JPH08147314A - Recognition type document filing apparatus and control method thereof - Google Patents

Abstract

(57) [Summary] [Purpose] A feature vector is stored for a character whose recognition result is suspicious, so that an accurate discrimination calculation can be performed later. [Structure] The document image read by the image scanner 3 is stored in the RAM 4, and characters are cut out and feature vectors are extracted. The extracted feature vector calculates the Euclidean distance from the average feature vector, obtains a predetermined number of candidate character groups in the order of closeness, and stores them in the large classification unit 11.
Thereafter, for each candidate, a pseudo Bayes identification formula for each typeface is calculated, and a predetermined upper number of each candidate is output. Then, the difference between the value of the first candidate's pseudo-Bayes discrimination formula and the value of the second candidate is determined.
Candidates are judged as correct, otherwise uncertain. For a character that is determined to be uncertain, the feature vector for that character is stored.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a recognition type document filing apparatus for reading a document from an original document to recognize characters and storing and storing the original document, and a control method thereof.

[0002]

2. Description of the Related Art Conventionally, in a document filing apparatus, a document is read by an image scanner and stored as bit map image data. In this case, even if the document image is compressed, the amount thereof is large, and a storage medium having a huge capacity is required to store a large amount of document images. Further, since the stored information is not text data, there is a problem that a word in a document cannot be used as a keyword for searching.

Therefore, there has been proposed a recognition type document filing apparatus which recognizes characters of a document by a character recognition device (OCR) and stores and stores them in the form of text data.

[0004]

[Problems to be Solved by the Invention] However, OCR
Although the recognition rate has improved, unfortunately it has not reached 100% accurate recognition of characters. Therefore, the conventional recognition-type document filing apparatus has a problem in that the text data includes characters that are erroneously recognized. Also, a means for correcting the recognition result can be used as a post-processing function of the OCR, but in that case, there is a problem that a lot of trouble is required for the correction.

[0005]

[Means for Solving the Problems] and

The present invention has been made in view of the above problems, and a recognition-type document filing apparatus and a recognition-type document filing apparatus capable of performing accurate identification calculation later by storing a feature vector for a character whose recognition result is suspicious. It is intended to provide the control method.

To solve this problem, for example, the recognition type document filing apparatus of the present invention has the following configuration.
That is, a recognition type document filing apparatus for recognizing and filing a character image cut out from input image data, and an identification dictionary storing standard feature data of a character image in association with a character code for each character type. An identification means for calculating a dissimilarity for each character type from the characteristic vector of the cut-out character image and the standard feature data, and selecting a plurality of candidate character types in ascending order of dissimilarity, and uncertainty of recognition accuracy. Check candidate determination means for determining
A text data generating means for generating text data from the information of the candidate character type selected by the identifying means; and a check candidate character feature extracting means for extracting a feature vector corresponding to the candidate character type which is a check candidate by the check candidate determining means. The text data generating means stores the feature data of the check candidate in the text data in association with the candidate character type of the text data, and stores the text data.

According to a preferred embodiment of the present invention,
The check candidate determination means calculates a pseudo Bayes identification formula for the cut-out character image and the recognition candidate character,
It is desirable to determine the uncertainty by whether or not the value of the pseudo-Bayes discriminant for the candidate character and the second candidate character exceeds a predetermined threshold value.

[0008]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An example of an actual sister according to the present invention will be described in detail below with reference to the accompanying drawings.

FIG. 1 is a hardware block diagram of an embodiment representing the present invention. In the figure, 1 is a CPU that performs overall control, identification calculation, character cut-out, etc., 2 is an address / data bus, 3 is an image scanner for reading a document and performing photoelectric conversion to capture it as an image, and 4 is an image scanner for reading. Stored document image, CPU1
RAM (Random Access
Memory). Reference numeral 5 is a feature extraction unit for extracting features from a character image, 6 and 7 are detailed dictionaries A, B and 8 for storing standard patterns used for detailed identification, and display units of a display for displaying document images and recognition results. , 9 are pointing devices for designating an arbitrary position on the display unit 8, 1
0 is R that stores the boot program and character font
OM (Read Only Memory), 11 is a large classification unit that roughly narrows down candidate characters, 12 is a large classification dictionary that stores standard patterns used in large classification, and 13 is a document image or recognition result. An external storage unit 14 that stores or stores the OS and the processing procedure (program) described later.
Is a keyboard for inputting commands for giving instructions to the CPU and various data, and 15 is a printer for printing document images, recognition results, and the like.

