character recognition unit 13.

Next, the operation of the character recognition device shown in FIG. 13 will be described. The character extracting unit 12 receives a binary image from the image recording medium 11 in which a binary image such as a document or a form in which characters are described is stored, and cuts out the binary image in units of one character. The image storage unit 51 receives a binary image in units of one character from the character extracting unit 12 and stores the binary image. The contour point extraction unit 52 receives the binary image of one character unit from the image storage unit 51, and extracts the contour points of the binary image by the same processing as in the first embodiment. On the other hand, the line width detection unit 53 detects the line width W of the binary image by referring to the one-character unit binary image stored in the image storage unit 51. Note that the method of detecting the line width W is the same as in the second embodiment.

FIG. 14 is a diagram showing the relationship between the line width and the end points. In the case of a character line having a general thickness, the higher the resolution, the thicker the line width. Therefore, the size of the line width gives an indication of the resolution. In other words, the larger the line width W is, the higher the resolution is, and therefore, the thinning-out scanning threshold can be increased. However, as shown in FIG. 14, even if the resolution is high, when the line width W is originally extremely small, a slight shift in feature point extraction using the thinning-out scanning threshold may greatly affect. In the case of such a thin line, it is desirable to use a thinning-out scanning threshold even if the resolution is high. Conversely, line width W
If is large enough, a slight shift has little effect. For such thick lines, it is desirable to use a large thinning scan threshold to significantly reduce processing time. Therefore, it is considered that the thinning-out scanning threshold should be determined according to the size of the line width W.

The thinning-out scanning threshold detector 54 receives the line width W detected by the line width detector 53, and receives the following (12),
The first and second thinning-out scanning thresholds MTH1 and MTH2 are calculated by the equation (13). MTH1 = W / N3 (Truncation below decimal point) (12) MTH2 = W / N4 (Truncation below decimal point) (13) where N3, N4; constant (for example, 2) MTH1 when MTH1 = 0 = 1 When MTH2 = 0, MTH2 = 1.

In the feature point extracting section 55, the contour point extracting section 52
, The line width W from the line width detecting unit 53, the thinning scanning threshold values MTH1 and MTH2 from the thinning scanning threshold value detecting unit 54, and feature points are extracted by the following processing. First, the feature point extraction unit 55 calculates the threshold values TH1 and TH2 using the line width W as in the second embodiment. Further, the subsequent processing is performed with the threshold value TH.
A feature point is extracted by thinning-out scanning similar to the third embodiment using the thinning-out scanning thresholds MTH1 and MTH2. When the feature points are extracted by the feature point extraction unit 55, the contour table in which the feature points are written is sent to the unevenness detection unit 56. The unevenness detecting unit 56 receives the contour table from the feature point extracting unit 55, receives the line width W from the line width detecting unit 53, and calculates the threshold value TH3 from the line width W in the same manner as in the second embodiment. Is a convex feature point or a concave feature point. The unevenness detecting unit 56 gives a contour table indicating whether the feature point is a convex feature point or a concave feature point to the character recognizing unit 13. Recognize.

As described above, in the fourth embodiment,
A line width detection unit 53 and a thinning scanning threshold value 54 are provided, and the threshold values TH1 and TH1 used by the feature extraction unit 55 in accordance with the character line width W.
TH2 is calculated, and the feature points are extracted. Therefore, since it is not affected by the line width W of the character, a stable feature point can be always extracted, and the processing time can be suppressed even when the resolution is high. Further, the unevenness detecting section 56 also calculates the threshold value in accordance with the line width W of the character and detects the unevenness of the characteristic point, so that the information of the unevenness of the characteristic point is always stable without being affected by the line width W of the character. Can be detected. In addition, since the thinning-out scanning thresholds MTH1 and MTH2 change in accordance with the line width W, when the line width W is small and the deviation of the feature point has a large effect, a small thinning-out scanning threshold is used to determine the feature point. In the case where extraction is performed as accurately as possible and the line width W is large and the shift of the feature point is not so affected, a large thinning-out scanning threshold value is used, so that the processing time can be greatly reduced. Therefore, for example, a form in which characters with non-uniform line widths W are mixed because of writing with a plurality of writing tools, or a document in which characters with non-uniform line widths are mixed due to being printed with a plurality of fonts, is targeted. In this case, since the thinning-out scanning changes according to the character line width, the processing effect is particularly large.

