KR100718185B1 - Apparatus for recognizing character and method for operating the apparatus - Google Patents

Abstract

The character recognition method according to the present invention comprises the steps of: reading feature information and format information of a predetermined single image; Generating initial information, neutral information, or final information of the single image according to the feature value information and format information; Generating a plurality of pseudo words for reading the single image according to the initial information, the neutral information, or the final information, and calculating a non-split confidence value corresponding to each of the pseudo words; If the difference between the first-order non-partitioned reliability having the highest value and the second-order non-partitioned reliability having the next higher value is greater than or equal to a predetermined threshold value, the pseudo-first-partitioned reliability Recognizing a piece as the piece image; And when the difference between the first-rank reliability and the second-rank reliability is less than the threshold value, subdivides the single image to calculate a divided reliability corresponding to each of the pseudo words, and the divided reliability and the ratio corresponding to each of the pseudo words. And a step of recognizing a pseudo piece having the highest reliability as a result of multiplying the reliability to be obtained as the single image.

Character Recognition, Phoneme Segmentation, Phoneme Segmentation, Font, Holistic, Component

Description

Character recognition apparatus and its method {APPARATUS FOR RECOGNIZING CHARACTER AND METHOD FOR OPERATING THE APPARATUS}

BRIEF DESCRIPTION OF THE DRAWINGS Fig. 1 is a diagram showing an example of a method of character segmentation character recognition according to the prior art.

2 is a flow chart showing the overall flow of the phoneme-based character recognition method according to an embodiment of the present invention.

3 is a block diagram showing the configuration of a character recognition apparatus according to an embodiment of the present invention.

4 is a diagram illustrating a principle of generating single image feature value information according to an embodiment of the present invention;

FIG. 5 is a diagram illustrating a primitive recognizer for calculating each primitive class score according to an embodiment of the present invention.

FIG. 6 is a diagram illustrating a phoneme-divided image in which a phoneme division part phoneme-divided a piece image according to an embodiment of the present invention; FIG.

7 is an internal block diagram of a general purpose computer system that may be employed to implement the character recognition method in accordance with the present invention.

<Explanation of symbols for the main parts of the drawings>

210: feature value reading step

220: format information reading step

230: non-partitioned reliability calculation step

240: comparison of the first-order non-partitioned reliability and the second-order non-partitioned reliability difference and threshold value

250: partition reliability calculation step

260: Step Recognition

The present invention relates to a method and a device for character-based character recognition, and more particularly, to classify subtitle characters of a video into characters, and to recognize the characters as they are without separating the characters of the characters. Phoneme-based character that can recognize characters more accurately regardless of changes or inaccuracies of character fonts by appropriately utilizing the improved segmentation recognition method that recognizes the word by separating the phoneme correctly. A recognition method and apparatus are provided.

Recently, the IT industry has been spreading various video media such that it is no exaggeration to enter the video service and device war. As new video services such as satellite DMB broadcasting, terrestrial DMB broadcasting, data broadcasting, and internet broadcasting have begun, the video service industry is emerging as blue ocean in all areas of IT such as telecommunication, internet service, and digital equipment.

Satellite / terrestrial DMB broadcasting has opened the era of TV in the hand, and mobile carriers are also expanding video services through their data broadcasting through partnerships with the content industry. In addition, Internet portal sites also provide users with their own videos and affiliated videos through their own production or partnership with content companies.

In addition, TV portals are recently being serviced. TV portals are a preliminary step of Internet TV, and users can download movies or dramas provided by the portal site in the form of video on demand (VOD) through a PC, a laptop or a mobile terminal. It means a service that can be streamed and watched. In the future, the demand for video content will increase further when the Triple Play Service (TPS) service, which can use the Internet, broadcast, and telephone all over the Internet through a broadband integrated network, is in full swing.

As such, for younger generations who are familiar with the video culture, video is not an option but an essential feature, and the video industry is now becoming the most competitive IT industry. Accordingly, the market for moving picture playback terminals such as DMB terminals and PMPs is expanding day by day.

Mobile terminal companies are competitively launching satellite DMB phones and terrestrial DMB phones, and MP3 player companies are developing and launching various models of PMPs that support DMB broadcasting. Recently, MP3 players also support a video playback function by mounting a small LCD such as 2 inches as a display means. These various video supporting terminals are sure to evolve into convergence products that support each video service in one terminal.

