JPH1011533A - Character recognizing device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To detect the inclination of the area independent from the size of a partial area of a picture by collating a detected inclination pattern and an angular pattern stored in a storing means so as to calculate the result of collation. SOLUTION: A partial area extracting means 1 divides a picture on a first page into some parts at first. A straight line component candidate extracting means 2 extracts a straight line candidate coordinate point from a partial area picture extracted here. Next a straight line candidate coordinate point pattern is judged by a line segment detecting matrix. The line segment detecting matrix represents the inclination of virtual straight lines respectively combining the group of the straight line candidate coordinate points and an inclination representing the line segment detecting matrix discriminated before is selected. As an erroneous detection is easily generated by only discriminating the pattern of one straight line candidate coordinate point, the generation probability of the pattern of a present straight line candidate coordinate point based on a past detecting result is weighted to improve detecting precision.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an apparatus for recognizing characters of a document image read by an image scanner or the like.

[0002]

2. Description of the Related Art At present, in a character recognition apparatus, a technique of cutting out characters one by one from a document image read by a scanner is indispensable. At this time, if the document image read by the scanner is tilted, a highly accurate cutout can be performed by detecting the tilt angle of the image and correcting the cutout position according to the tilt angle.

In recent years, in this type of document image inclination detecting apparatus, a horizontal or vertical run length is obtained while rotating the inclined image at a predetermined angle, and the inclination of the binary image is obtained based on the statistical value. Was. For example, the rotation angle when the horizontal white run length is maximized is defined as the image inclination angle. Japanese Patent Application Laid-Open No. 2-56081 discloses such a technique.

In addition, connected components of individual characters are integrated by using labeling or the like, and a contour point in contact with a straight line inclined by a predetermined angle from the horizontal and vertical directions with respect to the character string is formed as a temporary character string. Then, the inclination angle of the image is obtained by obtaining the inclination angle of the provisional character string from the positional relationship between the points. Japanese Patent Application Laid-Open No. 7-192085 discloses such a technique.

[0005]

A device, such as a facsimile, for inputting paper by a roller and reading it by a scanner is widely used. However, in this type of device, the quality and condition of the paper to be read or the characteristics of a paper feeding mechanism are considered. A state called skew occurs where the tilt angle of the image is not uniform in the read image.

In the configuration of the first prior art, since the image is rotated at a predetermined angle and the inclination angle of the entire image is calculated from the statistical value of the white run length, the calculation time is problematic. Since the difference in run length becomes clearer and the detection accuracy improves as the size of the region becomes larger, it is inconvenient to detect the inclination angle of a part of the image.

In the configuration of the second prior art, a tentative character string can be obtained even in a wide area including a certain amount of characters, even if the outline points slightly vary. However, there is a problem that the inclination angle of the provisional character string cannot be obtained because variation occurs depending on the type of the character string.

SUMMARY OF THE INVENTION The present invention has been made in view of the above problems, and is a character recognition device capable of detecting a tilt angle of a partial area of an image without depending on the size of the area and capable of coping with skew. The purpose is to provide.

[0009]

SUMMARY OF THE INVENTION In order to solve the above-mentioned problems, the present invention divides a read image into a predetermined area and extracts the partial area, and a detection means for detecting an inclination pattern of the predetermined area. Storage means for storing a plurality of angle patterns in advance; matching means for matching the inclination pattern detected by the detection means with the angle pattern stored in the storage means; and calculating the result of matching by the matching means. It is provided with a configuration of control means.

[0010] Further, a partial area extracting means for dividing the read image into a predetermined area and extracting the same, a detecting means for detecting an inclination pattern of the predetermined area, a storage means for storing a plurality of angle patterns in advance, A matching unit that matches the inclination pattern detected by the unit with the angle pattern stored in the storage unit; a pattern table that stores a past occurrence probability of the inclination pattern; a matching result obtained by the matching unit and stored in the pattern table And a control means for determining a collation result based on the occurrence probability to be performed.

Further, a partial area extracting means for dividing the read image into predetermined areas and extracting the same, a bit operating means for adding predetermined bits in the horizontal direction to the image extracted by the partial area extracting means, and a bit operating means Detecting means for detecting the processed inclination pattern of the predetermined area; storage means for storing a plurality of angle patterns in advance; and collating the inclination pattern detected by the detection means with the angle pattern stored in the storage means. It is provided with a configuration of a matching unit and a control unit that calculates a result of the matching by the matching unit.