Fifth Embodiment FIG. 15 is a block diagram of a character recognition device provided with a feature extracting device according to a fifth embodiment of the present invention. This character recognition device includes an image storage medium 11, a character cutout unit 12, and a character recognition unit 13 similar to those of the first embodiment, and is connected between the character cutout unit 12 and the character recognition unit 13. A feature extraction device 60 different from that of the first embodiment is provided. As in the first embodiment, the feature extraction device 60 includes an image storage unit 61 that stores the binary image given from the character extraction unit 12.
And an outline point extraction unit 62 connected to the image storage unit 61. On the output side of the image storage unit 61,
A line width detection unit 63 for detecting the line width W of the binary image stored in the image storage unit 61 is connected. Contour point extraction unit 6
2, a thinning-scanning-threshold detecting unit 64 that detects a thinning-scanning threshold corresponding to the resolution is connected.

On the output side of the contour point extracting section 62, the line width detecting section 63 and the thinning scanning threshold detecting section 64, a feature point extracting section 65 is provided.
Is connected to the output side of the feature point extraction unit 65 and the line width detection unit 63. The feature point extracting unit 65 scans the contour points extracted by the contour point extracting unit 62 and extracts characteristic portions of the character as feature points. Is to detect whether the extracted feature point is a convex feature point or a concave feature point. The output side of the unevenness detection unit 66 is connected to the input side of the character recognition unit 13.

Next, the operation of the character recognition device shown in FIG. 15 will be described. The character extracting unit 12 receives a binary image from the image recording medium 11 in which a binary image such as a document or a form in which characters are described is stored, and cuts out the binary image in units of one character. The image storage unit 61 receives and stores a binary image in units of one character from the character extraction unit 12. The contour point extracting unit 62 receives the binary image of one character unit from the image storage unit 61, and extracts the contour points of the binary image by the same processing as in the third embodiment. The thinning-out scanning threshold value detection unit 64 includes the contour point extraction unit 62
And detects the first and second thinning-out scanning thresholds MTH1 and MTH2 in the same manner as in the third embodiment. On the other hand, the line width detection unit 63 refers to the one-character unit binary image stored in the image storage unit 61,
The line width W of the binary image is detected. The method of detecting the line width W is the same as in the second embodiment.

In the feature point extracting section 65, the contour point extracting section 62
, The line width W from the line width detecting unit 63, the thinning scanning threshold values MTH1 and MTH2 from the thinning scanning threshold value detecting unit 64, and feature points are extracted by the same processing as in the fourth embodiment. Then, the feature points are written in the contour table. The unevenness detecting unit 66 includes the feature point extracting unit 6
5, the line width W is received from the line width detection unit 63, and in the same manner as in the second embodiment, whether the characteristic point is a concave characteristic point or a convex characteristic point is obtained, and written in the outline table. The outline table is provided to the character recognition unit 13, and the character recognition unit 13 performs character recognition while referring to information on outline points and feature points in the outline table.

As described above, in the fifth embodiment,
A line width detection unit 63 is provided, and thresholds TH1 and TH2 used by the feature extraction unit 65 are calculated according to the line width W of the character to extract feature points. Therefore, it is possible to always extract a stable feature point without being affected by the line width W of the character. In addition, the unevenness detecting unit 66 calculates a threshold value in accordance with the line width W of the character and calculates the unevenness of the characteristic point. Is detected, the information of the unevenness of the feature point can always be detected stably without being affected by the line width W of the character.