As such, as the functions of the video service and the terminal are developed day by day, the desire of the user for the convenience of the service is also increasing. That is, the user no longer requires the terminal to simply play the video, but wants to be provided with a video service supporting more various additional functions.

For example, there is a video summary service. The video summary service refers to a service that generates a summary image of the video and provides it to the user when the user does not have time to watch all the videos for several hours in a busy daily life. . Such a video summary service is suitable for the daily life of busy modern people who watch a video through their mobile terminals by using a short break while traveling, such as commuting, it is expected that the video summary service will be gradually expanded.

The video summary service or video search service as described above may be implemented through predetermined image information. Subtitle characters of the video may be used as the image information. For example, subtitle text information may be used to summarize and search for a video such as news or sports.

In order to implement a video summary and search service using the caption characters, the caption characters should be recognized quickly and accurately. In addition, scanning operations such as business card recognition and book scanning using a scanner or the like also assume accurate character recognition.

A conventional character recognition method includes a phoneme segmentation character recognition method that recognizes each phoneme by dividing a phoneme of a single letter and then recognizes the phoneme by combining the recognized phonemes. According to the character recognition method of the phoneme unit, as shown in (a) of FIG. 1, since the case of recognition is a stronger character than the case of the letter unit, the advantage of enabling more accurate character recognition is provided. have.

However, since the fonts of the characters are various, an error occurs frequently in the segmentation operation. For example, as shown in (b) of FIG. 1, when using the projection profile method, the words "three" may be separated into the letters "s", "s", and "s". In addition, as shown in Fig. 1C, in the case of " much " and " m " As described above, an error occurs frequently in the conventional phoneme segmentation character recognition method, and it is pointed out that it is difficult to expect an accurate phoneme segmentation especially in the segmentation operation of vowels.

In addition, in the case of recognizing a letter or a character through the conventional type classification, the type classification is performed through the extraction of the stroke, and thus, in the case of a font having a thick stroke, the stroke extraction operation itself is difficult. In addition, since the type classification operation itself does not accompany accuracy, errors frequently occur, and thus, there is a high possibility of overall misrecognition.

In addition, in the conventional vowel structure and rule-based phoneme separation method, the phoneme is separated by extracting the horizontal and vertical vowels. However, in the case of a printed body having a thick stroke, the stroke extraction operation is difficult, and an error occurs frequently in the segmentation.

In addition, in the case of the conventional phoneme separation method using the background thinning method, the background thinning algorithm is so sensitive to noise that it is difficult to apply to a video with a high noise such as a video character.

According to the problems of the prior art, there is a demand for the development of a character recognition method capable of recognizing any character more accurately and quickly, regardless of the change in the character font.

The present invention has been made to improve the prior art as described above, by reducing the execution time of the character recognition operation by recognizing the character through a non-divisional character recognition method that recognizes the character as it is without dividing the character into characters. It is an object of the present invention to provide a phoneme-based character recognition method and apparatus.

In addition, the present invention, if the non-divisional phoneme character recognition result is ambiguous, after recognizing the character through a divisional phoneme recognition method for dividing the recognition of the phoneme of the character, and comparing the result with the non-divisional phoneme recognition result By recognizing, it is an object of the present invention to provide a phoneme-based character recognition method and apparatus capable of accurately recognizing a character regardless of a change in a character font.

According to the present invention, when the phoneme is divided and recognized, the phoneme is more accurately detected by using a phoneme type information and curvature information to detect a touch point, which is a connection point with a vowel, among the phonemes. An object of the present invention is to provide a method and apparatus for characterization-based character recognition capable of performing division.

In order to achieve the above object and to solve the problems of the prior art, the character recognition method according to the present invention comprises the steps of: reading feature information and format information of a predetermined single image; Generating initial information, neutral information, or final information of the single image according to the feature value information and format information; Generating a plurality of pseudo words for reading the single image according to the initial information, the neutral information, or the final information, and calculating a non-split confidence value corresponding to each of the pseudo words; If the difference between the first-order non-partitioned reliability having the highest value and the second-order non-partitioned reliability having the next higher value is greater than or equal to a predetermined threshold value, the pseudo-first-partitioned reliability Recognizing a piece as the piece image; And when the difference between the first-rank reliability and the second-rank reliability is less than the threshold value, subdivides the single image to calculate a divided reliability corresponding to each of the pseudo words, and the divided reliability and the ratio corresponding to each of the pseudo words. And a step of recognizing a pseudo piece having the highest reliability as a result of multiplying the reliability to be obtained as the single image.