Further, a partial area extracting means for dividing the read image into a predetermined area and extracting the same, a detecting means for detecting a tilt pattern of the predetermined area, and a degree of approximation of the tilt pattern detected by the detecting means by a neural network. Is provided as a neural network means for detecting the

In addition, the present invention is provided with a configuration in which the inclination of the image is corrected by rotating the image based on the calculated inclination angle of the predetermined area, characters are cut out one by one from the corrected image, and the characters are recognized. .

Further, reading means for reading learning data in which the inclination pattern is described over a plurality of predetermined areas, and learning a teacher signal indicating the inclination pattern read by the reading means and the corresponding inclination angle. Learning means,
A storage means is provided for storing the inclination pattern and the inclination angle obtained by the learning means in association with each other.

According to the present invention, in order to detect an inclination angle of an image in a partial area, first, a line candidate coordinate point is obtained, and then a pattern of the line candidate coordinate point is determined by a line detection matrix. . Each of the line segment detection matrices represents the inclination angle of a virtual straight line connecting the line candidate coordinate point group, and the representative inclination angle of the previously determined line segment detection matrix is selected. Since a detection error is likely to occur only by discriminating the pattern of one straight line candidate coordinate point, the detection accuracy is increased by weighting the occurrence probability of the current straight line candidate coordinate point pattern based on the past detection result. Thus, even if the input image has a skew, accurate inclination correction can be performed, and highly accurate character cutout and character recognition can be performed.

In addition, when the character space in the character string in the input image is wide, the bit OR in the horizontal or vertical direction of the image is performed in order to facilitate the calculation of the line candidate coordinate points. Thus, the straight line candidate coordinate points can be stably calculated without being affected by the character type.

Further, by using a neural network when discriminating the pattern of the straight line candidate coordinate points by the line segment detection matrix, the amount of memory required for processing can be saved, and high-speed processing can be performed. Further, at the time of learning, an inclination pattern is formed over a plurality of regions in the image data and is read at a time, so that the operation for learning can be facilitated.

[0018]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, a first embodiment of the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram of a document image inclination detecting apparatus according to an embodiment of the present invention, FIGS. 2 and 3 are schematic explanatory diagrams of a linear component candidate extracting means in FIG. 1, and FIG. 4 is a detailed explanatory diagram thereof. is there.

First, an original is read by a reading means (not shown) and stored in a memory. When one page or a certain amount of image information is stored in the memory, the partial area extracting means 1 starts dividing the image. In the partial area extracting means 1, first, as shown in FIG.
Divide 1 into several. N × M divided image 22
It is called a partial area image. FIG. 3 shows an enlarged view of this image.

From the extracted N × M partial area image, the line component candidate extracting means 2 extracts a line candidate coordinate point.
In FIG. 3, the detection direction is upward as an example. Arrow 32
A change point 34 where a pixel on the image changes from white to black from the start point 33 of the above is detected as a straight line candidate coordinate point.

The details of FIG. 3 will be described with reference to FIG. As shown in FIG. 4, q00 to qnm (n = N / point interval, m = M /
(Point interval) is set as the detection start point of the line candidate coordinate point. These start points 33 are arranged at a fixed interval from neighboring start points. The pixel is scanned in any of the predetermined upper, lower, left, and right directions from the start point 33, and a point at which a white pixel changes to a black pixel is detected. These change point coordinates are stored in a N ′ × M ′ pixel area as shown in FIG. 5 as straight line candidate coordinate points. That is, q00 to qm shown in FIG.
A change point is detected upward starting from the detection start point of n.
Then, an N ′ × M ′ image area 51 (M ≧ M) shown in FIG.
M ′, N ≧ N ′) is converted into an N × M image area. Specifically, only the abscissa of the change start point is extracted and converted from size N to size N ', and the ordinate is converted from size M to size M'. At this time, the coordinates of the change point are also converted according to the conversion rate. Then, the means 4 performs collation with the line segment detection matrix. Here, the coordinates are converted in the vertical and horizontal directions, but the coordinates may be converted only in the horizontal direction. However, converting the coordinates in the vertical direction can reduce the memory capacity.

Next, a method of collating the line detection matrix will be described. FIG. 6 shows an example of the n × m line segment detection matrix. 61 is an attention point pixel, 62 is a straight line component candidate point,
Other matrix groups not shown here are also representative feature patterns for checking how much the local candidate coordinate point locally has an inclination. The amount of local inclination of these patterns is calculated using the target point pixel as cp and the pattern (a).
= 0 degree, pattern (b) = 10 degrees, pattern (c) = 2
Defined as 0 degrees.