Further, a thinning-scanning threshold detecting section 64 is provided, and thinning-scanning thresholds MTH1 and MTH corresponding to the number of contour points are provided.
2 is obtained and the feature points are extracted by performing the scan thinned by the thinning scan threshold values MTH1 and MTH2.
It is possible to reduce the processing time when the resolution is high and the number of contour points is large. Note that the characteristic points may be slightly shifted when the thinning scanning is simply performed. When the resolution is high, a slight shift is not noticeable, but when the resolution is low, even a slight shift greatly shifts from a characteristic point. However, in the present embodiment, since the thinning amount is determined according to the contour point amount Rsum, the problem can be sufficiently dealt with. Therefore, for example, a form in which characters with irregular sizes are mixed due to being entered in a plurality of entry frames, or a sentence in which characters with irregular sizes due to being printed in multiple font sizes are mixed. Even in the case of a target, a feature point can be extracted with an appropriate thinning amount.

The present invention is not limited to the above embodiment, but can be variously modified. For example, there are the following modifications. (1) In the contour point extraction units 22, 32, 42, 52, and 62, contour points are obtained by performing a contour tracking method as shown in FIG. 3, but the invention is not limited to this. Method may be used. (2) In the detection of the line width W, the observation window of two rows and two columns in FIG. 11 is used, but another method may be used as long as the line width W can be detected. (3) The thresholds TH1, TH2, and TH3 are determined from the line width W. Other equations may be used as the equations (7) to (9). (4) Thinning-out scanning threshold M from the contour point count value Rsum
Expressions (10) and (11) for obtaining TH1 and MTH2 may use other expressions. Similarly, other expressions may be used for the expressions (12) and (13) for obtaining the thinning-out scanning thresholds MTH1 and MTH2 from the line width W.

(5) In the first to fifth embodiments, the feature point extraction result and the irregularity detection result are written in the contour table and passed to the character recognition unit 13. However, the feature point extraction result and the irregularity detection are used. Each result may be passed separately from the contour table. (6) In the first to fifth embodiments, the segment sg
When there are a plurality of segments sg1 and sg
2, the contour points N satisfying (condition 1) and (condition 2)
Although R has been extracted, when extracting the contour point NR, the contour table of all segments may be targeted.
That is, the point of interest C and the contour point NR satisfying the condition may be different segments sg1 and sg2. (7) The character extracting section 12 inputs a binary image from the image recording medium 11, but is not limited to this, and may input an image from a scanner or a facsimile.

[0064]

As described above in detail, according to the first aspect, the image storage unit, the contour point extracting unit and the feature point extracting unit are provided, and the contour point sequence is formed from the points of interest C _n and C _{n + 1.} The contour points NR _n and NR _{n + 1} at the positions separated by the first threshold value TH1 and closest to the points of interest C _n and C _{n + 1} are obtained, and the contour points NR _n and NR _{n + 1} are determined. If the distance on the contour point is greater than or equal to the second threshold value TH2, the attention point C _{n + 1} is extracted as a feature point. Therefore, regardless of the direction change of the contour point sequence, the feature point is obtained from the distance. In addition to extracting end points and steep branches and intersections as feature points as well as extracting gradual arcs that are not suitable as feature points as feature points, Feature points can be reliably extracted from intersections.

According to the second aspect, the feature extracting apparatus includes the same image storage unit and contour point extracting unit as in the first aspect, and further includes a line width detecting unit and a characteristic point extracting unit to reduce the line width. Since the corresponding first and second thresholds TH1 and TH2 are determined and the feature points are extracted using the first and second thresholds TH1 and TH2, even if the line width of the character changes, The feature points can be reliably extracted without being affected by the influence.
According to the third aspect, the image storage unit and the contour point extracting unit are provided, and the thinning scanning threshold value detecting unit and the feature point extracting unit are provided, and the first and second thinning scanning threshold values MTH1 according to the amount of the contour points are provided. , MTH2, and thinning-out scanning are performed using the first and second thinning-out scanning thresholds TH1 and TH2 to detect a feature point. An increase in the throughput is suppressed.

