JPH08241378A - Recognizing method for low-quality character - Google Patents

Abstract

PURPOSE: To read even a document consisting of deteriorated characters at a low misread rate by opening a feature dictionary to be used for reading according to the character quality of the document inputted from an image scanner. CONSTITUTION: A character recognition processor 22 temporarily displays document image input data, inputted by the scanning of the image scanner 21, on the display device 31 of a host computer 3 before reading. Read conditions are set according to the display contents. Then character vectors are extracted from the document image input data. An area corresponding to the character quality label on the read condition setting screen is opened in the feature dictionary containing feature character vectors by using character quality labels as indexes on the basis of the selective specification of the character quality label on the read condition setting screen. The cut character vector which is extracted is collated with feature character vectors stored in the quality label area according to similarities extract read candidate characters, which are selected and then outputted as determined recognized character after specific postprocessing.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition method for a character reader, and more particularly to a recognition method for low quality characters.

[0002]

2. Description of the Related Art A character reading device (OCR) 1 includes an image scanner 21 and a character recognition processor 22 as shown in FIG.
Character reading recognition unit 2 and host computer 3
The character recognition processor 22 receives the read data obtained by optically scanning the document to be read by the image scanner 21 as an input, and first observes the read data according to the process of the flow shown in FIG. Then, the character pattern data area to be processed as one character is cut out, the cut-out character pattern data is analyzed to extract the characteristic parameter contained in the character pattern data, and this characteristic parameter is set to each character in the reading target range. Correspondingly, the read character is recognized by comparing with the dictionary of the characteristic parameters belonging to each character prepared in advance and extracting the characters having the matching characteristic parameters, and the character classification code assigned to the character is output as the character information. The action is the basic function. At this time, the host computer 2 functions as a man-machine interface for setting the reading conditions of the document to be read, displaying the reading result, and the like, and edits and proofreads the document based on the document information obtained by the reading. Do the work well.

In the character recognition processor 22, the character pattern data area as illustrated in FIG. 8A which is input from the image scanner 21 and is cut out as one character has an appropriate fineness as shown in FIG. 8B. The cells are divided into cells of length L and width I, and the density value of each cell is a function f (Xp,
Yp) (1.ltoreq.p.ltoreq.I, 1.ltoreq.q.ltoreq.L) is read, and a set of density pattern values obtained corresponding to one character pattern data area is used as approximate data of the read character. Then, such a character pattern data area is divided into squares, and the number of squares (I × L = K) obtained in the squares is set to a group of density pattern values f (Xp, Yp). Numbers are assigned in order, and the density pattern value xi = f (Xp, Y
p) are arranged, and this is interpreted as a vector X expressed by the equation 1 stretched in the K-dimensional space.

[0004]

[Expression 1] X = (x1, x2, xk) k = I × L (1) As described above, one character has a dimension equal to the number k of cells dividing one character area corresponding to the reading resolution. Is represented by the character vector of, the K-dimensional reference character vector representing the character by the above-mentioned grid division is obtained in advance for each of the characters to be read, and the set of the reference character vectors is used as a character vector dictionary. In principle, the character vector of the recognition target character obtained by scanning the document to be read with the image scanner is compared with the reference character vector stored in the dictionary and the matching vector is selected. Can recognize the character.

By the way, even if a standardized character such as a character printed by a printer or printed characters,
When comparing the fonts of the same characters printed in the same typeface in detail, it is common to see slight differences due to differences in the character designs of the printer or the type manufacturer.
Even if the same character is printed, two characters printed in different fonts have somewhat different density pattern distributions in the character pattern area, and thus can be obtained by sampling the character pattern data area into cells. The character vectors are similar but not completely identical between two characters with different fonts. Therefore, if the condition of judgment recognition is a perfect match between the character vector of the character to be recognized and the reference character vector registered in the dictionary, most characters cannot be read and recognized.

In order to avoid the above-mentioned inconvenience that occurs when a perfect match of character vectors is used as a judgment condition for character recognition, and to recognize a character correctly without being affected by a slight difference in font,
Instead of matching two vectors in the character vector space, a distance d (X, Y) represented by equation (1), which means a separation between two vectors X and Y, or a similarity represented by equation (2). A method of using the degree s (X, Y) for recognition judgment is in practical use.

[0007]

[Equation 2]

[0008]

(Equation 3)

The distance d (X, Y) takes a value of 0≤d≤∞,
It becomes 0 when X = Y. On the other hand, the similarity s (X, Y) is equal to the cosine value cos θ of the angle θ formed by the two vectors X and Y, and takes a value of −1 ≦ s ≦ 1, and becomes 1 when X = Y. In character recognition, a method that uses the degree of similarity represented by Expression 2 that is less affected by the difference in print density as a criterion is often used.