FIG. 7 illustrates a method of referring to the line segment detection matrix. Reference numeral 51 denotes an N ′ × M ′ binary pattern in which a set of straight line candidate coordinate points is arranged in a certain area, and is created by the straight line component candidate extracting means 2. 72 is n
A reference pixel a00 in which the center point pixels 61 of the line detection matrix are arranged at equal intervals in an area where the nxm line detection matrix 72 refers to the line candidate coordinate point. a01,..., anm are set.

The collation is performed by comparing the distance between the n × m line segment detection matrix 72 and the pattern of the n × m line candidate coordinate points overlapping this area, and comparing the reference pixel a
This is performed based on the two occurrence probabilities from the line segment detection matrix determined in nm.

First, an n × m line segment detection matrix 72
Then, the distance of the pattern of the n × m straight line candidate coordinate point overlapping this area is calculated. As a result of weighting the occurrence probability with respect to this distance, the line segment detection matrix that has become the minimum is defined as the coincident line segment detection matrix. Here, a method of calculating the occurrence probability will be described. The probability of the next occurrence from the previously determined line segment detection matrix of k times is preset. FIG. 8 shows an example where k = 3.
When determining the first three line segment detection matrices, since the number of past determination data is three or less, all the occurrence probabilities are determined to be the same. N indicates a matrix having a tilt angle of 0,
Uh and Lh (h = 1, 2,... H: H is the number of components) indicate matrices having inclinations in the counterclockwise or clockwise rotation directions, respectively. Indicates that Here, H = k.

Next, a method of determining the line segment detection matrix being referred to will be described. The occurrence probabilities of Uh, Lh (h = 1, 2,... K) and N are defined as P (Uh), P (Lh), and P (N), and the distance between Uh, Lh and N in the area currently being referred to is d. (Uh), d (Lh) and d (N). The product of the occurrence probability and the distance is calculated for all the line segment detection matrices.

Dis (s) = d (s) × P (s) where s is Uh, Lh or N. A line segment detection matrix having the highest product dis (s) is selected.

When the line segment detection matrices are determined for all the reference pixels, the inclination angle of the image of the partial area is calculated from these matrixes as follows.

When the line segment detection matrix detected in the reference pixel component (i, j) has a local inclination amount t (degree), lm (i, j) = t, where t: -tw ≦ t ≦ tw tw: maximum allowable inclination angle i: number of horizontal reference pixel components in N ′ × M ′ straight line candidate coordinate point pattern j: number of vertical reference pixel components in N ′ × M ′ straight line candidate coordinate point pattern , And the inclination angle Tout is Tout = MAX (lm (i, j)) where MAX (k) is obtained by the element of k indicating the maximum appearance frequency.

By rotating the image using the tilt angle of the image, an image without tilt can be obtained. The characters described in this image are cut out character by character. Thereafter, character recognition is performed.

Next, a second embodiment of the present invention will be described with reference to the drawings. FIG. 9 is a block diagram of a document image inclination detecting apparatus according to a second embodiment of the present invention. The same reference numerals as those in FIG. 1 denote the same components.

First, the N extracted by the partial area extracting means 1
The × M partial area image undergoes image deformation in the partial area bit operating means 11 of the straight line component candidate extracting means 2. This instruction is issued from the partial area element distance extracting means 12. The partial area element distance extracting means 12 extracts the minimum element of the partial area using the projection value in the vertical or horizontal direction, finds the average of the distances between the minimum elements, and finds out how much the character spacing in the character string as a whole is. Is estimated. Thereafter, as shown in FIG. 10A, the bit OR of the currently referred bit data and the immediately preceding bit data are performed in the horizontal direction from the start point a to the end point b in accordance with the estimated value. The result is as shown in FIG. As a result, the distance between the minimum elements almost disappears, and when extracting the line candidate coordinate points by the line segment detection matrix matching means 4, more characteristic points representing the linear elements can be extracted. In this case, it is assumed that the arrangement of characters is horizontal.

