JP3187899B2 - Character recognition device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition apparatus for improving the recognition accuracy in a field of pattern recognition by using character extraction processing from a document, recognition, and language processing in a complementary manner.

[0002]

2. Description of the Related Art As a conventional character recognition apparatus, for example, "Study of Handwritten Character String Recognition Method", Yayoi Sato and Jun Tsugumo, Proceedings of the 1990 IEICE Spring Conference,
D-490, 7-243 are known.

FIG. 11 shows a main configuration diagram of this conventional character recognition device.

First, an image character string 2 captured and input
01 is passed to the rectangularization processing unit 202 as two-dimensional data. Here, it is assumed that the number of pixels included in the character string component is 1, the number of pixels not included is 0, the character string direction is the horizontal axis direction for convenience, and the height of the character string is also known.

FIG. 12 is a diagram showing a specific configuration of the conventional rectangular processing section 202. As shown in FIG. The character string 201 input to the rectification processing unit 202 is stored in the character string data storage unit 211.
The margin distribution calculation unit 212 calculates a value obtained by accumulating the pixel values held in the character string data storage unit 211 in the vertical direction, and outputs the result to the Gaussian filter processing unit 213. The Gaussian filter processing unit 213 removes high frequency components of the cumulative histogram by performing smoothing by convolving a Gaussian function with the vertical cumulative histogram of the pixel values of the character string data obtained by the peripheral distribution calculating unit 212. , Reduce the effects of noise. The smoothed cumulative histogram is input to the left and right end detection unit 214. As shown in FIG. 13, the left and right edge detection unit 214 detects a portion where the value of the cumulative histogram is equal to or less than a certain value (threshold value), and sequentially sets the boundary coordinate positions as the left and right edge coordinates of the basic rectangle in the column direction peripheral distribution calculation unit. 215 and the basic rectangle storage unit 217.

At this time, it is assumed that the threshold is set so that the width of all the basic rectangles is equal to or less than a certain multiple of the height of the character string.

The column direction marginal distribution calculating section 215 receives the input from the left and right end detecting section 214, reads out the pixel values of the portion sandwiched between the left and right end coordinates of the basic rectangle from the character string data storage section 211, and performs horizontal oblique shading. And outputs the result to the upper and lower ends
6 is output. Further, as shown in FIG. 14, the upper / lower edge detecting section 216 obtains the maximum value and the minimum value of the portion formed by the character component of the oblique result from the column-direction marginal distribution calculating section 215, and compares the result with the upper and lower edge of the corresponding basic rectangle. Is output to the basic rectangle storage unit 217.

By the above processing, the basic rectangle storage unit 217
The basic rectangle is represented by its left and right ends and its upper and lower ends, and is stored in the order of detection.

FIG. 15 is a block diagram showing the rectangular integrated recognition unit 203 shown in FIG.
FIG. 16 is a specific configuration diagram of the weighted lattice creating unit 241 which is a part of the configuration. Hereinafter, the operations of the rectangle integration recognition unit 203 and the lattice creation unit 241 will be described with reference to the drawings.

The basic rectangle combination unit 251 reads out the coordinate values of the basic rectangles stored in the basic rectangle storage unit 217 and determines that the outline width of the rectangle obtained by combining the adjacent basic rectangles is not more than a fixed multiple of the character string height. Are detected, and the result is output to the weighted lattice storage unit 252.

The weighted lattice storage unit 252 stores a
As shown in the figure, the upper, lower, left and right coordinates of the combination rectangle are stored and retained for all the combinations of the basic rectangles detected by the basic rectangle combination unit 251, and the weight values and candidate characters corresponding to each combination rectangle are also stored and retained. To secure memory for

A weight value calculation unit 242 receives inputs from the character evaluation value calculation unit 204 and the rectangle evaluation value calculation unit 205 in FIG. 11 and receives weights corresponding to each combination rectangle stored in the weighted lattice storage unit 252. Find the value and candidate characters,
The result is output to the corresponding memory of the weighted lattice storage unit 252 again.

First, the operation of the character evaluation value calculation unit 204 and the rectangle evaluation value calculation unit 205 will be described.

FIG. 18 is a specific configuration diagram of the character evaluation value calculation unit 204. The feature amount extraction unit 271 refers to the upper, lower, left and right coordinates of the combination rectangle stored in the weighted lattice storage unit 252, reads out the pixels included in the coordinates from the character string data storage unit 211, and calculates the feature amount. And outputs it to the similarity calculation unit 273. As the feature amount extracted by the feature amount extraction unit 271, for example, a mesh feature vector or the like can be considered. The mesh feature vector is obtained as follows.

One side of a combination rectangle including input pixels is divided into n equal parts, the entire rectangle is divided into n × n squares, and the number of pixels included in a character component in each square is determined by the entire square. The n × n-dimensional vector formed by arranging the values of the ratio of the pixel number to the pixel number in order is defined as the mesh feature vector.

The character dictionary 272 stores and stores a representative mesh feature vector of each character. The representative mesh feature vector of each character is an average vector of the mesh feature vectors of a plurality of standard characters.

The similarity calculation unit 273 calculates the reciprocal value of the square error between the representative mesh feature vector of each character and the mesh feature vector input from the feature amount extraction unit 271, and calculates a similarity value indicating which character and how close it is. Is output to the output selection unit 274.

After the similarity values for all the characters are input, the output selection unit 274 outputs the candidate characters corresponding to the maximum value x to the weight value calculation unit 242.

FIG. 19 is a specific configuration diagram of the rectangle evaluation value calculation unit 205. The squareness calculating unit 281 calculates the aspect ratio (squareness r) of the rectangle according to (Equation 1) by referring to the upper, lower, left and right end coordinates of the combination rectangle stored in the weighted lattice storage unit 252, and Output to calculation section 283.