According to the fourth aspect, an image storage unit, an outline point extraction unit, and a line width detection unit are provided, and a thinning scanning threshold value detection unit and a feature point extraction unit are provided. 2 thinning scanning thresholds MTH1 and MTH2 are determined,
First and second thresholds TH1 and TH2 corresponding to the line width are determined, and the first and second thinning scanning thresholds MTH1 and MT are determined.
Since the feature points are extracted at the first and second threshold values TH1 and TH2 according to the line width while performing the thinning-out scanning at H2, only the feature points can be reliably extracted even if the line width changes. At the same time, the processing amount can be suppressed. In addition, even if thinning scan is performed, it becomes thinning according to the line width,
The feature point deviation can be maintained at a value that can be ignored.

According to the fifth aspect, an image storage unit, an outline point extraction unit, a line width detection unit, and a thinning scanning threshold value detection unit are provided, and a feature point extraction unit similar to that of the fourth aspect is provided. And second thinning-out scanning threshold MTH according to the amount of
1, MTH2, and first and second thresholds TH1 and TH2 corresponding to the line width, and performing the thinning-out scanning at the first and second thinning-out scanning thresholds MTH1 and MTH2. Since the feature points are extracted using the first and second threshold values TH1 and TH2, only the feature points can be reliably extracted even if the line width changes, and the processing amount can be suppressed. In addition, even if the thinning-out scanning is performed, the thinning-out is performed according to the amount of the contour points, so that the deviation of the feature points can be maintained at a negligible value. According to the sixth and seventh aspects of the present invention, since the unevenness detecting section is provided, it is possible to perform an appropriate process such as character recognition.

[Brief description of the drawings]

FIG. 1 is a configuration diagram of a character recognition device including a feature extraction device according to a first embodiment of the present invention.

FIG. 2 is a diagram showing a problem in conventional feature extraction.

FIG. 3 is a diagram showing an outline of contour point extraction performed by a contour point extracting unit 22 in FIG. 1;

FIG. 4 is a diagram showing a configuration example of a contour table.

FIG. 5 is a diagram showing a character configuration example.

FIG. 6 is a diagram showing a procedure of feature point extraction performed by a feature point extraction unit 23.

FIG. 7 is a diagram showing an outline of feature point extraction.

FIG. 8 is a diagram illustrating scanning of a contour table by the unevenness detection unit 24;

FIG. 9 is a diagram showing detection results of concave feature points and convex feature points.

FIG. 10 is a configuration diagram of a character recognition device including a feature extraction device according to a second embodiment of the present invention.

FIG. 11 is a diagram showing an observation window for line width detection.

FIG. 12 is a configuration diagram of a character recognition device including a feature extraction device according to a third embodiment of the present invention.

FIG. 13 is a configuration diagram of a character recognition device including a feature extraction device according to a fourth embodiment of the present invention.

FIG. 14 is a diagram showing a relationship between a line width and an end point.

FIG. 15 is a configuration diagram of a character recognition device including a feature extraction device according to a fifth embodiment of the present invention.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 11 Image recording medium 12 Character extraction part 13 Character recognition part 20,30,40,50,60 Feature extraction device 21,31,41,51,61 Image storage part 22,32,42,52,62 Contour point detection part 33, 53, 63 Line width detectors 43, 54, 64 Thinning scan threshold detectors 23, 34, 44, 55, 65 Feature point extractors 24, 35, 45, 56, 66 Unevenness detectors

Claims (7)