The feature dictionary that serves as a reference in character recognition determination is a collection of features of optically read data extracted from standard samples such as Mincho typeface, Gothic typeface, italic typeface, and textbook typefaces (each having a plurality of fonts). Is. Therefore, the recognition rate of the character type / character style not registered in the feature dictionary becomes low. On the other hand, unnecessarily registering the character type / font in the feature dictionary relatively lowers the similarity of other character types / characters when calculating the similarity, which adversely affects recognition. Also, if the feature dictionary is too large, many similarity calculations will have to be performed,
It takes more processing time than necessary.

Therefore, the reference character vector corresponding to each character registered in the feature dictionary is not determined based on one character selected as a reference, but a large number of case samples of the same character belonging to the same genre are collected and obtained. The method of registering the average vector of each character vector as a character reference vector is being implemented.

[0012]

If the document to be read is a document resulting from repeated copying, the quality of the characters deteriorates and the characters are "blurred" or "blurred". The optical feature obtained from the character is significantly different from the corresponding character in the feature dictionary according to the related art and the degree of similarity is low, so that a correct recognition result cannot be obtained.

FIG. 9 shows an example of a difference in image due to deterioration of characters having the same font. FIG. 9A shows a character of normal quality,
The figure (b) is a poor quality character, and the figure (c) is a remarkably poor character. Although the word "use" is included in each case, the horizontal component is "blurred" in FIG. 9 (b) and hard to read in FIG. 9 (b), and the horizontal component in FIG. 9 (c). It is completely lost, and when the character as shown in (c) of FIG. 9 is recognized by OCR by the conventional method, the probability of being recognized as “river” is very high. Since the similarity to "use" is low and it does not enter the upper rank of the candidate character, it is very difficult to derive the accuracy by post-processing. Conventionally, it has been difficult to recognize such poor quality characters.

Further, if the degree of similarity is too low, candidate characters are selected (during the recognition process, a plurality of candidates for the recognition result of the character to be read are first selected. These are called candidate characters).
The correct recognition result is not output inside.
In such a case, it is difficult to derive a correct answer by post-processing such as Japanese processing using a pattern of combination of candidate characters.

Further, in the feature dictionary, a plurality of fonts are registered for one character type, and by adding an optical feature standard pattern having feature amounts such as "blurred" and "blurred",
It is also possible to recognize characters that are “blurred” or “blurred”, but the added optical feature standard pattern becomes closer to the optical feature standard pattern of other characters, has a high degree of similarity, and It may adversely affect the recognition result.

The present invention is a method for recognizing a deteriorated character which can read a document including a character whose quality is deteriorated and read it as a correct character or a character close to the correct character in a short processing step. To improve the recognition rate of deteriorated characters in a character recognition device.

[0017]

In order to solve the above problems, in the present invention, pattern data of a document to be read which is optically read by a character recognition method in a character recognition device by an image scanner and stored in a buffer is used. A reference character vector generated based on the read data of an image scanner of a character sample of different quality in advance is used as a character quality classification. Feature dictionary prepared by collecting and prepared for search
When setting the reading conditions, the characters of the document to be read are collated and recognized based on the selected character quality classification.

The feature dictionary corresponding to the character quality classification used in the character recognition process is prepared by collecting reference character vectors generated in advance based on image scanner read data of character samples of different qualities for each character quality classification. It is a feature dictionary by multiple quality levels or a single integrated dictionary in which the same character is classified according to the character quality level, and the quality referred to by the character quality label indicating the character quality level selected for the characters of the document to be read. Select the reference area of the level-specific feature dictionary or integrated dictionary. In addition, the characteristic dictionary corresponding to degraded characters is preset in accordance with the type of font and quality level for the reference character vector corresponding to individual characters recorded in the reference feature dictionary generated based on high-quality character samples. It may be obtained by multiplying by a mask for weighting the specified vector component.

The setting of the character quality label corresponding to the character quality level of the character of the document to be read is performed by the reading condition setting screen having the input image display area for displaying the image image scanned and input by the image scanner as it is. The input character image displayed in the input image display area is displayed on the window provided on the reading condition setting screen, and the case of the character of each character quality level prepared in advance corresponding to the character product level classification is displayed. Specify a quality label that represents a character quality level equivalent to the quality of.

[0020]

[Function] The characters in the document image displayed on the reading condition setting screen are equivalent to the character quality of the document to be read in comparison with the example of the deteriorated characters displayed corresponding to the character quality label in the window of the screen. When the character quality label is selected and designated, the feature dictionary corresponding to the designated character quality label is opened and character recognition processing is performed.