[0020]

[Mathematical formula-see original document] r = vertical length of rectangle / horizontal length of rectangle The blankness calculation unit 282 refers to the upper, lower, left and right coordinates of the combination rectangle stored in the weighted lattice storage unit 252, and determines the character string data storage unit 211. , The pixels within the rectangle are read out, and the ratio of the number of pixels not included in the character component to the total number of pixels of the rectangle (blank degree s) is calculated and output to the matching degree calculation unit 283.

The goodness-of-fit calculating section 283 calculates the value y shown in (Equation 2)
Is obtained as a fitness value, and is output to the weight value calculation unit 242.

[0022]

Y = ρ1 (r) · ρ2 (s) where r is an input from the squareness calculator 281, s is an input from the blankness calculator 282, ρ1 and ρ2 are appropriate functions, and It does not change depending on the target.

As described above, the character evaluation value calculating section 20
4, when the output selection unit 264 outputs x and the corresponding candidate character from the output selection unit 264 and the rectangle evaluation value calculation unit 205 outputs y from the suitability calculation unit 273 as an output, the weight value calculation unit 242
Calculates the value w shown in (Equation 3) as the weight value of the weighted lattice storage unit 252 of the corresponding combination rectangle.

[0024]

## EQU00003 ## w = 1 (x.gtoreq..alpha.2) .rho. ((C0 + C1.x + C2.y + C3.x.y) / C4) (x <.alpha.2) where C0 = .beta.1. (. Alpha.2-1) (α2-α1) -α1 · β2 , C1 = α1 · β1-α2 · β1 + β2, C2 = α1 · α2-α2 2 + α2, C3 = α2-α1-1, C4 = (α2-α1) · (β2-β1 ), Ρ (z) = z (z> 0), ρ (z) = 0 (z ≦ 0).

The processing parameters α1, α2, β1, β2
Is a constant value and does not change depending on the processing target.

(Equation 3) satisfies the following four conditions, and it is considered that the larger the value, the more likely the character is. 1. If the evaluation value x obtained from the character evaluation value calculation unit 251 is smaller than a certain value α1, and the evaluation value y obtained from the rectangle evaluation value calculation unit 252 is smaller than a certain value β1, it is determined that the character is not a character, and w = 0. And 2. If x is larger than a certain value α2, it is determined that the character is character-like, and w = 1. 3. When y is larger than a certain value β2, it is highly possible that the character is a character. 4. The weight value w monotonically increases with respect to x and y.

Next, the calculated weight value w and the output selection unit 264
Lattice storage unit 25 stores candidate characters input from
2 to the corresponding memory.

After the weight values for all the combination rectangles and the candidate characters at that time are stored and held in the weighted lattice storage unit 252 as described above, the cutout position search unit 2
43 determines the cutout position and the recognition result as follows.

From the paths from the start to the end of the weighted lattice storage unit 252, the path with the largest weight per rectangle on the path is selected, the combined rectangle at that time is set to one character, and the candidate character is Output as a recognition result 206.

[0030]

However, in the above configuration, a candidate character string is obtained from a similarity value obtained only from pattern matching of a single character.
It is not always linguistically correct. Further, when one or more characters are included in one basic rectangle in the rectification process, the character cutout result and the recognition result are greatly incorrect. In addition, since the evaluation is performed according to a certain rectangular evaluation criterion, when the optimum rectangular evaluation criterion differs for each document and each region, such as a printed document or a handwritten document including a character string of an indefinite pitch, an error in the rectangular evaluation value may occur. It is easy to occur.

For the above reasons, there is a problem that the recognition accuracy is deteriorated due to the inappropriateness of the character evaluation criterion of the entire processing system.

The present invention has been made in view of the above problems of the conventional character recognition apparatus, and aims to improve the processing accuracy of the entire processing system, reduce erroneous segmentation of characters, and provide a character recognition apparatus with good recognition accuracy. I do.

[0033]

According to a first aspect of the present invention, a character string component is divided by a portion having a small histogram value formed by projecting an input character string component onto an axis in the character string direction. A plurality of candidate characters are selected by performing a feature extraction from character components inside the rectangles and calculating a similarity representing which character is close to which character for each character. Candidate character selecting means, language processing means for calculating the evaluation value of certainty as a sentence by comparing the result of the candidate character selecting means with a word dictionary and a grammar dictionary, and determining the character as a character from the result of the language processing means A determined character detecting means for detecting the determined rectangle; a character rectangle information extracting means for extracting a feature of the determined rectangle as a character from the result of the determined character detecting means; Based on the rectangle information obtained by the character rectangle information extraction unit, a rectangle not yet determined as a character is integrated and divided, and the result is fed back to the rectangularization processing unit. A character recognition device comprising a recognition result output unit that outputs a result of the detection unit as a recognition result, wherein the language processing unit includes a plurality of candidate characters that are combined with each other to check consistency with a word dictionary. Candidate word determining means for determining candidate words, and a candidate phrase determining means for estimating a phrase from the result of the candidate word determining means and determining a plurality of candidate phrases by checking the consistency between the estimated phrase and the grammar dictionary Means, for the candidate clauses listed as candidates by the candidate clause determining means, the similarity average of the characters constituting the candidate clauses, Square of the difference between the number of characters and a predetermined value of words formed, the number of characters in the candidate phrase,
After calculating an evaluation value based on an evaluation function represented by the frequency of words constituting the candidate clause, a phrase group having a fixed length formed by combining the continuous candidate clauses, the evaluation per phrase A phrase evaluation unit for selecting a phrase group having the best value, determining the head phrase of the phrase group, and using an evaluation value of the determined phrase as an evaluation value representing linguistic correctness as a phrase, The rectification processing means receives the feedback from the rectangle division integration means, outputs a rectangle newly generated by the rectangle division integration means to a candidate character selection means in a part which has not been determined as a character yet, and The processing of the processing unit, the confirmed character detection unit, and the character rectangle information extraction unit is performed until all rectangles are confirmed as characters or repeated a certain number of times. A character recognition apparatus characterized.