<Description of Processing> Pre-Processing Next, the processing will be described with reference to the flowchart of FIG. First, in step S210, the document set in the image scanner 3 is read, photoelectrically converted and binarized to form bitmap image data (referred to as a document image) in the RAM.
Store in 4.

In step S220, the read document image is displayed on the display unit 8. In step S230, the position is specified using the keyboard 14 and pointing device (hereinafter referred to as PD) while looking at the display 8.
Erase unnecessary parts of the document image such as noise. In step S240, a block of an area for character recognition and an area for image processing is designated. Designation is performed by using PD to enclose a desired block (rectangular area) by dragging the PD.

In step S250, the processing mode is set for each area. The processing mode is information on processing conditions such as how to process the area (character recognition or image compression). The processing conditions are vertical / horizontal writing in the case of character recognition.
Specifying how to insert a line feed code (whether it is inserted in each line, whether it is inserted in each paragraph or not at all), the character type to be recognized (Japanese or alphanumeric), detailed dictionary (which dictionary to use), and You can choose. Further, in the case of image compression, it is possible to select the compression method (MMR, MH, MR or other). In step S260, compression processing is performed on the area for which the image compression mode is set, the image is encoded, and the encoded data is temporarily stored in the RAM 4.

In the character cutting step S270, character cutting processing is performed on the area (image data) in the character recognition mode. This method is often used by taking a histogram of black pixels in the row direction and making line breaks in the valleys of the histogram, and taking a histogram of black pixels in the direction perpendicular to the rows for each line and making character breaks in the valleys of the histogram. This is a known method.

In the present embodiment, first, all line segmentation processing is performed, and then character segmentation is performed for one line. The coordinate data of the character cutting frame for the character that has been cut is stored in the RAM 4. The coordinate data of the character cutting frame is the coordinates of the upper left corner of the rectangle surrounding one character (x, y
Coordinates) and the length of the horizontal side and the length of the vertical side. It is assumed that the coordinate data of the character cutting frame is not erased even if the character cutting of a certain line is completed and the character cutting of the next line is performed, and is sequentially left. When the character sequence number is r, the structure array of the coordinate data of the character cutting frame is H [r], and is defined as follows in the C language. However, the sequence number r of the characters is assumed to be the numbers assigned to the characters in the order in which the first line, the second line, ... #define MAX_NUMS 100000 / * maximum size of H [r] * / struct H_DATA {int Point_x; / * upper left vertex x coordinate * / int Point_y; / * upper left vertex y coordinate * / int Side_x; / * horizontal side Length * / int Side_y; / * Length of vertical side * /} H [MAX_NUMS]; Feature extraction Next, in step S280, feature extraction is performed from the character image to obtain a feature vector. This is CPU1 RAM4
1 is extracted from the document image in FIG. 2 based on the coordinate data of the character cutting frame obtained in step S270 and is transferred to a character buffer (not shown) in the feature extraction unit 5, which is disclosed in Japanese Patent Publication No. 2-59507. After the feature extraction unit 5 finds a feature vector, which is a weighted direction histogram, and stores it in the character buffer, the CPU 1
And the CPU 1 stores the feature vector in the RAM 4. Here, details of the inside of the feature extraction unit 5 will not be described, but a method of processing by a microprocessor may be used, for example. The obtained feature vector is Xi (i = 1,
2, ..., N). The dimension number n of the vector is 72 in Japanese Patent Laid-Open No. 2-59507.

