JPH1139429A - Character recognition part - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve the precision of character recognition by generating a segment pattern in a specific direction from a character pattern, extracting feature data from the generated segment pattern, and analyzing the feature data and deciding whether or not the data are a deletion line. SOLUTION: The observation part 2 of the character recognition device 1 reads characters entered into a document A by a scanner, etc., and converts them into binary image data, a character segmentation part 3 separates the binary image data into respective characters, and a preprocessing part 4 removes noise from the segmented characters to normalize the sizes, etc. A segment pattern generation part 91 generates a segment pattern in a specific direction from the normalized character patterns and a segment extraction part 92 extracts feature data from the generated segment pattern. Further, a deletion line decision part 93 analyzes the feature data on the segment pattern to decide whether the data are a deletion line. Therefore, the precision of the character recognition can be improved.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

[0001] The present invention relates to an optical character recognition device (OCR). In a character recognition device for recognizing handwritten characters, it is required to reduce misread characters and reduce restrictions on form entry.

[0002]

2. Description of the Related Art FIG. 9 is a conceptual diagram of a conventional character recognition device (OCR). Each part of the character recognition device 1 operates as follows. First, the observation unit 2 reads a character (and a character frame) written on the form A with a scanner or the like, and after photoelectric conversion, converts the character into binary image data. Character cutout part 3
Separates each character from the form image one by one. The preprocessing unit 4 performs noise removal, size normalization, and the like on the extracted characters. The feature extracting unit 5 extracts a feature value for character recognition. The dictionary unit 6 stores a dictionary of characteristic values for each character type. The dictionary matching unit 7 compares the extracted characteristic value of the character with the dictionary, selects the character type having the closest characteristic value as a candidate, and the result output unit 8 outputs the recognition result B accordingly.

There are various types of feature extraction methods.
FIG. 10 shows an example. In this example, as shown in FIG. 10A, a pixel of a contour portion of a pixel pattern of a character is extracted, and a direction code indicating a direction of the contour is attached to each pixel. In this example, the image is divided into four directions: horizontal, upward diagonal, vertical, and left diagonal. Then, the pixels are counted for each direction, and the arrangement is treated as a feature vector (FIG. 10).
(See (2)). The dictionary unit 6 also stores vector values for each character type in the same manner. The distance (feature distance) between the input character vector and the dictionary vector is calculated and candidates are set in ascending order. At this time, if the distance from the candidate character is long, or if the difference between the first and second candidates is small, the recognition may not be possible.

Conventionally, some OCR forms have a reject function in which a line segment (strikethrough) penetrating a character string is entered for a character string entered incorrectly so that the character string is not read. (Japanese Unexamined Patent Publication No. 61-368)
74). However, this is an effective method under the condition of "sufficiently long line segment" compared to the width of one character. is not.

[0005] When writing a strikeout for one character,
For example, a rule is used to fill the inside of the character frame or draw a long horizontal line that crosses the character frame. If this rule is obeyed, the difference in characteristic distance from the original character is clear and can be distinguished, but since this rule places a burden on the writer, it is often not observed. FIG. 11 shows an example of a form in which handwritten characters are input. Therefore, there is a possibility that a considerable percentage of the characters that the writer intends to have a strike-through are misrecognized without being rejected. For example, “0” with a short horizontal line crossed out is mistaken for “8”. This is "0 with a strikethrough" and "8"
This is because it is not easy to distinguish them because the feature distance is short.

In order to avoid this problem in the conventional method, it is necessary to judge which character is equivalent to the original character, and also to judge the character from being erased. Needs to be extracted. The feature dictionary will also be large.

[0007]

SUMMARY OF THE INVENTION The present invention focuses on the fact that the form of a strike-through line is most frequently determined by horizontal, vertical and oblique line segments, An object of the present invention is to improve the accuracy of character recognition by determining whether or not a character pattern cut out for each character from an image has a strike-through line, and rejecting when it is determined to be. .

[0008]

FIG. 1 is a block diagram showing the principle of a character recognition apparatus according to the present invention. The observation unit 2, the character cutout unit 3, the preprocessing unit 4, the feature extraction unit 5, the dictionary unit 6, the dictionary collation unit 7, and the result output unit 8 may be the same as the conventional one. Note that these portions may have configurations different from those in this drawing.

A first aspect of the present invention is a line segment pattern generating section for extracting a line segment pattern in a predetermined direction from a character pattern.
91, a line segment extraction unit 92 that extracts feature data from the generated line segment pattern, and a line eraser determination unit 93 that analyzes characteristic data of the line segment pattern and determines whether or not the line is a strikeout line. The configuration is as follows. The line segment extraction unit 92 extracts a pattern corresponding to a line segment in a predetermined direction, for example, a horizontal direction, from the character pattern, and generates a line segment pattern. This generation is performed, for example, as in claim 2 below.