According to a second aspect of the present invention, the rectangle dividing and integrating means includes an aspect ratio of a rectangle which is a circumscribed rectangle candidate of a character generated by a character extraction process from an input character string, an area of the rectangle, and a character occupying the rectangle area. A divided rectangle candidate selecting means for selecting a rectangle to be divided (a rectangle including one or more character components) based on an evaluation function represented by a ratio of an area of a portion other than a component, and the divided rectangle candidate Rectangle dividing means for dividing the rectangle selected by the selection means,
2. The character recognition device according to claim 1, further comprising: a rectangle integration unit that determines an optimal combination of rectangle integration by combining all rectangles including a rectangle generated by the rectangle division unit. It is.

According to a third aspect of the present invention, a character string component is divided into a plurality of rectangles by dividing an input character string component into small portions of a histogram value formed by projecting the character string component on an axis in the character string direction. Rectangularization processing means, and candidate character selection means for extracting a plurality of candidate characters by performing feature extraction from character constituent components inside the rectangles and calculating a similarity representing how close to each character for each character, A language processing means for calculating an evaluation value of certainty as a sentence by comparing the result of the candidate character selection means with a word dictionary and a grammar dictionary; and a determination for detecting a rectangle determined as a character from the result of the language processing means Character detection means, character rectangle information extraction means for extracting a rectangular feature determined as a character from the result of the fixed character detection means, and the fixed character detection means A writer estimating means for determining a writer of an input character string or a font of a printed character based on the result and the rectangle information obtained by the character rectangle information extracting means, and outputting the result to the candidate character selecting means; Based on the rectangle information obtained by the character rectangle information extraction unit, a rectangle not yet determined as a character is integrated and divided, and the result is fed back to the rectangularization processing unit. A recognition result output means for outputting a result of the detection means as a recognition result, wherein the writer estimation means comprises a character written by each writer or a font habit of a printed character. (The distribution of the characteristic amount of each character, the aspect ratio of the circumscribed rectangle of the character, and the size of the character). A writer evaluation value calculating means for calculating an evaluation value indicating how close to a writer or a font of a printed character of the writer dictionary by checking consistency with the determined character and the writer A writer determining unit that specifies the writer of the input character string or the font of the print character according to the result of the evaluation value calculating unit, wherein the rectification processing unit receives feedback from the rectangular division integration unit and determines the character as a character. The rectangle newly generated by the rectangle dividing and integrating unit is output to the candidate character selecting unit in the portion that has not been processed, and the processing of the language processing unit, the fixed character detecting unit, and the character rectangle information extracting unit is performed in a manner that all the rectangles are characters. The character recognition device is characterized in that the character recognition device is determined or repeated up to a certain number of times.

[0036]

According to the first aspect of the present invention, a character string which is considered to be linguistically correct in a character string composed of a plurality of candidate characters output by the candidate character selecting means is set to "per character in a phrase". The higher the degree of similarity, the higher the probability of correctness as a phrase. "," The greater the number of characters in a word in a phrase, the higher the possibility of correctness as a phrase. "," The closer the number of characters in a phrase to a certain value, the more correct the phrase. Probability is high.",
Based on the assumption that the higher the frequency of occurrence of words that form a phrase in the input character string is, the more likely it is to be correct as a phrase, the result of calculating the linguistic certainty evaluation value for each phrase is determined. The character detection means detects the determined character, and the recognition result output means outputs the result as a recognition result, while the character rectangle information detection means detects the rectangle information of the determined character and optimizes the rectangle evaluation function. The rectangle division integrating means obtains a new basic rectangle by dividing and integrating the basic rectangles of the undecided parts based on the optimized rectangle evaluation function, and performs the processing after the candidate character selection means again.

The above processing is repeated until all the rectangles are finally determined linguistically as characters.

As described above, since the character evaluation result is further evaluated by language processing and the recognition result is determined, the recognition accuracy is improved. In addition, since the result is used to estimate the evaluation function itself for evaluating the character likeness of a rectangular character, optimal processing can be performed for each document.

According to the second aspect of the present invention, when a rectangle that is a circumscribing rectangle candidate for a character includes one or more character components, the rectangle is written horizontally (vertically). Vertical), the area of the rectangle is likely to be larger than other rectangles, and the ratio of the area of the portion other than the character component to the area of the rectangle is likely to be constant. " , A rectangle (1
By dividing a rectangle including character components equal to or more than characters, erroneous cutout of characters can be reduced.

According to the third aspect of the present invention, with the above-described structure, the first
Since the character evaluation result is further evaluated by language processing in the same manner as in the present invention, the recognition result is determined, so that the recognition accuracy is increased. Further, by using the result to specify the writer or the font of the printed character, the specified writer or the character dictionary for the font of the printed character is used as the character dictionary when determining the candidate character in the candidate character selecting means. By doing so, the accuracy of the candidate character selecting means is improved, and more optimal processing can be performed.

That is, the fixed character detecting means detects a character in a part considered to be linguistically correct as a fixed character, and the recognition result output means outputs the result as a recognition result. And optimize the rectangle evaluation function. The rectangle division integrating means obtains a new basic rectangle by dividing and integrating a basic rectangle of a part which has not yet been determined based on the optimized rectangle evaluation function, and performs the processing after the candidate character selection means again. The writer estimation unit optimizes the dictionary used by the candidate character selection unit by using information obtained from the determined character.