Large classification When the feature extraction processing is completed, the process proceeds to the next step S290.

By the way, in the large classification dictionary 12, an average value (referred to as an average feature vector) of feature vectors obtained from a fixed number (sufficiently large number, for example, 500) of sample character images for each character is stored in advance. It is stored in association with the code. The character code for the character s is c (s), and the average feature vector is mi (s) (i = 1,
2, ..., N), c (s) and mi (s) (i =
The correspondence table of 1, 2, ..., N) is stored in the large classification dictionary 12. Here, s is a sequential number attached to the character, and s = 1, 2, ..., N, and these s
A set of this character is called a recognition target character set. N is, for example, 3500, which is a scale including JIS first-level kanji and non-kanji.

Now, in step S290, step S
The feature vector Xi is set between the character image cut in the processing of 270 and all the characters in the recognition target character set.
And Euclidean distance D (s) defined by the following equation of the average feature vector mi (s) is calculated.

[0018]

[Equation 1]

The character set to be recognized is sorted so that D (s) is in ascending order, and the character number s, the character code c (s), and the distance value D (s) are stored in the RAM 4 for the top 50 characters. These 50 characters are called the candidate character set.

Detailed identification The detailed dictionary A and the detailed dictionary B are identification dictionaries used in the detailed identification A and the detailed identification B, which are respectively processed subsequently, and the character code c ( s) and the standard pattern Ψ (s). Here, the detailed dictionary A is for Mincho and the detailed dictionary B is for Gothic.

The standard pattern Ψ (s) has an average feature vector mi (s), an eigenvector φij (s), and an eigenvalue λj (j
, 1, 2, ..., K) are all generic names. The eigenvector φij (s) and the eigenvalue λj (j = 1, 2, ..., K) are
It is defined in Japanese Patent Publication No. 2-59507. However, the detailed dictionary A stores a standard pattern obtained from a learning sample of Mincho type, and the detailed dictionary B stores a standard pattern obtained from a learning sample of Gothic type. In addition, k is 3 here.

Now, in step S300, step S2
For all the characters cut out in 70 and the candidate character set, the pseudo Bayesian discriminant B1 (s) is calculated using each data of the detailed dictionary A, and sorted in ascending order to the value of B1 (s),
Character number s and character code c1 (s) of the top 10 characters
And the identification value B1 (s) are stored in the RAM 4, and a check candidate flag area and a typeface code area to be used later are secured.

Similarly, in step S310, step S2
The pseudo-base discriminant B2 (s) is calculated using each data of the detailed dictionary B for the characters cut in 70 and all of the candidate character sets, and sorted in ascending order of B2 (s) values. , The character number s of the upper 10 characters, and the character code c
2 (s) and the identification value B2 (s) are stored in the RAM 4, and a check candidate flag area and a typeface code area to be used later are secured.

FIG. 3 illustrates the data format in the RAM 4 of these identification result data for one character. The detailed identification A and the detailed identification B have the data of FIG. 3 separately.

In the font type discrimination step S320, the pseudo Bayes identification formula Bi obtained for each typeface (Mincho typeface / Gothic typeface)
(S) (i = 1, 2) is compared for the first candidate, and the typeface with the smaller value of the pseudo-Bayes identification formula is used as the result of the typeface determination, and the detailed identification corresponding to the typeface that is determined to be correct. The identification result data of (1) is copied to another area of the RAM 4, and 0 is set for Mincho typeface and 1 is set for Gothic typeface code. The copied identification result data will be referred to as recognition data hereinafter.

Check candidate determination In the detailed identification dictionaries A and B, the character code c (s) and the standard pattern Ψ for each character of the recognition target character set are described.
In addition to (s), the reliability determination threshold value Wi (s) (i = 1,
2) is included in correspondence. The reliability determination threshold value Wi (s) can be experimentally determined in advance so that the following holds.