[0021]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 shows a process flow in an embodiment of a character recognition method using a similar character discrimination method according to the present invention, and the method of the present invention will be described with reference to FIG. The configuration of the character reading device that executes the method of the present invention is equivalent to that of the device illustrated in FIG.
In the following description, the reference numerals given in FIG. 6 will be referred to when necessary for the explanation.

The character recognition processor 22 of the character reader is
Document image input data scanned and input by the image scanner 22 is displayed on the display device 31 of the host computer 3 before reading.
Is displayed as it is (S2). Then, in consideration of the contents displayed on this input display screen, the reading direction and other reading conditions for reading the document image input data are set through the host computer 3 (S3).
1) The method of the present invention includes a step (S32) of selecting and designating the quality level of the character image input in the reading condition setting item in accordance with a predetermined character quality label category.

When the reading condition is set as described above, a character area as one character is cut out from the document image input data input to the character recognition device based on the set reading condition, and the character pattern data of the cut out character area is analyzed. Then, the cut-out character vector is extracted (S4, S5). Then, based on the selection specification of the character quality label on the reading condition setting screen, a character vector generated in advance based on the data read by the image scanner of character samples of different quality levels is set according to the quality level classification. An area corresponding to the character quality label of the feature dictionary in which the quality label is recorded as a feature character vector in the index is opened (S
6) After the characteristic character vector recorded in the quality label area is collated with the extracted cut-out character vector according to the degree of similarity to sort and extract the read candidate characters (S7), the candidate character characters extracted by the sorting are selected. (S8)
Post-processing (S9) consisting of correction by knowledge processing according to the prior art is added and then output as a decision recognition character (S).
Ten).

Then, a method of creating a feature dictionary searched by a character quality label corresponding to the quality level of a character used in the process of the character recognition basic flow in the method of recognizing a low quality character according to the present invention described above. Will be explained. In a document including characters of poor quality as illustrated in FIGS. 9B and 9C, other characters also show similar deterioration in quality as a whole. Such a document is often a document that is repeatedly copied. From this, it is considered that the degree of deterioration of character quality can be quantified by the number of copies. Therefore, copying is repeated based on the character image sample of normal quality, and the character image sample for creating the feature dictionary in which the character quality deteriorates according to the number of copies is created.

The deteriorated character image sample thus created is input to the character reading apparatus described with reference to FIG. 6 to extract characteristic parameters, and the degree of deterioration of character quality is expressed by the number of copies. Create a feature dictionary for each quality by attaching it to the label. The basic method of creating the above-mentioned feature dictionary is shown in FIG.

In the feature dictionary creating method illustrated in FIG. 2, the feature dictionary is configured by the volumes that are divided in the order of character quality and labeled with the character quality, but as illustrated in FIG. The characteristic character vectors obtained by different character quality samples of the same character are recorded in the characteristic dictionary of one volume for each character in the order of quality, and only the characteristic character vector specified in the character quality label is referenced when referring to the dictionary. May be referred to.

Next, the basic method of generating a feature dictionary using a weighted table prepared in advance is shown in FIG. 4 and this method will be described. For example, if the characters to be read are characters in Mincho type, the horizontal component is finer than the vertical component, so if you repeat copying on a document printed in Mincho type. The degree of “blurring” of the horizontal component is larger than that of the vertical component.
(See FIGS. 9A and 9B). Therefore, when calculating the similarity, a mask is created that weights the vertical and horizontal components so as to emphasize the remaining vertical components without blurring, and this weighting mask is applied to each character vector recorded in the feature dictionary. It is made to act and a practical feature vector is generated.

As described above, the mask corresponding to the character quality registered in the weighting table has a separate filter for extracting and attenuating the stroke portion which is theoretically easily deteriorated based on the character of the character. Either adopting a method of creating a sample of deteriorated characters without collecting it, or conversely, creating a filter that superimposes samples of deteriorated characters collected in advance and extracts the stroke part that remains in common without deterioration. be able to.

As the character dictionary to which the mask is applied, an individual feature dictionary belonging to a specific quality label generated by the deteriorated character sample may be used so as to correspond to a more deteriorated state. Is not sufficiently obtained and only the reference character dictionary for character reading is prepared, it is possible to apply a mask to the reference character dictionary so as to deal with the deteriorated character.

Finally, a method of selecting and designating the character quality label in setting the reading conditions will be described. As described at the beginning of the embodiment, in the character reading device 1, the document image data scanned by the image scanner 21 and input to the character recognition processor 22 is once displayed as it is on the display device of the host computer 3. It is usually done on the screen of the reading condition setting screen.