[0042]

Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 shows a configuration diagram of a character recognition device according to an embodiment of the third invention. The first and second aspects of the present invention constitute constituent elements of the third aspect of the present invention, and will be described simultaneously in Examples of the third aspect of the present invention.

In FIG. 1, an input character string 201 is an input image character, a rectification processing unit 202 is a processing unit for rectifying the input character, and a candidate character selection unit 1 is a unit for selecting candidate characters. The language processing unit 2 linguistically processes the candidate character, the fixed character detection unit 3 detects a fixed character, the character rectangle information extraction unit 4 detects character rectangle information, and the rectangle division integration unit 5 Is a unit that divides and integrates a rectangle, the recognition result output unit 6 is a unit that outputs a recognition result, and the writer estimation unit 7 is a unit that estimates a writer.

The operation of the character recognition device of the present embodiment configured as described above will be described below.

The rectification processing unit 202 has the same components as in the conventional character recognition device, and stores the basic rectangle in the basic rectangle storage unit 217 in the same manner as in the conventional character recognition device.

FIG. 2 is a specific block diagram of the candidate character selection unit 1. The feature extraction unit 271 is a unit for extracting features, the character dictionary 272 is a dictionary storing character information, and the similarity calculation unit 273 is a character extraction unit. A unit that calculates the similarity with reference to the dictionary 273, the character selection unit 21 selects a character with reference to the similarity, and the candidate character storage unit 22 stores a candidate character.

The operation of the candidate character selection section 1 is as follows.
is there. The feature extraction unit 271 is based on the basic rectangle storage unit 217.
Referring to the coordinate position of this rectangle, the pixel within the coordinate position is rectangular.
Read from the character string data storage unit 211 in the shaping processing unit 202.
In the same way as the conventional character recognition device,
And outputs it to the similarity calculation unit 273. Next, similarity calculation
The output unit 273 obtains a representative feature vector of each character from the character dictionary 272.
And read it in the same way as with a conventional character recognition device.
To calculate the similarity with each character and output it to the character selection unit 21
You. The character selection unit 21 selects m characters in descending order of similarity.
The character and its similarity are output to the candidate character storage unit 22, and the candidate
The character storage unit 22 stores m candidate characters per rectangle and the similarity.
Remember.

FIG. 3 is a specific configuration diagram of the language processing unit 2 according to the first embodiment of the present invention. In FIG. 3, a candidate word determination unit 31 determines a candidate word, and a candidate phrase determination unit 32
A candidate phrase determining unit, a phrase evaluation determining unit 33 is a unit for evaluating phrases, a candidate character combining unit 301 is a unit for combining candidate characters, a word matching unit 302 is a unit for matching words, and a word dictionary 303 is a word dictionary. And a candidate word storage unit 304 is a unit for storing candidate words and a candidate phrase output unit 305
Is a candidate phrase output unit, a phrase matching unit 306 is a unit that matches phrases, a grammar dictionary 307 is a dictionary that stores grammar, a candidate phrase storage unit 308 is a unit that stores candidate phrases, and a phrase evaluation value calculation unit 309 is a phrase evaluation value. Calculating unit, candidate phrase combining unit 3
Reference numeral 10 denotes a unit for combining candidate phrases, and a phrase determining unit 311 denotes a unit for determining a phrase.

The candidate character combination unit 301 selects continuous rectangular candidate characters from the candidate character storage unit 22 one by one for each rectangle, and outputs the result of combining one or more selected characters to the word matching unit 302. . The above combinations are comprehensively performed, and all combinations are output to the word matching unit 302. In the word dictionary 303, each word is expressed in a tree structure so that each word can be matched, and it is assumed that each word has its part of speech and discrimination symbol attached and stored. The word matching unit 302 determines whether the combination of the input character and the word stored in the word dictionary 303 match or not for each character, and determines the character string of the "character combination" in which all the characters match. And outputs the position, classification symbol, and part of speech to the candidate word storage unit 304. The candidate word storage unit 304 sequentially stores and holds the inputs from the word matching unit 302.

The candidate phrase output unit 305 stores the candidate word memory 30
4 is read out, the candidate phrase is determined from the candidate words at adjacent positions, and output to the phrase matching unit 306. It is assumed that the grammar dictionary 307 stores information such as the inflected forms of words and connections with particles. The phrase matching unit 306 determines whether the input candidate phrase matches information stored in the grammar dictionary 307, and determines the position of the matched phrase and the candidate character storage unit 22.
From the number k of characters constituting the phrase, the average similarity d and the number 1 of characters constituting the words in the phrase, and the frequency q of the word in the character string, the candidate phrase storage unit 30
8 is output. The candidate clause storage unit 308 includes a memory for storing and holding the above information of the input candidate clause and a memory for storing and holding the evaluation value fed back from the clause evaluation value calculation unit 309 as described below. Is stored in those memories.

The phrase evaluation value calculation unit 309 reads the above-described phrase information stored and held in the candidate phrase storage unit 308, and calculates H in (Equation 4) as an amount representing the linguistic correctness of the phrase. Is output to the candidate phrase storage unit 308 as the phrase evaluation value of the corresponding phrase.

[0053]

(Equation 4) H = a1 · d + a2 · k + a3 · (L−1) ^Two + A4 · q Where d is the average similarity of the characters constituting the phrase, k is the number of characters of a word constituting a phrase, L is a constant value, l is the number of characters in the phrase, q is the frequency of the words that make up the phrase, a1, a2, a3, a4 are normalization coefficients, and a1, a2,
a4 has a positive value and a3 has a negative value.