Hereinafter, with reference to the accompanying drawings will be described an embodiment of the present invention;

First, after briefly explaining the entire flow of the phoneme-based character recognition method according to the present invention with reference to FIG. 2, the phoneme-based character recognition method of the character recognition apparatus according to the present invention will be described in detail with reference to FIGS. 3 to 5. Shall be.

2 is a flowchart illustrating the overall flow of a phoneme-based character recognition method according to an embodiment of the present invention.

The character recognition apparatus according to an embodiment of the present invention may recognize a character according to a non-partitioned phoneme recognition method or a divisional phoneme recognition method. When a single image such as a caption character of a moving image is input, the character recognition device reads feature information of the single character (step 210). The feature value information may be implemented as vector information of a pixel included in each mesh image obtained by dividing the single image into a plurality.

The character recognition apparatus reads the format information of the letter (step 220). The format information of the single image may be set to any one of a Korean format consisting of first to sixth formats and a seventh format of an English letter, a number, or a special character.

The character recognition apparatus generates initial information, neutral information, or final information of the single image according to the feature value information and format information. The character recognition apparatus generates a plurality of pseudo words for reading the single image according to the initial information, the neutral information, or the final information, and generates a non-divided confidence value corresponding to each of the pseudo words. Calculation (step 230).

The character recognition apparatus compares the difference between the first-order non-partitioned reliability having the highest value and the second-order non-partitioned reliability having the next higher value among the non-partitioned reliability corresponding to each pseudo word with a predetermined threshold value (step ( 240)). In step 240, when the difference of the reliability is greater than or equal to the threshold value, the pseudo word having the first non-segmented reliability is recognized as the piece image (step 260).

In operation 240, when the comparison result shows that the difference in reliability is less than the threshold value, the character recognition device may subdivide the single image to calculate a split reliability corresponding to each of the pseudo characters (step 250). In this case, the character recognition apparatus multiplies the split reliability and the non-segment reliability corresponding to each of the pseudo digits with each other and recognizes the pseudo gram having the highest reliability as the single image (step 260). .

Hereinafter, a configuration of a character recognition apparatus and a phoneme-based character recognition operation according to an embodiment of the present invention will be described in detail with reference to FIGS. 3 to 5.

3 is a block diagram showing a configuration of a character recognition apparatus according to an embodiment of the present invention.

Character recognition apparatus according to the present invention is a PVR (Personal Video Recorder), home server (Home Server), Smart Mobile Server (Smart Mobile Server), DVD player / recorder, PC, laptop, PDA, digital TV, or mobile communication terminal It can be implemented as either.

Character recognition apparatus according to an embodiment of the present invention is characterized in that the feature value information unit 310, the format information unit 320, the non-divisional phoneme recognition unit 330, the piece recognition control unit 340, and the divided phoneme recognition unit 350 It can be configured to include. The non-partitioned phoneme recognizer 330 may include a first recognizer 331, a neutral recognizer 332, and a final recognizer 333. The divided phoneme recognizer 350 may include a phoneme divider 351, a phoneme feature value information unit 352, a first recognizer 353, a neutral recognizer 354, and a seed recognizer 355.

The feature value information unit 310 reads feature value information of a predetermined single image among subtitle characters of an input video. The feature value information may be implemented as vector information of a pixel included in each mesh image after dividing the single image into a plurality of mesh images. This will be described in detail with reference to FIG. 4.

4 is a diagram illustrating a principle of generating single image feature value information according to an embodiment of the present invention.

According to an embodiment of the present invention, the feature value information unit 310 divides the input single image into a plurality of mesh images. For example, as illustrated in FIG. 4A, the feature value information unit 310 may divide the single image into 36 mesh images of 6 * 6.

The feature value information unit 310 reads vector information of pixels included in each of the divided mesh images. For example, a direction vector may be set as shown in FIG. 4B in a boundary area between a region where characters are displayed and a region where characters are not displayed among 5 * 3 mesh images 410 among the 36 mesh images.

The feature value information unit 310 reads angle information of the respective direction vectors displayed on the 5 * 3 mesh image 410. That is, the direction vector may be a first angle (1 degree to 45 degrees), a second angle (46 degrees to 90 degrees), a third angle (91 degrees to 135 degrees), a fourth angle (136 degrees to 180 degrees), Read which angle belongs to the fifth angle (181 degrees to 225 degrees), the sixth angle (226 degrees to 270 degrees), the seventh angle (271 degrees to 315 degrees), and the eighth angle (316 degrees to 360 degrees) do.