A strike-out determination unit 93 extracts feature data from the generated line segment pattern. The feature data may be extracted by a histogram method as described in claim 3, a method of extracting feature data from a character pattern, or the like. The strike-through determination unit 93 analyzes the feature data and determines whether there is a line segment that is not a part of the character pattern. If there is such a line segment, it is regarded as a strike-through line and the character is deleted by the result output unit.

A second aspect of the present invention: a line segment pattern generator 91
Is moved on a character pattern using a rectangular extraction window that is long in the direction of extraction, and if the number of black pixels in the extraction window is less than or equal to a predetermined value, all pixels in the extraction window are assumed to be white pixels, and For example, a line segment pattern is generated by converting all pixels as black pixels.

FIG. 2 is an explanatory diagram of line segment extraction. If there are many black pixels in the elongated extraction window, there is a high possibility that the portion is a line segment in that direction. If there are few black pixels, there is a high possibility that the portion is not a line segment in that direction. Therefore,
As shown in FIG. 2A, by generating a pattern replaced with a black pixel and a white pixel, pixels forming a line segment in the extraction window direction are extracted as shown in FIG. 2B. Become.

The eraser determination unit 93 scans the line segment pattern in the direction to be extracted to generate a histogram of the number of pixels, and determines the width and height of the peak in a predetermined area of the generated histogram. It is configured to determine whether or not it is a strike-through line.

FIG. 3 is an explanatory diagram of the relationship between characters and line segment extraction. Since all the pixels present in the line segment pattern are not necessarily derived from the strike-out line, those having a high possibility of being strike-through lines are extracted. For example, since there are many characters having a horizontal line segment near the upper and lower boundaries (see FIG. 3 (2)), and since the strike-out line rarely exists near the upper and lower boundaries, when determining the histogram in the horizontal direction, The accuracy is higher when peaks near the upper and lower boundaries are removed.

A line segment extracting section 92 extracts feature data from a line segment pattern in the same manner as feature data extraction from a character pattern. Is analyzed to determine whether or not it is a strike-through line.

In the invention, the strike-out determination unit 93 analyzes the characteristic data of the line segment extracted by the line segment extraction unit 92 and corresponds to the character type recognized from the characteristic data of the input character pattern. It is configured to determine whether or not it is a strike-out line by using the determination criteria set as above.

For example, the pattern of "0" or "8" has no peak exceeding the threshold value in the histogram by the horizontal scanning, but the standard pattern of "4" as shown in FIG. Has one peak near the center exceeding the threshold.
Therefore, if the peak of the histogram exists near the center of the pattern recognized as “0” or “8”, it is highly likely that a strike-out line is attached, but the center of the pattern recognized as “4” is high. Even if there is a peak in the histogram, it is unlikely that the line is struck out. Therefore, the determination of the presence of a strike-through line is more accurate when it is changed for each character type.

A sixth aspect of the invention: a line segment pattern generating section 91
Generates a plurality of line segment patterns, a line segment extraction unit 92 extracts line segment characteristic data for each line segment pattern, and a strikeout line determination unit 93 detects the presence or absence of a strikeout line in any direction. The configuration is as follows.

In addition to the horizontal direction, the vertical direction, the diagonally upward right direction,
The same may be applied to the case of a diagonally up left direction. It should be noted that if a rule can be determined such that the strikeout line is a horizontal line, the speed can be increased by detecting only the horizontal direction.

The invention according to claim 7: a line segment pattern generating section 91.
Is configured to rotate a character pattern by a predetermined angle, if necessary, before generating a line segment pattern.
For example, if it is rotated 45 degrees and performed four times,
Subsequent processes need only be repeated the same.

In the character recognition apparatus, a line segment pattern in a predetermined direction is extracted from the character pattern, feature data of the extracted line segment pattern is extracted, and the feature data of the line segment pattern is analyzed. A character recognition program is recorded in a computer-readable recording medium, wherein the character recognition program performs a process of determining whether the character is a strike-out line.

With the above-described configuration, the character can be removed by appropriately determining the erased line entered by the writer to indicate that the entry has been made incorrectly. Therefore, the character that should have been erased is less likely to be erroneously recognized as another character, and the accuracy can be improved.

[0023]

DESCRIPTION OF THE PREFERRED EMBODIMENTS An example of an optical character recognition device (OCR) to which the present invention is applied will be described. FIG. 4 is a block diagram showing the configuration of a character recognition device which mainly allows handwritten numerals to be crossed out by a line segment having a character width.