(Equation 4) indicates that "the greater the similarity of the constituent characters, the more correct the phrase. The greater the number of words, the more correct the phrase. The closer the number of characters in the phrase to L, the better. The higher the frequency of the word, the more reliable the phrase. It seems. "

After the evaluation values have been calculated for all the candidate phrases stored in the candidate phrase storage unit 308, the candidate phrase combination unit 310 uses the position information of the candidate phrases stored in the candidate phrase storage unit 308 to calculate the evaluation values. A “phrase combination” composed of a plurality of continuous phrases so as to have a fixed length is extracted, and the result is output to the phrase determination unit 311. Therefore, a plurality of “phrase combinations” that start at the same position and end at the same position are input to the phrase determination unit 311. In the initial state, the phrase combination is assumed to be extracted in order from the beginning of the input character string 201. Further, each time a phrase is determined by the phrase determining unit 311 as described below, the beginning of the extracted phrase is determined. Move to the next position.

The phrase determining unit 311 selects a “phrase combination” that maximizes the phrase evaluation value of each of the constituent phrases from the input “phrase combination”, and sets the first phrase as the determined phrase as position information. The candidate character and the phrase evaluation value are output to the confirmed character detection unit 3.

By the above processing, the input character string 201 is divided into clause units, and each clause is given a clause evaluation value as a quantity representing the linguistic correctness of the clause. The confirmed character detection unit 3 determines candidate characters in a phrase having a phrase evaluation value equal to or greater than a certain value among the input phrases as a confirmed character of the recognition result.
The information is output to the character rectangle information extraction unit 4 and the recognition result output unit 6.
It is assumed that the character rectangle information extraction unit 4 and the recognition result output unit 6 also output position information of characters in a phrase.

The recognition result output unit 6 outputs the determined character input from the determined character detection unit 3 as a recognition result at the position where it is input along with the determined character.

The character rectangle information extracting section 4 extracts rectangle information from the determined character. FIG. 4 is a specific configuration diagram of the character rectangle information extraction unit 4. The squareness calculation unit 281 calculates the squareness, the blankness calculation unit 282 calculates the blankness, and the rectangular area calculation unit 41 calculates the squareness. A section for calculating a rectangular area, a squareness evaluation function determining section 42 for determining a squareness evaluation function,
The blankness evaluation function determining unit 43 is a unit that determines a blankness evaluation function, the rectangular area evaluation function determining unit 44 is a unit that determines a rectangular area evaluation function, and the evaluation function storage unit 45 is a unit that stores an evaluation function.

The squareness calculator 281 calculates the squareness of the rectangle of the determined character from the basic rectangle storage unit 217 with reference to the input positional information of the determined character, and outputs it to the squareness evaluation function determining unit 42. Similarly, the blankness calculating unit 282 refers to the position information of the input fixed character, and obtains, from the character string data storage unit 211 in the rectification processing unit 202, a pixel not included in the character component among the pixels in the coordinate position. The ratio of the number of pixels to the total number of pixels (blank degree) is calculated and output to the blank degree evaluation function determining unit 43. The rectangular area calculation unit 41 refers to the input positional information of the determined character, obtains the total number of pixels in the coordinate position from the character string data storage unit 211 in the rectangularization processing unit 202, and sends it to the rectangular area evaluation function determination unit 44. Output.

For example, FIG.
And the histogram of the input squareness values as shown
To determine the form of the evaluation function. And the histogram values
Squareness evaluation by determining the evaluation value by normalization
Determine the valence function ρ1. Determine blankness evaluation function in the same way
The unit 43 calculates blanks from the histogram of the input blankness values.
The degree evaluation function ρ2 is determined. Rectangular area evaluation function determination unit 4
4 is a rectangular surface from the histogram of the input rectangular area values.
The product evaluation function ρ3 is determined.

The evaluation function determined by the squareness evaluation function determination unit 42, the blankness evaluation function determination unit 43, and the rectangular area evaluation function 44 is stored in the evaluation function storage unit 45.

The evaluation function determined as described above is
Estimated from linguistically determined characters from force string 201
It is considered that the accuracy as an evaluation function is very high.
available.

FIG. 6 is a specific configuration diagram of the rectangular division integration section 5 of the second invention.

The divided rectangle candidate selecting section 61 reads a rectangle that has not yet been determined as a character among the rectangles stored and held in the basic rectangle storage section 217 and reads the squareness r,
The blankness s and the area M are obtained. Next, a rectangle satisfying (Equation 5) is selected from the evaluation functions stored in the evaluation function storage unit 43 as a rectangle considered to include a plurality of characters.

[0066]

Ρ1 (r) <λr and ρ2 (s)> λs and ρ3 (M)> λM where λr, λs and λM are appropriate values.

(Equation 5) indicates that "a rectangle that is longer than a certain length, has a large area, and has a large blank degree is likely to include a plurality of characters."

The divided rectangle candidate selecting section 61 outputs positional information of the selected rectangle to the rectangular dividing section 62.

FIG. 7 is a specific configuration diagram of the rectangular division section 62.

The marginal distribution calculating section 212 calculates a value obtained by accumulating the values held in the character string data storage section 211 corresponding to the position input from the divided rectangle candidate selecting section 61 in the vertical direction, and calculates the result by Gaussian cyan. Output to the filter processing unit 213. The Gaussian filter processing unit 213 performs smoothing by convolving a Gaussian function with the vertical cumulative histogram of the character data obtained by the marginal distribution calculating unit 212, thereby removing high-frequency components of the cumulative histogram and removing noise. Reduce the effect. The smoothed cumulative histogram is input to the division point detection unit 71. The division point detection unit 71 continuously detects a portion where the value of the cumulative histogram is equal to or less than a certain value (threshold), and further divides the coordinate position where the value of the cumulative histogram becomes minimum in that portion into a rectangle which is currently focused on. The data is output to the divided rectangle column direction peripheral distribution calculation unit 72 and the divided rectangle storage unit 74 as points.