When angle information of the direction vector is read, the feature value information unit 310 generates feature value information of the 5 * 3 mesh image 410 using the angle information as shown in FIG. The feature value information includes a first angle region 421, a second angle region 422, a third angle region 423, a fourth angle region 424, a fifth angle region 425, and a sixth angle region ( 426, a seventh angular region 427, and an eighth angular region 428.

In the case of the 5 * 3 mesh image 410 as described above, three direction vectors having a sixth angle, two direction vectors having a seventh angle, and two direction vectors having an eighth angle among a total of eight direction vectors Since there are three, the feature value information unit 310 may generate feature value information of "00000323" corresponding to the 5 * 3 mesh image 410.

The feature value information unit 310 generates feature value information corresponding to all mesh images included in the single image. When the single image is divided into 36 mesh images as described above, the feature value information unit 310 may generate 288 digit feature value information corresponding to the single image.

3 again, the format information unit 320 reads format information of the single image. The format information may be implemented as format information according to a seven-type classification system consisting of alphabets, numbers, or special characters in Korean six formats. That is, the first to sixth forms correspond to Hangul, and the seventh form corresponds to alphabetic characters, numbers, or special characters.

Hangul included in the first format may be ga, b, da, or the like. Hangul included in the second format may be phrase, nu, or do. Hangul included in the third form may be an ear, a brain, a back, or the like. Hangul included in the fourth format may be Gan, Non, Dan, and the like. Hangul included in the fifth form may be a group, a rice field, or a money. The Hangul included in the sixth form may be Gin, divided, or Han. The first to third forms may be implemented in Korean, which does not include a base, and the fourth to sixth forms may be implemented in Korean, including a base.

The type information unit 320 may classify the type information by using a multi-class SVM.

The non-segmented phoneme recognition unit 330 generates the initial information, the neutral information, or the final information of the single image by referring to the feature value information and the format information. The non-partitioned phoneme recognizer 330 generates the initial information of the single image through the initial recognizer 331, generates the neutral information of the single image through the neutral recognizer 332, and uses the final recognizer 333. Generates final information of the single image.

Each letter of Hangeul may be configured to include any one or more of the initial, neutral, or final. The first class can be divided into 19 classes, the neutral class can be divided into 5 to 9 classes, and the final class can be divided into 27 classes.

The first classifier 331 includes 19 first classifiers respectively corresponding to the 19 first class. The neutral recognizer 332 includes five to nine neutral classifiers corresponding to the five to nine neutral classes, respectively. The species identifier 333 includes 27 species classifiers each corresponding to the 27 species classes.

Each of the initial classifier, the neutral classifier, and the final classifier may maintain class feature value information corresponding to each initial class, the neutral class, and the final class. For example, in the case of the a classifier among the initial classifiers, feature value information corresponding to each of the various types of a fonts may be maintained. This may be preset in various ways in advance according to the judgment of those skilled in the art.

The non-segmented phoneme recognition unit 330 substitutes all the single image into each class classifier to calculate a class score corresponding to each class classifier. For example, when the single image is "a", the a classifier compares a feature value information or type information of "a" with a feature value information maintained by the a classifier and compares the a according to the similarity. Class scores can be calculated. The a class score may be included in the initial information of the piece image.

As shown in FIG. 5, the a classifier 511 included in the initial recognizer 510 may calculate a class score corresponding to the image “a”. The classifier 512 may also calculate a class class score corresponding to "A". The classifier 519 may also calculate a class score corresponding to "A". In this case, each class score value may be calculated in order of a class score, a class score, and a class score.

In this manner, the non-partitioned phoneme recognizer 330 substitutes the single image into all class classifiers included in the initial recognizer 331, the neutral recognizer 332, and the final recognizer 333, respectively. Corresponding to each other, the class score of each class classifier can be calculated.

3 again, the piece recognition control unit 340 generates a plurality of pseudo words for reading the piece image according to the initial information, the neutral information, or the final information, and the ratio corresponding to each of the pseudo words. The confidence value to calculate is calculated.

The piece recognition controller 340 may convert the 19 class scores included in the initial information into probability values. For example, P (a) may be calculated as a probability value corresponding to the a-class score. The word recognition control unit 340 may perform the transformation of the probability value of the class score through Equation 1 below.