[Claims] 1. A feature extraction device for inputting an image of a character drawn by a character line and extracting a feature point representing a characteristic of the character in the image, wherein the image extracted in units of one character is input. An image storage unit for storing, a contour point extraction unit for extracting a contour point sequence forming the contour of the character line in the stored image, and scanning the contour point sequence extracted by the contour point extraction unit. Detecting a contour point NR _n separated from the notice point C _n, which is one of the points, by the number of dots on the contour point sequence by the first threshold value TH1 or more and the closest coordinate distance from the notice point C _n ; Further, after shifting the point of interest from the point of interest C _n to a point of interest C _{n + 1} shifted by one dot on the line of contour points, the contour point sequence extracted by the contour point extraction unit is scanned, and the point of interest is scanned. apart C n _{+ 1} from the first threshold value TH1 or more in the number of dots on the contour point sequence and the noticed Contour points N with the C n _{+ 1} to the closest distance coordinate distance
R _{n + 1} is detected, and when the number of dots on the contour point sequence between the contour point NR _n and the contour point NR _{n + 1} is larger than a second threshold TH2, the target point C _{n + 1} is determined. And a feature point extracting unit that sets the feature points. 2. A feature extraction device for inputting an image of a character drawn by a character line and extracting a feature point representing a feature of the character in the image, wherein the image extracted in units of one character is input. An image storage unit for storing, an outline point extraction unit for extracting an outline point sequence forming the outline of the character line in the stored image, and a line width for detecting a line width of the character line of the stored image A first threshold value TH1 and a second threshold value corresponding to the detection unit and the line width;
Determining the threshold value TH2, the scanning the contour point sequence extracted by the contour point extracting unit, from the target point C _n is one of the outline point of the first in dots on the contour point sequence threshold TH1 more distant and detects the contour points NR _n in from the noted point C _n to the closest distance coordinate distance further target point C _n of the point of interest from the noted point C _n shifted 1 dot on the contour point sequence _After moving to ₊₁ , scan the outline point sequence extracted by the outline point extraction unit,
From the noted point C _{n + 1} apart first threshold TH1 or more in the number of dots on the contour point row and detecting contour points NR n _{+ 1} in the the noted point C _{n + 1} to the closest distance coordinate distance and, the noted point when the number of dots on the contour point sequence is greater than the threshold value TH2 of the second between the contour points NR _n and contour point NR n _{+ 1} C
a feature point extraction unit having _{n + 1} as the feature point. 3. A feature extraction device for inputting an image of a character drawn by a character line and extracting a feature point representing a characteristic of the character in the image, wherein the image extracted in units of one character is input. An image storage unit for storing, an outline point extraction unit for extracting an outline point sequence forming the outline of the character line in the stored image, and a first area corresponding to an amount of the outline points extracted by the outline point extraction unit Thinning scanning threshold MTH1 and second thinning scanning threshold MTH
And thinning the scanning threshold detector for determining a 2, a contour point string extracted by the contour point extracting unit scans for each of the second thinning scanning threshold MTH2, said the target point C _n is one of the contour points The first threshold T is determined by the number of dots on the contour point sequence.
H1 or away and detects the contour points NR _n in from the noted point C _n to the closest distance coordinate distance, further, the point of interest from the noted point C _n on the contour point row first decimating scanning threshold After moving to the point of interest Cn _{+ MTH1} shifted by _MTH1 , the contour point sequence extracted by the contour point extraction unit is scanned for each of the second thinning-out scanning thresholds MTH2, and the contour point sequence is extracted from the point of _{interest Cn + MTH1.} detecting contour points NR n _{+ MTH1} that apart first threshold TH1 or more in dots on the contour point sequence and from the noted point C _{n + MTH1} the closest coordinate distance, the contour point NR _n and the contour Point NR
_The number of dots on the outline point sequence between _{n + MTH1} is equal to the second threshold value TH.
A feature point extraction unit that uses the point of _interest C _{n + MTH1} as the feature point when the value is larger than two. 4. An image of a character drawn by a character line is input.