At this time, it is assumed that the threshold is set so that all the rectangles formed by the division do not satisfy (Equation 5).

Receiving the input from the division point detecting section 71, the division rectangle column direction peripheral distribution calculation section 72 calculates the left end coordinates of the basic rectangle and the input division points, and the input division points and the right end coordinates of the basic rectangle. The character data of the portion to be sandwiched is shaded separately in the horizontal direction, and the result is output to the divided rectangle upper / lower end detection unit 73. When there are a plurality of division points, the same processing is performed on the character data between the division points.

As shown in FIG. 13, the divided rectangle upper / lower edge detecting section 73 obtains the maximum value and the minimum value of the portion formed by the character component of the oblique result from the divided rectangular column direction peripheral distribution calculating section 72, and The result is output to the divided rectangle storage unit 74 as the upper and lower ends of the corresponding divided rectangle.

The divided rectangle storing section 74 displays the divided rectangles by their left and right ends and upper and lower ends by the above processing, and stores them in the order of detection.

FIG. 8 is a specific configuration diagram of the rectangular integration unit 63. The operation of the rectangular integration unit 63 will be described below with reference to the drawings.

FIG. 9 is a specific configuration diagram of the integrated weighted lattice creating section 81 of the rectangular integrating section 63, where 91 is a rectangular combination section, and 92 is an integrated weighted lattice storage section. The rectangle combination unit 91 reads coordinate values of rectangles that are not determined as characters held in the basic rectangle storage unit 217 and the divided rectangle storage unit 74, and determines the outline width of a rectangle obtained by combining adjacent rectangles with a character. All combinations that are equal to or less than a fixed multiple of the column height are detected, and the results are output to the integrated weighted lattice storage unit 92.

The integrated weighted lattice storage unit 92 stores the data in FIG.
As shown in FIG. 6, in order to store and hold the coordinates of the upper, lower, left and right ends of the combination rectangle and to store and hold the weight values corresponding to each combination rectangle, for all the combinations of rectangles detected by the rectangle combination unit 81 as shown in FIG. Allocate memory for

The integrated weight value calculation unit 82 is provided in the evaluation function storage unit 4
5 is calculated from the evaluation function stored in the storage unit 5 and the information on the combination rectangle stored in the integrated weighted lattice storage unit 92, and the integrated weight is calculated as a weight value corresponding to the combination rectangle. The result is output to the lattice storage unit 92.

The cut-out position search unit 243 selects a route having the maximum weight per rectangle on the route from among the routes from the start end to the end of the weighted lattice storage unit, and integrates the combined rectangle at that time into a rectangle. As a result, the basic rectangle storage unit 2
17 is output.

In the character recognition device according to the first aspect of the present invention, the above-described processes after the candidate character selection unit 1 are repeated again until all rectangles are determined as characters and output by the recognition result output unit 6. . If the number of times of the repetition processing becomes equal to or more than a certain value, the phrase determining unit 31 at that time is used.
1 may be output from the recognition result output unit 6 as a recognition result, or the recognition of a part that has not been determined as a character may be rejected.

As described above, according to the first embodiment of the present invention, the candidate character selection unit 1 first converts the input character string divided into appropriate basic rectangles by the rectangularization processing unit 202 into one basic rectangle. A plurality of candidate characters are added, and the language processing unit 2 determines a character string formed by the candidate characters. Then detected as linguistically correct the confirmed character detecting unit 3 is determined character a character portion seems, while the recognition result output unit 6 is output as the recognition result, the rectangle information of the character rectangle information extracting unit 4 is fixed character And optimize the rectangle evaluation function. Based on the optimized rectangle evaluation function, the rectangle division integration unit 5 obtains a new basic rectangle by dividing and integrating a basic rectangle of a part that has not been determined yet, and again performs the processing of the candidate character selection unit 1 and subsequent steps. Do. Characters determined linguistically in this way are used as recognition results, and those characters are used to determine the evaluation function being processed, so that a highly accurate evaluation function can be automatically determined for each document.

In the character recognition device according to the third aspect of the present invention, the writer estimating unit derives the writer or print character from the character part linguistically determined by the language processing means of the first aspect of the present invention as described below. Is determined, the content of the character dictionary 272 in the candidate character selection unit 1 is changed from the result, and the processing from the candidate character selection unit 1 described above is repeated. Therefore, at the time of the repetition processing in the candidate character selection unit 1, the character dictionary 272 is optimized according to the document, so that more accurate processing can be performed.

As shown in FIG. 1, the writer estimating unit 7 has inputs from the rectangularization processing unit 202, the fixed character detecting unit 3, the character rectangle information extracting unit 4, and an output to the candidate character selecting unit 1.

FIG. 10 is a specific configuration diagram of the writer estimating unit 7 according to the third embodiment of the present invention.

The scribe dictionary 101 is, as shown in FIG.
By writer who is expected to use a character recognition device in advance,
Or sentences by writer group with similar distribution of character features
Character dictionaries 111 _{1 to} 111 _t (t is a writer or a group of scribes
Number) and consists squareness dictionary 112 ₁ to 112 _t
You. However, each character has the squareness dictionary 112 _{1 to} 112 _t
It is assumed that the value of the squareness is stored and held.

The writer evaluation value calculation unit 102 includes a feature extraction unit 271 and a similarity average calculation unit 121 as shown in FIG. 10, and calculates an evaluation value as follows.

First, the feature extracting unit 271 reads a pixel value from the character string data storage unit 211 in the rectification processing unit 202 based on the positional information of the fixed character input from the fixed character detecting unit 3 and selects candidate characters. In the same manner as in the unit 1, the feature amount vector of the determined character is obtained and output to the similarity average calculating unit 121.