[Equation 1]

P (W _i | X) = 1 / (1 + exp (af (x) + b))

W _i : Class i classifier

-X: feature information

f (x): class score

a, b: constant

After converting the class score into a class probability through Equation 1, the piece recognition control unit 340 multiplies each initial class probability, the neutral class probability, or the final class probability by combining the classes corresponding to each probability product. Generate pseudo-seed according to.

For example, the piece recognition control unit 340 may multiply the "a class probability" by the initial class probability and the "class class probability" by the neutral class probability. In this case, in response to the probability product, the piece recognition controller 340 may generate a pseudo piece of "a". The product of each of these class probabilities may be performed for all class probabilities to produce all combinable pieces. Accordingly, a probability product obtained by multiplying each of the class probabilities may be calculated in correspondence with each generated pseudo word. The probability product may be set to a non-segmented confidence value of each pseudo shoe.

The piece recognition control unit 340 selects the first-order pseudo word having the highest value of the non-partitioned reliability and the second-order pseudo word having the next highest value of non-partitioned reliability among all the generated pseudo words. Thereafter, the piece recognition control unit 340 calculates a difference between the undivided reliability of the first-order pseudo word and the non-divided reliability of the second-order pseudo word.

As a result of the calculation, when the difference between the first rank undivided reliability and the second rank undivided reliability is greater than or equal to a predetermined threshold value, the piece recognition controller 340 recognizes the first rank pseudo piece as the piece image. The threshold value may be set by a person skilled in the art to a predetermined value by analyzing a frequency of error occurrence through a predetermined experiment.

As a result of the calculation, when the difference between the first rank undivided reliability and the second rank undivided reliability is less than the threshold value, the piece recognition controller 340 may recognize the piece image through the divided phoneme recognizer 350.

When the difference between the first-order non-partitioned reliability and the second-order non-partitioned reliability is less than the threshold value as described above, the divided phoneme recognizer 350 subdivides the single image. That is, the divided phoneme recognizer 350 divides the phoneme of the single image through the phoneme divider 351.

The phoneme dividing unit 351 sets a phoneme segmentation area through the format information of the piece image, and converts the piece image into the initial image and the neutral image according to a touching point of the piece image included in each of the divided areas. , Or a final image.

For example, as illustrated in FIG. 6, the phoneme dividing unit 351 generates a first phoneme divided image 620 that performs contour smoothing on the single image “system” 610. Thereafter, the phoneme dividing unit 351 detects a dominant point of the first phoneme-divided image to generate a second phoneme-divided image 630. The dominant point refers to a point having a high curvature, that is, a portion that is bent a lot in the single image. The phoneme dividing unit 351 may remove a neighborhood point from the second phoneme divided image, extract a nonlinear cutting path, and generate a phoneme divided image 640 of the single image. have.

After the phoneme segmentation of the single image, the phoneme feature value information unit 352 reads each feature value of the initial image, the neutral image, or the final image from which the single image is divided. The feature value information reading may be implemented in the same manner as the feature value reading operation of the feature value information unit 310 described with reference to FIG. 4.

Thereafter, the divided phoneme recognizing unit 350 selects N pseudo characters in the order of high degree of non-division reliability among all the pseudo characters generated by the character recognition control unit 340. The division phoneme recognizing unit 350 reads each of the initial class, the neutral class, or the final class each of the N pseudo words.

The segmented phoneme recognizer 350 may include the first pseudo image, the neutral image, or the final image in which the single image is divided by the initial recognizer 353, the neutral recognizer 354, and the final recognizer 355, respectively. The initial class score, the neutral class score, or the final class score which have in correspondence to are calculated.

For example, when the single image is "ga" and the N pseudo-words are "gi, ga, g, g, p, key, g, wave, ka" in undivided reliability order, the segmentation recognition is performed. The unit 350 calculates a class score for each of the 10 pseudo characters of the initial image " a " and the neutral image " '"

That is, when calculating the class score of the single image "a" for the first order pseudo word "group", the first class recognizer 353 calculates the initial class score of the initial image "a" through the "a classifier". Calculate. In addition, the neutral recognizer 352 calculates a neutral class score of "ㅏ" through the "|" classifier.