On the reading condition setting screen, as shown in FIG. 5, the input document image data display area ID, the reading range to be specified in reading, the reading direction and other setting items and their setting items are displayed. The area CR and the area CR. Therefore, in the present invention, a window WI for displaying the character quality level case is provided in the reading condition setting screen, and the character quality label and an example of representative deteriorated characters belonging to the character quality label are displayed in this window. Then, the degraded character of this representative example is compared with the character in the document image displayed in the input document image data display area, and the code of the character quality label having similar character quality is designated and input.

The window for displaying the character quality level case may be provided in a part of the area CR of the setting item, or may be opened in the input document image data display area ID by a command if necessary. .

[0033]

In the document reading apparatus to which the low-quality character recognition method of the present invention is applied, the characteristic dictionary used for reading is selected and opened corresponding to the character quality of the document input from the image scanner, so that the deterioration occurs. The effect of being able to read even a document consisting of written characters with a low misreading rate is first obtained, and the usage range of the feature dictionary is selected by the quality label so that it is not necessary to collate the entire range of the dictionary. Therefore, it is possible to obtain the effect that the processing speed does not decrease even when reading the deteriorated character.

According to the method of generating the individual characteristic dictionary for each character quality level by the character vector obtained by reading the character sample collected for each quality classification, the characteristic dictionary is automatically generated when the character sample can be collected. It will be possible. Further, according to the method of generating the feature dictionary at the use stage by using the weighting mask, even if a sufficient character sample is not obtained in advance to establish the feature dictionary corresponding to the character quality level, it may occur depending on the character style. Since the quality deterioration that tends to occur is generated by the mask, the effect that a feature dictionary adaptable to reading deteriorated characters can be obtained.

According to the method in which the window is opened on the reading condition setting screen and the character quality label and the case of the character belonging to the quality label are displayed, the document image to be read and the quality label are classified. Since an example of deteriorated characters is displayed on the same screen, it is possible to easily and appropriately perform the selection of the character quality label when the reading condition selection is designated.

[Brief description of drawings]

FIG. 1 is a process flow chart of a low quality character recognition method according to the present invention.

FIG. 2 is an explanatory diagram of a method for creating an individual feature dictionary by quality level.

FIG. 3 is an explanatory diagram of an integrated feature dictionary creating method for recording character vectors by quality level.

FIG. 4 is an explanatory diagram of a method of generating a characteristic dictionary corresponding to a deteriorated character by using a mask of a weighting table.

FIG. 5 is a configuration diagram of a reading condition setting screen according to the present invention.

FIG. 6 is a block diagram of a character reading device

FIG. 7 is a basic flowchart of character recognition processing in the character reading device.

FIG. 8 is an explanatory diagram of a character vector.

FIG. 9 is a diagram showing an example of printed characters having different character qualities.

[Explanation of symbols]

 1 character reading device 2 character reading recognition unit 21 image scanner 22 character recognition processor 3 host computer

Claims (5)

[Claims] 1. A cut-out character vector, which is an ordered set of values of a density pattern obtained by observing pattern data of a document to be read optically scanned by an image scanner and stored in a buffer and cut out as a single character area, is different in advance. A characteristic dictionary prepared by collecting searchable reference character vectors corresponding to character quality classifications based on data read by an image scanner of quality character samples, and selecting the characters of the document to be read when setting the reading conditions A low-quality character recognition method characterized in that character recognition is performed by matching under the specified character quality classification. 2. The feature dictionary comprises a plurality of quality level-based feature dictionaries prepared by collecting reference character vectors generated based on image scanner read data of character samples of different qualities in advance for each character quality category. The low-quality character according to claim 1, wherein a character dictionary having a quality level corresponding to a character quality level selected for a character of a document to be read is selected and collated with a cut-out character vector to perform character recognition. Recognition method. 3. The characteristic dictionary is a single integrated dictionary in which the character quality levels of the same characters are separately recorded based on image scanner read data of character samples of different qualities in advance. 2. The low-quality character according to claim 1, wherein a reference character vector registered in the integrated dictionary corresponding to the character quality level selected for the character is selected and collated with the cut-out character vector for character recognition. Recognition method. 4. A reference character vector corresponding to an individual character recorded in a reference characteristic dictionary generated based on a high-quality character sample in the characteristic dictionary is preset in accordance with the type of font and quality level. The method for recognizing a low quality character according to claim 1, wherein the character vector is configured by a character vector obtained by multiplying a mask for weighting a specific vector component. 5. A reading condition is set on a reading condition setting screen having an input image display area for displaying an image image scanned and input by an image scanner as it is, and a window portion provided on the reading condition setting screen is set. , An example of characters of each character quality level prepared in advance corresponding to the character product level classification is displayed, and a quality label representing the character quality level equivalent to the quality of the input character image displayed in the input image display area is specified. The method for recognizing low quality characters according to any one of claims 1 to 4, characterized in that the character quality classification relating to the feature dictionary is set accordingly.