FIG. 5 shows a flow chart of the overall processing of this embodiment. s1 to s6: The character recognition device 1 reads a character (and a character frame) written on a form with a scanner or the like, and converts it into binary image data after photoelectric conversion. A character extracting unit 3 for separating each character by a conventional method, a preprocessing unit 4 for removing noise from the extracted characters, normalizing the size, and the like; a feature extracting unit 5 for extracting a character recognition feature value; A dictionary unit 6 in which a dictionary of feature values for each character type is stored, and a dictionary matching unit 7 that matches the extracted character recognition feature values with the dictionary and outputs an intermediate recognition result. Up to this point, it is the same as the conventional method.

S7 to s10: The intermediate recognition result and the normalized character pattern generated by the preprocessing unit are sent to a line segment pattern generation unit 91 and a deletion line determination unit 93 to perform a deletion process. The erased line determination is performed by detecting a line segment in a horizontal direction, a vertical direction, a diagonally rising right direction, and a diagonally rising left direction with respect to the normalized character pattern. In addition, a threshold file 94 in which a threshold for determination is set for each character type and for each direction is prepared, and a threshold set corresponding to the character type set as the first candidate in the intermediate recognition result is taken out. To determine whether there is a strike-through.

If it is determined that the character is a strike-through character, the recognition result character type code is replaced with a reject code, and the result is output as the final recognition result. If it is determined that the character is a normal character without a strikeout line, the intermediate recognition result is output as it is as the final recognition result.

In the following, description will be made focusing on the process of determining a strike-through line. The strike-out processing is performed by the line segment pattern generation unit 9 shown in FIG.
1, line segment extraction unit 92, strikeout determination unit 93, threshold file 94,
And a line segment pattern rotation unit (not shown).

The line segment pattern generator 91 extracts a line segment pattern from the normalized character pattern as follows. FIG. 6 shows a flow chart of the strike-through processing.

(1) Extraction of vertical, horizontal, upward-sloping diagonal, and left-upward diagonal line segments A strike-out line to be determined is a strike-out line composed of vertical, horizontal, and diagonal line segments. Here, the oblique direction is the diagonal direction of the circumscribed rectangle of the character. FIG. 7 shows the directions of the character circumscribed rectangle and the extracted line segments. 7 (1) and (2) are the circumscribed rectangles of the characters, and 1 in FIG. 7 (1) is the horizontal direction, 2 is the diagonal direction going up left, 3 is the vertical direction, and 4 is the diagonal direction going right up. .

The extraction of line segments will be described below by taking the horizontal direction as an example. To extract another direction, the angle of the diagonal direction of the circumscribed rectangle is calculated, the coordinate of the character pattern is converted by that angle, and then rotated by that angle (FIG. 7 (2)
See) and do the same. Alternatively, the rotation may be sequentially performed at a fixed angle of 45 degrees, 90 degrees, and 135 degrees.

In the character pattern area, the vertical and horizontal size nx
Consider an m-long rectangular (n <m) window. The number of black pixels in this area is counted, and the total number of pixels in the rectangle (n × m
), The ratio B of the number of black pixels to the number

When this threshold value is set to TH1, B> TH
In the case of 1, all the pixels inside the rectangular area are replaced with black pixels. When B ≦ TH1, pixels inside the rectangular area are
All are replaced with white pixels. The window is moved all over the character circumscribed rectangle to extract a horizontal line pattern. FIG.
(A) shows an example in which n × m = 3 × 8 and TH1 = 0.7.

(2) Counting the number of line segments in each direction A histogram of the number of pixels is generated from the line pattern obtained above, and the number of line segments in a predetermined area is counted. Similarly, a description will be given using a horizontal line segment as an example.

A black pixel count histogram is created by scanning the horizontal line segment pattern in the horizontal direction and counting black pixels.
However, as shown in FIG. 3 (3), since the uppermost or lowermost character line is often detected as a horizontal line segment such as “5” or “2” in the case of a numeral, the uppermost or lowermost character line is detected. From the bottom, a histogram is created excluding the area of width LE = LH × TH2. TH2 is the width of the exclusion area with respect to the height of the circumscribed rectangle. Note that the histogram may be created for the entirety, and may be excluded at the time of determination. Alternatively, it may be excluded when a line segment pattern is generated. FIG. 6 shows an example thereof.

If the width of the circumscribed rectangle of the character pattern is LW, the histogram is cut at a position LC satisfying LC = LW × TH3. TH3 is a ratio of the cut portion width to the circumscribed rectangular width LW.

Assuming that the height of the remaining cut peak portion (the checkered portion in FIG. 3A) from the cut portion is H and the width of the cut portion is W, a portion satisfying H / W> TH4 is horizontal. Judge as a line segment in the direction.