The similarity average calculating section 121 includes the feature extracting section 27
1 and the character vector of the determined character from the character rectangle information extraction unit 4. The similarity average calculating unit 121 performs the similarity calculating unit 27 in the candidate character selecting unit 1 with respect to the feature amount vectors of all the input fixed characters.
3, the similarity between the corresponding determined character in the character dictionaries 111 _{1 to} 111 _t and the representative vector is calculated, and the sum μ of the similarity for each character dictionary is calculated. And the squareness of the squareness of the corresponding fixed character in the squareness dictionaries 112 _{1 to} 112 _t are calculated, the sum ν of the squared errors for each of the character dictionaries is calculated, and Ei represented by (Equation 6) is obtained. Is output to the writer determining unit 103 as a writer evaluation value indicating how close the determined character is to the character dictionary 111 _i .

[0089]

Ei = (μ + b1 · ν) / N where b1 is a normalization constant, and N is the number of input confirmed characters.

The writer determining section 103 selects a character dictionary 111max in the writer dictionary 101 having the largest writer evaluation value Emax from the input from the writer evaluation value calculating section 102.
The contents of the character dictionary 272 in the candidate character selection unit 1 are converted into the contents of the character dictionary 111max.

Note that this conversion is based on the maximum writer evaluation value E.
It may not be performed until max exceeds a certain value.
Also, even if the writer dictionary 101 has a dictionary of other information obtained from the character string (inter-character pitch, height position of the base of the circumscribed rectangle of the character, height position of the center of gravity of the circumscribed rectangle of the character, etc.) It is clear that the effect can be obtained.

The various means of the present invention are generally realized by software using a computer, but may be realized by using a hardware circuit having each function.

[0093]

As described above, according to the present invention,
The evaluation function representing the character-likeness for processing can be automatically optimized for each document, the processing accuracy and the recognition rate can be improved, and the practical effect is large.

[Brief description of the drawings]

FIG. 1 is a configuration diagram of a character recognition device according to an embodiment of the present invention.

FIG. 2 is a configuration diagram of a candidate character selection unit 1 according to the present invention.

FIG. 3 is a configuration diagram of a language processing unit 2 according to the present invention.

FIG. 4 is a configuration diagram of a character rectangle information extraction unit 4 according to the present invention.

FIG. 5 is an explanatory diagram of a method of determining a squareness evaluation function according to the present invention.

FIG. 6 is a configuration diagram of a rectangular division integration unit 5 according to the present invention.

FIG. 7 is a configuration diagram of a rectangular division unit 62 according to the present invention.

FIG. 8 is a configuration diagram of a rectangular integration unit 63 in the present invention.

FIG. 9 shows an integrated weighted lattice creation unit 8 according to the present invention.
1 is a configuration diagram.

FIG. 10 is a configuration diagram of a writer estimation unit 7 in the present invention.

FIG. 11 is a configuration diagram of a character recognition device in a conventional example.

FIG. 12 is a configuration diagram of a rectangularization processing unit 201 of a character recognition device in a conventional example.

FIG. 13 is an explanatory diagram of an operation of a left and right end detection unit 214 in a conventional example.

FIG. 14 is an explanatory diagram of an operation of an upper and lower end detection unit 216 in a conventional example.

FIG. 15 is a configuration diagram of a rectangle integration recognition unit 203 in a conventional example.

FIG. 16 shows a weighted lattice creating unit 24 in a conventional example.
1 is a configuration diagram.

FIG. 17 shows a weighted lattice storage unit 25 in a conventional example.
FIG.

FIG. 18 is a configuration diagram of a character evaluation value calculation unit 204 in a conventional example.

FIG. 19 is a configuration diagram of a rectangular evaluation value calculation unit 205 in a conventional example.

[Explanation of symbols]

 1 candidate character selection unit (means) 2 language processing unit (means) 3 fixed character detection unit (means) 4 character rectangle information extraction unit (means) 5 rectangle division integration unit (means) 6 recognition result output unit (means) 7 writing Estimator (means) 21 Character selection unit (means) 22 Candidate character storage unit (means) 31 Candidate word decision unit (means) 32 Candidate phrase decision unit (means) 33 Clause evaluation decision unit (means) 301 Candidate character combination unit (Means) 302 Word matching unit (Means) 303 Word dictionary 304 Candidate word storage unit (Means) 305 Candidate phrase output unit (Means) 306 Clause matching unit (Means) 307 Grammar dictionary 308 Candidate phrase storage unit (Means) 309 Candidate clause Combination unit (means) 310 Clause evaluation value calculation unit (means) 311 Clause determination unit (means) 41 Rectangular area calculation unit (means) 42 Squareness function decision unit (means) 43 Blankness evaluation Function determination unit (means) 44 Rectangular area evaluation function determination unit (means) 45 Evaluation function storage unit (means) 61 Divided rectangle candidate selection unit (means) 62 Rectangular division unit (means) 63 Rectangular integration unit (means) 71 Division inspection Outgoing unit (means) 72 Divided rectangular column direction peripheral distribution calculating unit (means) 73 Divided rectangular upper / lower end detecting unit (means) 74 Rectangular storage unit (means) 81 Lattice creation unit with integrated weights (means) 82 Integrated weight calculating unit ( Means 91 Rectangle combination unit (Means) 92 Lattice storage unit with integrated weight (Means) 101 Writer dictionary 102 Writer evaluation value calculation unit (Means) 103 Writer determination unit (Means) 111 Character dictionary in Writer dictionary 101 112 Squareness dictionary in the scribe dictionary 101 121 Similarity average calculation unit (means) 201 Input character string 202 Rectangularization processing unit (Means) 203 Rectangular integration recognition unit (Hand) Stage) 204 character evaluation value calculation unit (means) 205 rectangular evaluation value calculation unit (means) 206 recognition result 211 character string data storage unit (means) 212 marginal distribution calculation unit (means) 213 Gaussian filter processing unit (means) 214 left and right Edge detection unit (means) 215 Column-direction marginal distribution calculation unit (means) 216 Upper and lower end detection units (means) 217 Basic rectangle storage unit (means) 241 Weighted lattice creation unit (means) 242 Weight value calculation unit (means) 243 Extraction position search unit (means) 251 Basic rectangle combination unit (means) 252 Weighted lattice storage unit (means) 271 Feature extraction unit (means) 272 Character dictionary 273 Similarity calculation unit (means) 274 Output selection unit (means) 281 Squareness calculator (means) 282 Blankness calculator (means) 283 Suitability calculator (means)