If the character s is the first candidate, the character s
Identification value Bi (s) and the identification value Bi of the second-ranked candidate character p
The difference from (p) is expressed as δi (s) = Bi (p) −Bi (s) (Equation-2). When this is called the candidate discriminant value difference, the candidate discriminant value difference δ
If i (s) is larger than the reliability determination threshold value Wi (s), it is considered that the first candidate is stochastically correct, and the candidate identification value difference δi
If (s) is not larger than the reliability determination threshold Wi (s), the first candidate is considered to have poor reliability, and the validity of the recognition result should be rechecked by some means later. Is determined. In step S330, this is determined, and 1 is set in the check candidate flag area secured in the RAM 4 if it is a check candidate, and 0 is set if it is not a check candidate.

The processing from steps S280 to S330 is performed on the character image for one line that has been character-cut in step S270. Then, the process returns to step S270 and character cutting is performed on the next line, and steps S280 to S33 are performed.
If there is no line for which character cutout / character recognition is not performed by repeating the process up to 0, that is, if the above process is completed for all lines, the process proceeds to step S340.

Text Data Creation / Saving / Display Next, in step S340, only the first rank candidate is extracted from the recognition data (FIG. 3) stored in the RAM 4, and the structure type array Z [r] (r = 1, 2, ,,, u; u is the number of characters). At this time, the character code of the first candidate and the address pointer pointing to the recognition data corresponding to the character code and the sequence number v (described later) of 3
Structure type array Z with each element as one array data
Generate [r]. Therefore, in addition to the identification value / check candidate flag / typeface code associated with the first rank candidate, those pieces of information regarding the second rank and lower rank candidates can be easily obtained if the sequence number r of the input character is determined. However, the character sequence number r is the first line, the second line,
It is assumed that the characters are numbered in the order in which the characters were cut out.

For all the recognized characters whose check candidate flag is 1, the original document image (RAM4) is used by using the structure array H [r] of the coordinate data of the character cutting frame obtained in step S270. (Remembered in)
The individual character images are extracted to obtain the feature vector, and RA
The feature vector is continuously stored in M4. The sequence numbers pointing to these feature vectors are represented by v (= 1, 2, ...) And these feature vectors are called check candidate character features.

Then, the above-mentioned structure type array Z [r] has the above-mentioned sequence number v indicating the feature vector corresponding to the character as a structure element. However, when the check candidate flag is 0, v = 0 is set.

From the structure type array Z [r] described above, the character code, the sequence number v, and the feature vector of the check candidate character stored in the RAM 4 are output to the external storage device 13 for all r.

In this way, the character code of the recognition result and
For the check candidate characters, the feature vectors are associated and stored.

Next, in step S350, the font of the corresponding character is generated from a character generator (not shown) by referring to the character code from Z [r], and the characters are displayed on the display unit 8 in order. At this time, the flag is 1 when the check candidate flag of the recognition data is seen from the address pointer of Z [r].
If so, the color of the character is displayed in red, and if the flag is 0, it is displayed in white. The background color of the display portion is black, and the text display portion of the recognition result occupies approximately half the area of the entire display portion. The display of the document image displayed in S230 remains on the other half of the display unit. Then, based on the coordinate data of the character cutting frame obtained in step S270, the character frame is overlaid on the document image to draw a line.

Swap Candidate / Save Document Next, in step S360, the validity check is easily performed because the check candidate is cheap to distinguish from the others, and if the operator finds a character that is erroneously recognized, it is checked by the PD. Click the part. When the CPU1 detects this, Z
The candidate character is arranged from the information of the recognition data from the address pointer of [r] and displayed on the display unit, and the operator corrects the candidate character and selects the character with the PD. As a result, the CPU 1 performs a process of replacing the displayed character of the recognition result with the selected candidate character.

In step S370, when the check is completed as described above, the displayed data of the recognition result is output to the external storage unit 13 as a character code string. The data of the recognition result stored here is obtained in step S34.
It should be stored in an area different from the one saved with 0.