Next, the straight line component candidate extracting means 2 extracts straight line candidate coordinate points. As shown in FIG. 4, q00 to qnm,
(N = N / point interval, m = M / point interval) is set as the detection start point of the line candidate coordinate point. These starting points are arranged at a fixed interval from neighboring starting points. Pixels are scanned in any of the predetermined upper, lower, left, and right directions from this start point to detect points where white pixels change to black pixels. These change point coordinates are set as straight line candidate coordinate points and stored in the N ′ × M ′ pixel area of the straight line component candidate storage unit 3. In the line segment detection matrix matching means 4, matching with the line segment detection matrix is performed. Subsequent processing is the same as in the first embodiment.

Next, a third embodiment of the present invention will be described with reference to the drawings. FIG. 11 is a configuration diagram of a document image inclination detecting apparatus according to a third embodiment of the present invention, and FIG. 12 is a schematic explanatory diagram of creation of learning data of the neural network in FIG.

First, the N extracted by the partial area extracting means 1
The straight line component candidate extracting means 2 extracts straight line candidate coordinate points from the × M partial area image. As shown in FIG.
Let m, (n = N / point interval, m = M / point interval) be the detection start points of the line candidate coordinate points. These starting points are arranged at a fixed interval from neighboring starting points. Pixels are scanned in any of the predetermined upper, lower, left, and right directions from this start point to detect points where white pixels change to black pixels. These change point coordinates are set as straight line candidate coordinate points and stored in the N ′ × M ′ pixel area of the straight line component candidate storage unit 3. The neural network 17 performs collation with the line detection matrix.

Next, a method of detecting a line segment element from a straight line candidate coordinate point by a neural network of the neural network means 17 will be described with reference to FIGS. 7, 11 and 12. FIG.

FIG. 12 illustrates a method for creating neural network learning data. Reference numeral 81 denotes image data, and a straight line 83 is rotated by t degrees. The learning data is an m × n dot width area indicated by 82, and the center line 8 of the area is
4 along d. Learning is performed while supplying the learning image data 82 at that time using the inclination angle t as a teacher signal. At this time, when the desired inclination angle is not output, the operator changes the neural network (operation coefficient) so as to output the desired inclination angle. Then, learning is repeatedly performed in the same manner while changing the inclination angle t of the straight line by Δt. As a result, the output becomes t (degree) with respect to the input of various patterns having the same inclination angle t as shown in FIG.
It can be. Here, since the recognition is limited to the inclination pattern, the operation for learning is facilitated by reading the inclination patterns written over a plurality of areas in the image data 81 and sequentially recognizing them as shown in FIG. I have to. Further, even with the same inclination, learning is performed by changing the position, and the operation is further facilitated.

After the learning data is created, the inclination pattern is recognized using the data. In a neural network, the network (coefficient) can be changed so that expected output data can be obtained for input data by learning. In this embodiment, the input is a pattern of a certain area of the N ′ × M ′ pixel area, and the expected output for the pattern is the pattern closest to this pattern.
That is, any line segment detection matrix having a similar shape may be used.

In the case of detecting a line segment element, the area 72 in FIG. 7 is input to the neural network, and the output is the detected amount of inclination t (degree) of the detected line segment matrix.

When the line segment detection matrices are determined for all the reference pixels, the inclination angle of the image of the partial region is calculated from these matrixes. The calculation formula is as follows. When the line segment detection matrix detected by the neural network in the reference pixel component (i, j) has a local inclination amount t (degree), lm (i, j) = t, where t: −tw ≦ t ≦ tw tw: maximum allowable tilt angle value i: number of horizontal reference pixel components in N ′ × M ′ straight line candidate coordinate point pattern j: number of vertical reference pixel components in N ′ × M ′ straight line candidate coordinate point pattern , And the inclination angle Tout is Tout = MAX (lm (i, j)) where MAX (k) is obtained by the element of k indicating the maximum appearance frequency.

The feature of this method is that the memory of the line-segment detection matrix table storage means 5 of the first embodiment becomes unnecessary, and the processing is performed at high speed because learning is performed in advance.

[0043]

As is clear from the above description, according to the present invention, the pattern of the straight line candidate coordinate points can be determined by the line segment detection matrix.

Furthermore, by weighting the current occurrence probability from the past detection result, it is possible to prevent an error in individual pattern discrimination, and it is possible to obtain an effect that more accurate tilt angle detection can be performed. .

When the character space in the character string in the input image is wide, the bit OR in the horizontal or vertical direction of the image is performed in order to facilitate the calculation of the straight line candidate coordinate point. Thus, the straight line candidate coordinate points can be stably calculated without being affected by the character type.