And extract feature points representing the features of the characters in the image.
In the feature extracting device, an image for inputting and storing the image cut out in character units
Forming an outline of the character line in the stored image;
A contour point extracting unit for extracting a contour point sequence to be extracted, and a line width for detecting a line width of the character line of the stored image.
A detection unit; a first thinning scanning threshold MTH1 corresponding to the line width;
Thinning-scanning-threshold detector for determining the second threshold MTH2
And a first threshold value TH1 and a second threshold value T corresponding to the line width.
H2 is determined, and a contour point sequence extracted by the contour point extracting unit
Is scanned every second thinning-out scanning threshold MTH2,
Attention point C, one of the points_nFrom the number of dots on the contour point sequence
At least the first threshold value TH1 and the point of interest C_nKara
Contour point NR closest to target distance _nIs detected
In addition, the point of interest is the point of interest C_nOn the contour point sequence from the first
Point of interest C shifted by thinning scanning threshold MTH1_{n + MTH1}Moved to
Then, the contour point sequence extracted by the contour point extracting unit is
2 for each thinning scanning threshold MTH2,
_{n + MT H1}From the number of dots on the outline point sequence to the first threshold TH
At least one point away and the point of interest C_{n + MTH1}Nearest coordinate distance from
Contour point NR at a long distance_{n + MTH1}And the contour point NR
_nAnd contour point NR_{n + MTH1}The number of dots on the contour point sequence between
When the target point C is larger than the second threshold value TH2,
_{n + MTH1}And a feature point extraction unit that sets the feature point as the feature point. 5. An image of a character drawn by a character line is input.
And extract feature points representing the features of the characters in the image.
In the feature extracting device, an image for inputting and storing the image cut out in character units
Forming an outline of the character line in the stored image;
A contour point extracting unit for extracting a contour point sequence to be extracted, and a line width for detecting a line width of the character line of the stored image.
A detecting unit; a first unit corresponding to the amount of the contour points extracted by the contour point extracting unit;
Thinning scanning threshold MTH1 and second thinning scanning threshold MTH
A first threshold value TH1 and a second threshold value T corresponding to the line width.
H2 is determined, and a contour point sequence extracted by the contour point extracting unit
Is scanned every second thinning-out scanning threshold MTH2,
Attention point C, one of the points_nFrom the number of dots on the contour point sequence
At least the first threshold value TH1 and the point of interest C_nKara
Contour point NR closest to target distance _nIs detected
In addition, the point of interest is the point of interest C_nOn the contour point sequence from the first
Point of interest C shifted by thinning scanning threshold MTH1_{n + MTH1}Moved to
Then, the contour point sequence extracted by the contour point extracting unit is
2 for each thinning scanning threshold MTH2,
_{n + MT H1}From the number of dots on the outline point sequence to the first threshold TH
At least one point away and the point of interest C_{n + MTH1}Nearest coordinate distance from
Contour point NR at a long distance_{n + MTH1}And the contour point NR
_nAnd contour point NR_{n + MTH1}The number of dots on the contour point sequence between
When the target point C is larger than the second threshold value TH2,
_{n + MTH1}And a feature point extracting unit that uses the feature point as the feature point. 6. A contour point _{R.sub.n + TH.sub.3} which has advanced the contour point sequence in the forward direction by a preset third threshold value TH.sub.3 from a contour point _R.sub.n corresponding to the feature point and only the third threshold value TH.sub.3. The contour point R _n-TH3 returned in the opposite direction is obtained, and the contour point R _n ,
The positional relationship of the contour point R _{n + TH 3} and contour point R _n-TH3, the position of the contour point R _n is provided with a concave-convex detecting part for detecting whether become concave or has a convex in the character The feature extraction device according to claim 1 or 3, wherein: 7. Determine the third threshold TH3 from said detected line width by the line width detecting section, the contour point sequence by the threshold TH3 of the third from the contour point R _n corresponding to the feature point A contour point R _{n + TH3} which has advanced in the forward direction and a contour point R _n-TH3 which has returned in the reverse direction by the third threshold value TH3 are obtained, and the contour point R _n , the contour point R _{n + TH3} and the contour point R are obtained. the positional relationship of the _n-TH3, claim 2, 4 or 5, characterized in that the position of the contour point R _n is provided with a concave-convex detecting part for detecting whether become concave or has a convex in the character A feature extraction device as described.