──────────────────────────────────────────────────の Continued on the front page (72) Inventor Shigeo Sakagami 1006 Kadoma Kadoma, Osaka Prefecture Inside Matsushita Electric Industrial Co., Ltd. (72) Inventor Toshio Niwa 1006 Kazuma Kadoma, Osaka Prefecture Inside Matsushita Electric Industrial Co., Ltd. JP-A-63-271588 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G06K 9/00-9/82

Claims (3)

(57) [Claims] A rectifying unit configured to divide the character string component into a plurality of rectangles by dividing the input character string component into small portions of a histogram value formed by projecting the input character string component on an axis in the character string direction; A candidate character selecting means for extracting a plurality of candidate characters by performing feature extraction from character constituent components inside the rectangles and calculating a similarity indicating how close to which character each character is, and the candidate character selecting means Language processing means for calculating an evaluation value of certainty as a sentence by comparing the result of the word dictionary with a grammar dictionary, and a fixed character detection means for detecting a rectangle determined as a character from the result of the language processing means, Character rectangle information extraction means for extracting a rectangular feature determined as a character from the result of the determined character detection means, and character rectangle information extraction means. Based on the obtained rectangle information, a rectangle not yet determined as a character is integrated and divided, and the result is fed back to the rectangularization processing means. A character recognition device comprising: a recognition result output unit that outputs a plurality of candidate words by combining candidate characters and checking consistency with a word dictionary. Word determination means, candidate phrase determination means for estimating a phrase from the result of the candidate word determination means, and determining a plurality of candidate phrases by checking the consistency between the estimated phrase and the grammar dictionary; and Means similarity average of the characters constituting the candidate clause for the candidate clauses listed as candidates by the means, the number of characters of the words constituting the candidate clause and a predetermined value After calculating an evaluation value based on the square of the difference, the number of characters of the candidate clause, and the frequency of the words constituting the candidate clause, a constant length that can be obtained by combining the consecutive candidate clauses For a phrase group, a phrase group with the best evaluation value per phrase is selected, the head phrase of the phrase group is determined, and the evaluation value of the determined phrase represents linguistic correctness as a phrase. The rectangularization processing means receives feedback from the rectangle division integration means, and generates a rectangle newly generated by the rectangle division integration means in a portion which has not been determined as a character yet. Output to the candidate character selection means, and determine whether the processing of the language processing means, the confirmed character detection means, and the character rectangle information extraction means is confirmed as all rectangles as characters. Alternatively, a character recognition device characterized by repeating a certain number of times. 2. The method according to claim 1, wherein the rectangular division integrating means includes an aspect ratio of a rectangle which is a circumscribed rectangle candidate of a character generated by a character cutout process from an input character string, an area of the rectangle, and a portion other than a character component occupying the rectangle area. A divided rectangle candidate selecting means for selecting a rectangle to be divided (a rectangle including one or more character components) based on the evaluation function represented by the area ratio of Rectangle dividing means for dividing the divided rectangle, and rectangle integrating means for determining an optimal combination of rectangular integration by combining all the rectangles including the rectangle generated by the rectangle dividing means. The character recognition device according to claim 1, wherein 3. A rectangle processing means for dividing a character string component into a plurality of rectangles by dividing an input character string component into small portions of a histogram value formed by projecting the character string component on an axis in the character string direction. A candidate character selecting means for extracting a plurality of candidate characters by performing feature extraction from character constituent components inside the rectangles and calculating a similarity indicating how close to which character each character is, and the candidate character selecting means Language processing means for calculating an evaluation value of certainty as a sentence by comparing the result of the word dictionary with a grammar dictionary, and a fixed character detection means for detecting a rectangle determined as a character from the result of the language processing means, Character rectangle information extracting means for extracting a feature of a rectangle determined as a character from the result of the determined character detecting means, and a result of the determined character detecting means and the character A writer estimating means for determining a writer of an input character string or a font of a printed character based on the rectangular information obtained by the shape information extracting means, and outputting the result to the candidate character selecting means; Based on the rectangle information obtained by the means, a rectangle not yet determined as a character is integrated and divided, and the result is fed back to the rectangularization processing means by a rectangular division integration means, and the result of the determined character detection means And a recognition result output means for outputting a character as a recognition result, wherein the writer estimation means comprises: a character written by each writer; A writer dictionary that stores and retains the distribution of the feature amount, the aspect ratio of the circumscribed rectangle of the character, and the size of the character, and a sentence that is determined to be linguistically correct with the writer dictionary. A writer evaluation value calculating means for calculating an evaluation value indicating how close to which writer or which font of a printed character is in the writer dictionary by seeing consistency with the writer dictionary; and Writer determination means for specifying the writer of the input character string or the font of the print character according to the result, wherein the rectangularization processing means receives feedback from the rectangular division integration means, and in a part which has not been determined as a character yet, A rectangle newly generated by the rectangle division integration unit is output to a candidate character selection unit, and the processing of the language processing unit, the fixed character detection unit, and the character rectangle information extraction unit is performed when all rectangles are determined as characters or A character recognition device characterized by repeating a certain number of times.