Further, by using a neural network when discriminating the pattern of the line candidate coordinate points by the line segment detection matrix, the amount of memory required for processing can be saved, and high-speed processing can be performed. Further, when learning the recognition using a neural network, a tilt pattern is formed in a plurality of areas in the image data, and this is read at a time, so that the learning operation can be facilitated.

[Brief description of the drawings]

FIG. 1 is a configuration diagram of a first embodiment.

FIG. 2 is an explanatory diagram of a partial area extracting unit.

FIG. 3 is an explanatory diagram of a straight line component candidate extraction method in a partial area.

FIG. 4 is a detailed explanatory diagram of a straight-line component candidate extracting unit;

FIG. 5 is an explanatory diagram of a linear component candidate storage unit.

FIG. 6 is an n × m line segment detection matrix diagram.

FIG. 7 is a detailed explanatory diagram of an nxm line segment detection matrix matching method;

FIG. 8 is an explanatory diagram of an occurrence probability table of a line segment detection matrix.

FIG. 9 is a configuration diagram of a second embodiment.

10A is an explanatory diagram of a partial area bit operating unit (before the start of operation); FIG. 10B is an explanatory diagram of a partial area bit operating unit (after the start of operation);

FIG. 11 is a configuration diagram of a third embodiment.

FIG. 12 is an explanatory diagram of a method for creating learning data in a neural network.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Partial area extraction means 2 Straight line component candidate extraction means 3 Straight line component candidate storage means 4 Line segment detection matrix collation means 5 Line segment detection matrix table storage means 6 Image inclination angle calculation means 7 Line segment detection matrix occurrence probability table storage means 8 Image Inclination correction means 9 Character cutout means 10 Character recognition means 11 Partial area bit manipulation means 12 Partial area element distance extraction means 13 Learning data storage means 14 Input data switching means 15 Teacher signal storage means 21 One page image 22 Partial area image 32 Arrow indicating the detection direction of the straight-line component candidate coordinates 33 Detection start point of the straight-line component candidate coordinates 34 Detection change point of the straight-line component candidate coordinates 51 N '× M' pixel area 52 Straight-line component candidate coordinate points 61 Attention point pixel 62 Straight-line component candidate Point 72 nxm line segment detection matrix 73 Reference pixel 81 Input image data 82 linear 84 centerline with the inclination of the training data 83 t of

Claims (6)

[Claims] 1. A partial area extracting means for dividing a read image into a predetermined area and extracting the same, a detecting means for detecting a tilt pattern of the predetermined area, a storing means for storing a plurality of angle patterns in advance, A character recognizing device comprising: matching means for matching the inclination pattern detected by the means with the angle pattern stored in the storage means; and control means for calculating the result of matching by the matching means. 2. A partial area extracting means for dividing a read image into a predetermined area and extracting the same, a detecting means for detecting a tilt pattern of the predetermined area, a storing means for storing a plurality of angle patterns in advance, A matching unit that matches the inclination pattern detected by the unit with the angle pattern stored in the storage unit; a pattern table that stores a past occurrence probability of the inclination pattern; a matching result obtained by the matching unit and stored in the pattern table And a control unit for determining a collation result based on the occurrence probability to be performed. 3. A partial area extracting means for dividing a read image into predetermined areas and extracting the same, a bit operating means for adding predetermined bits in a horizontal direction to the image extracted by the partial area extracting means, and a bit operating means. Detecting means for detecting the processed inclination pattern of the predetermined area; storage means for storing a plurality of angle patterns in advance; and collating the inclination pattern detected by the detection means with the angle pattern stored in the storage means. A character recognition device, comprising: a matching unit; and a control unit that calculates a result of the matching by the matching unit. 4. A partial area extracting means for dividing a read image into a predetermined area and extracting the same, a detecting means for detecting a tilt pattern of the predetermined area, and a degree of approximation of the tilt pattern detected by the detecting means by a neural network. And a neural network means for detecting a character. 5. The method according to claim 1, wherein the inclination of the image is corrected by rotating the inclination of the image based on the calculated inclination angle of the predetermined area, and characters are cut out from the corrected image one by one to recognize the characters. A character recognition device according to any one of claims 1 to 4. 6. A reading means for reading learning data in which a tilt pattern is described over a plurality of predetermined areas, and a teacher signal indicating a tilt pattern read by the reading means and a tilt angle corresponding thereto is learned. A character recognition device comprising: learning means; and storage means for storing an inclination pattern and an inclination angle obtained by the learning means in association with each other.