If the number of the obtained horizontal line segments is equal to or larger than the threshold value TH5, it is determined that the character has a strike-through line formed by horizontal line segments. In other words, when LH is the height of the circumscribed rectangle of the pattern, if (H / W> TH4) & (exists in LH-LE ≧ y ≧ LE), the line is a crossed line eraser candidate.

Number of strike-through candidates for horizontal line segments ≧ TH5
If it is, it is determined that the character is a character with a strike-through line by a horizontal line segment. (3) Here, each threshold of TH1 to TH5 is set to an optimal value for each direction of the line segment to be extracted and for each character type, and is prepared as a threshold file. A set of thresholds to be used is determined according to the primary recognition result character type of the character to be determined. Vertical, upward-sloping, and left-upward diagonal line segments are similarly processed. The search area for a strike-through candidate is within a circumscribed rectangle in the horizontal and vertical directions, but may be in a rectangle including a diagonal line of the circumscribed rectangle in an oblique direction.

FIG. 8 shows an example of a character with a strikethrough determined according to the present invention.

[0040]

As described above, according to the present invention, a character with a strike-through line can be appropriately determined, and misreading of a character with a strike-through line in a certain direction can be reduced.

[Brief description of the drawings]

Fig. 1 Principle configuration diagram

FIG. 2 is an explanatory diagram of line segment extraction.

[Figure 3] Extraction example of character and line segment diagram

FIG. 4 is a configuration block diagram of an embodiment.

FIG. 5 is a flowchart of the entire process.

FIG. 6 is a flowchart of a strike-through processing.

FIG. 7 is an explanatory diagram of line segment extraction.

FIG. 8 is an example of a strike-through judgment.

FIG. 9 is a conceptual diagram of a configuration of a character recognition device.

FIG. 10 is an explanatory diagram of character feature extraction.

[FIG. 11] An example of a handwritten input form

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Character recognition apparatus 2 Observation part 3 Character cutout part 4 Preprocessing part 5 Feature extraction part 6 Dictionary part 7 Dictionary collation part 8 Result output part 91 Line segment pattern generation part 92 Line segment extraction part 93 Strikeout line judgment part 94 Threshold file

Claims (8)

[Claims] 1. A character recognition device for recognizing a character by extracting feature data of an input character pattern and comparing the extracted data with dictionary data, wherein the line segment extracts a line segment pattern in a predetermined direction from the character pattern. It has a pattern generation unit, a line segment extraction unit that extracts feature data from the generated line segment pattern, and a strikeout determination unit that analyzes feature data of the line segment pattern to determine whether or not it is a strikeout line A character recognition device characterized in that: 2. The character recognition device according to claim 1, wherein the line segment pattern generation unit uses a rectangular extraction window that is long in an extraction target direction, moves the character pattern on the character pattern, and sets the black in the extraction window. A character recognition apparatus characterized in that when the number of pixels is equal to or less than a predetermined value, all pixels in the extraction window are converted to white pixels, and when the number of pixels is larger than the predetermined value, all pixels are converted to black pixels to generate a line segment pattern. 3. The character recognition device according to claim 1, wherein the line segment extraction unit scans the line segment pattern in a direction to be extracted to generate a pixel number histogram, and a strike-out line determination unit. Is a character recognition device characterized in that it is determined from a width and a height of a peak in a predetermined area of a generated histogram whether or not the line is a strike-out line. 4. The character recognition device according to claim 1, wherein the line segment extracting unit extracts feature data from the line segment pattern in the same manner as feature data extraction from a character pattern. A character recognition device characterized in that a strike-through determination unit analyzes the extracted feature data to determine whether or not it is a strike-through. 5. The character recognition device according to claim 1, wherein the strike-out determination unit analyzes the input data of the line segment extracted by the line segment extraction unit. A character recognition device characterized in that it is determined whether or not a character is a strike-out line using a determination criterion set in correspondence with a character type recognized from characteristic data of a pattern. 6. The character recognition device according to claim 1, wherein the line segment pattern generation unit generates a plurality of line segment patterns, and the line segment extraction unit generates a line segment for each line segment pattern. A character recognition device characterized by extracting the characteristic data of (1), wherein the strike-through determination unit detects presence / absence of strike-through in any direction. 7. The character recognition device according to claim 1, wherein the line segment pattern generation unit generates the line segment pattern after rotating the character pattern by a predetermined angle. Character recognition device. 8. A character recognizing device for recognizing characters by extracting feature data of an input character pattern, comparing the extracted data with dictionary data, and extracting a line pattern in a predetermined direction from the character pattern. A segment pattern generating unit, a line segment extracting unit that extracts feature data of the extracted line segment pattern, and a strike-through line determining unit that analyzes the feature data of the line segment pattern and determines whether or not it is a strike-through line. A computer-readable recording medium on which a character recognition program is recorded.