Further, when calculating each class score of the single image "a" for the 10th rank pseudo word "child", the first class recognizer 351 calculates the initial class score of the initial image "a" through the "new" classifier. In addition, the neutral recognizer 352 calculates the neutral class score of "ㅏ" through the "ㅏ classifier".

As such, when each class score of the piece image is calculated for each of the 10 pseudo pieces, the piece recognition controller 340 multiplies the initial class score, the neutral class score, or the final class score by each pseudo piece. Calculate. The product of each class score may be set as a split reliability of the pseudo pseudo image of the single image.

For example, the segmentation reliability of the first order pseudo word "group" of the single image "a" may include the initial class score value of the initial image "a" calculated through the "a classifier" and "| It may be set to the product of the neutral class score value of the neutral image "ㅏ" calculated through the "classifier".

As described above, when the split reliability of the respective pseudo pictures of the piece image is calculated, the piece recognition control unit 340 multiplies the split reliability and the non-divided reliability corresponding to each of the N pseudo pieces to calculate the pseudo picture reliability. do. In this case, the split reliability may be converted into a probability value and multiplied by the non-split reliability, as in the case of non-split reliability.

Thereafter, the piece recognition control unit 340 may recognize the pseudo piece having the highest pseudo degree reliability among the N pseudo pieces as the single image.

As described above, when it is determined that the accuracy of character recognition is not guaranteed because the difference between the non-divided reliability of the first-order pseudo stitches and the second-rank pseudo stitches is short, the character recognition apparatus according to the present invention uses a single image. It is possible to guarantee the accuracy of character recognition by recognizing characters by phoneme division and recognizing a character by referring to both the phoneme division result and the phoneme non-division result.

Further, when the non-divided phoneme recognition result indicates that the difference between the non-divided reliability of the first-order pseudo word and the second-order pseudo word is calculated to be greater than or equal to a threshold value, the first-order pseudo word is not subdivided. By recognizing the image, it is possible to obtain the effect of ensuring the speed and accuracy of character recognition at the same time.

The character recognition method according to the present invention may be implemented in the form of program instructions that can be executed by various computer means and recorded in a computer readable medium. The computer readable medium may include program instructions, data files, data structures, etc. alone or in combination. Program instructions recorded on the media may be those specially designed and constructed for the purposes of the present invention, or they may be of the kind well-known and available to those having skill in the computer software arts. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tape, optical media such as CD-ROMs, DVDs, and magnetic disks, such as floppy disks. Magneto-optical media, and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, flash memory, and the like. The medium may be a transmission medium such as an optical or metal wire, a waveguide, or the like including a carrier wave for transmitting a signal specifying a program command, a data structure, or the like. Examples of program instructions include not only machine code generated by a compiler, but also high-level language code that can be executed by a computer using an interpreter or the like. The hardware device described above may be configured to operate as one or more software modules to perform the operations of the present invention, and vice versa.

7 is an internal block diagram of a general purpose computer system that may be employed to implement the character recognition method according to the present invention.

Computer system 700 includes one or more processors 710 connected to main memory including random access memory (RAM) 720 and read only memory (ROM) 730. The processor 710 is also called a central processing unit (CPU). As is well known in the art, the ROM 730 serves to transfer data and instructions to the CPU unidirectionally, and the RAM 720 typically transfers data and instructions bidirectionally. Used to. RAM 720 and ROM 730 may include any suitable form of computer readable media. Mass storage 740 is bidirectionally coupled to processor 710 to provide additional data storage capabilities, and may be any of the computer readable recording media described above. The mass storage device 740 is used to store programs, data, and the like, and is typically an auxiliary memory device such as a hard disk which is slower than the main memory device. Certain mass storage devices such as CD ROM 760 may be used. The processor 710 may include one or more input / output interfaces, such as video monitors, trackballs, mice, keyboards, microphones, touchscreen displays, card readers, magnetic or paper tape readers, voice or handwriting readers, joysticks, or other known computer input / output devices. 750 is connected. Finally, the processor 710 may be connected to a wired or wireless communication network through the network interface 770. Through this network connection, the procedure of the method described above can be performed. The apparatus and tools described above are well known to those skilled in the computer hardware and software arts.

While specific embodiments of the present invention have been described so far, various modifications are possible without departing from the scope of the present invention.

Therefore, the scope of the present invention should not be limited to the described embodiments, but should be defined not only by the claims below, but also by the equivalents of the claims.