[0037]

[Other Embodiments] In the above embodiment, the text data is saved before "candidate replacement" and the feature vectors are saved for all check candidate characters, but the text data is saved after "candidate replacement". For the check candidate characters whose candidates have been replaced, v = 0 and the feature vector may not be stored. that way,
You may correct the characters so that the character code is correct as much as possible, and you are free to avoid the trouble as much as possible without correcting the characters.

As described above, according to the present embodiment, the feature vector is stored for the character whose check result is suspicious (check candidate character). Therefore, the accurate character recognition result can be obtained by performing the accurate identification calculation later. Is obtained. Therefore, it is possible to realize a convenient recognition type document filing apparatus in which the storage capacity is improved and information is not mistaken.

The present invention may be applied to a system composed of a plurality of devices or an apparatus composed of a single device as described above. Further, it goes without saying that the present invention can be applied to the case where it is achieved by supplying a program to a system or an apparatus.

[0040]

As described above, according to the present invention, it is possible to perform accurate discrimination calculation later by storing a feature vector for a character whose recognition result is suspicious.

[0041]

[Brief description of drawings]

FIG. 1 is a block configuration diagram of a recognition type filing apparatus according to an embodiment.

FIG. 2 is a flowchart illustrating the operation of the embodiment.

FIG. 3 is a diagram illustrating a storage format of recognition result data.

[Explanation of symbols]

 1 CPU 2 Address / Data Bus 3 Image Scanner 4 RAM (Random Access Memory) 5 Feature Extraction Section 6, 7 Detailed Dictionary 1 2 8 Display Section 9 Pointing Device 10 ROM (Read Only Memory) 11 Large Classification 12 Large Classification Dictionary 13 External storage unit 14 Keyboard 15 Printer

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁶ Identification code Office reference number FI technical display location 9194-5L G06F 15/403 350 Z

Claims (4)

[Claims] 1. A recognition-type document filing apparatus for recognizing and filing a character image cut out from input image data, wherein standard characteristic data of a character image is stored in association with a character code for each character type. An identification dictionary, an identification means that calculates the dissimilarity for each character type from the extracted feature image of the character image and the standard feature data, and selects a plurality of candidate character types from the smallest dissimilarity, and the recognition accuracy. Corresponding to the candidate character type which is a check candidate by the check candidate determining means, the text data generating means for generating text data from the information of the candidate character type selected by the identifying means, A check candidate character feature extracting means for extracting a feature vector; and the text data generating means, A recognition-type document filing apparatus characterized in that a candidate candidate feature vector is provided in text data in association with a candidate character type of text data, and the text data is stored. 2. The check candidate determination means calculates a pseudo Bayes identification formula for the cut out character image and the recognition candidate character, and a value of the pseudo Bayes identification formula for the first candidate character and the second candidate character is predetermined. The recognition-type document filing apparatus according to claim 1, wherein the uncertainty is determined based on whether or not the threshold value is exceeded. 3. A control method of a recognition-type document filing apparatus for recognizing and filing a character image cut out from input image data, wherein standard feature data of a character image is associated with a character code for each character type. By referring to an identification dictionary stored in advance, the dissimilarity is calculated for each character type from the characteristic vector of the cut-out character image and the standard characteristic data, and a plurality of candidate character types are selected from the smallest dissimilarity. An identification step, a check candidate determination step for determining the uncertainty of recognition accuracy, a text data generation step for generating text data from the information of the candidate character types selected in the identification step, and a check candidate determination means as a check candidate. And a check candidate character feature extraction step of extracting a feature vector corresponding to the candidate character type. Generation step, a feature vector checking candidates in correspondence with a candidate character type of text data to have the text data, the control method of the recognition type document filing apparatus characterized by storing the text data. 4. The check candidate determining step calculates a pseudo Bayesian discriminant for the cut out character image and the recognition candidate character, and a value of the pseudo Bayesian discriminant for the first candidate character and the second candidate character is predetermined. 4. The method for controlling a recognition-type document filing apparatus according to claim 3, wherein the uncertainty is determined by whether or not the threshold value is exceeded.