According to the phoneme-based character recognition apparatus and method of the present invention, an effect of shortening the execution time of a character recognition operation by recognizing the letter through a non-divisional letter recognition method that recognizes the letter as it is without dividing the letter into phonemes. Can be obtained.

Further, according to the phoneme-based character recognition apparatus and method of the present invention, if the non-divisional phoneme character recognition result is ambiguous, after recognizing the letter through a divisional phoneme recognition method for dividing and recognizing the phoneme, By recognizing the letter in comparison with the result of the subtractive recognition, it is possible to obtain the effect of accurately recognizing the character regardless of the change of the character font.

In addition, according to the phoneme-based character recognition apparatus and method of the present invention, when a phoneme is divided and recognized, a touching point which is a connection part with a vowel among the phonemes by using the type information and curvature information of the letter. By detecting the phoneme segmentation, it is possible to obtain the effect that the phoneme segmentation can be performed more accurately.

As described above, the present invention has been described by way of limited embodiments and drawings, but the present invention is not limited to the above-described embodiments, which can be variously modified and modified by those skilled in the art to which the present invention pertains. Modifications are possible. Accordingly, the spirit of the present invention should be understood only by the claims set forth below, and all equivalent or equivalent modifications thereof will belong to the scope of the present invention.

Claims (8)

In the character recognition method, Reading feature value information and format information of a predetermined single image; Generating initial information, neutral information, or final information of the single image according to the feature value information and format information; Generating a plurality of pseudo words for reading the single image according to the initial information, the neutral information, or the final information, and calculating a non-split confidence value corresponding to each of the pseudo words; If the difference between the first-order non-partitioned reliability having the highest value and the second-order non-partitioned reliability having the next higher value is greater than or equal to a predetermined threshold value, the pseudo-first-partitioned reliability Recognizing a piece as the piece image; And When the difference between the first-rank reliability and the second-rank reliability is less than the threshold value, the split image is subdivided to calculate split reliability corresponding to each of the pseudo words, and the split reliability and the non-division corresponding to the respective pseudo words are divided. Recognizing the pseudo word having the highest reliability as a result of multiplying each other by the reliability image Character recognition method comprising a. The method of claim 1, The feature value information of the single image may include vector information of pixels included in each mesh image obtained by dividing the single image into a plurality. The method of claim 1, The format information of the single image is set to any one of a Korean format consisting of a first format to a sixth format and a seventh format of an English letter, a number, or a special character. The method of claim 1, The step of generating the initial information, the neutral information, or the final information of the single image according to the feature value information and format information, First class score (Score) each having 19 Korean first class (corresponding to the initial number of the single image), Neutral class score each having 5 to 9 neutral classes (1) corresponding to the neutrality of the single image, or Calculating a final class score of each of the 27 Korean final classes corresponding to the single image finality; And Converting the initial class score, the neutral class score, or the final class score into an initial class probability, a neutral class probability, or a final class probability, respectively. Character recognition method comprising a. The method of claim 4, wherein Generating a plurality of pseudo words for reading the single image, and calculating non-segmented reliability corresponding to each of the pseudo words, Multiplying each of the initial class probabilities, the neutral class probabilities, or the final class probabilities by each other, and calculating the probability product of each class of the pseudo-shallow image according to the combination of each initial class, the neutral class, or the final class. Character recognition method comprising a. The method of claim 1, When the difference between the first-order non-partitioned reliability and the second-order non-partitioned reliability is less than the threshold value, the step of subdividing the single image to calculate the divided reliability corresponding to each pseudo-letter, Dividing the single image into an initial image, a neutral image, or a final image; Reading respective feature information of the initial image, the neutral image, or the final image; Selecting N pseudo pieces according to a high value undivided reliability order of the pseudo pieces, and reading an initial class, a neutral class, or a final class each of the N pseudo pieces; Calculating an initial class score, a neutral class score, or a final class score each of the initial image, the neutral image, or the final image corresponding to each of the N pseudowords; And Multiplying the initial class score, the neutral class score, or the final class score by each of the N pseudo pieces to calculate a segmentation reliability of the single image corresponding to each of the N pseudo pieces Character recognition method comprising a. The method of claim 6, The step of dividing the single image into an initial image, a neutral image, or a final image may include setting a division region through format information of the single image, and touching points of the single image included in the respective division regions. And dividing the image into the initial image, the neutral image, or the final image. A computer-readable recording medium having recorded thereon a program for executing the method of any one of claims 1 to 7.

Images