KR20090111202A - The Optical Character Recognition method and device by the numbers of horizon, vertical and slant lines which is the element of Hanguel - Google Patents

Abstract

PURPOSE: An optical character recognition method and a device by the numbers of horizon, vertical and slant lines which is the element of Hanguel is provided to make the apparatus which changes the captured document into TTS conversion audio for visually-handicapped person. CONSTITUTION: An optical character recognition method and a device by the numbers of horizon, vertical and slant lines which is the element of Hanguel forms a character image of one letter. Each characteristic is recorded for horizontal line, vertical line, and slant lines are extracted from the target character image. The unique code value which indicates the letter encodes each number. It is recognized as the target character after comparing the line number with the recorded character database.

Description

      The Optical Character Recognition method and device by the numbers of horizon, vertical and slant lines which is the element of Hanguel}

Recognition of Hangul is very difficult compared to English and Chinese characters, and the current technology does not continue research in related fields compared to various fonts, so it can only recognize Myeongjo and Gothic fonts. Various approaches such as pre-processing and post-processing to recognize fonts such as signs are being conducted.

In other words, since Perceptcom, a Korean document recognition SW-related company, launched the business card reader since 10 years ago, the market has not been formed on a large scale, so investment has not been made in recent years. In addition, the research on the recognition of characters in the surrounding environment, rather than documents, has been done through the character region verification technique and the text region extraction technique from the camera image.

OCR (Optical Character Recognition) technology for two-dimensional printing Korean documents has been increasingly applied to mobile environments with the development of mobile phone technology, but the recognition technology of printed documents has been continuously studied for the last 40 years. It is also true that it is still not as expected by the general public.

The present invention has to improve the recognition rate and recognition speed in order to incorporate the document recognition function of the Hangul into the mobile environment, the Hangul has a large number of 2350 characters that can be generated and a relatively small number of basic characters complicate one character Because it is formed in the form of, for example, the finality, there is a limit in recognition of the similarity of characters.

While many existing recognition methodologies that use the structural characteristics of Hangeul define the basic elements of Hangeul as strokes, the present invention uses Hangeul, which is composed of highly directional elements such as vertical lines, horizontal lines, and diagonal lines. It was provided to distinguish rather than to recognize.

That is, the present invention has been made in that it is necessary to develop a new character recognition algorithm in order to improve the character recognition rate and recognition speed in a mobile environment such as a mobile phone.

In particular, recently equipped with high-performance camera sensor of more than 3M pixels (English recognizable resolution), it is possible to apply the OCR technology of 3D camera acquisition document and background. The present invention has been made in that it can be applied as a real-time reading aid.

It is necessary to develop new algorithms based on the appropriate combination and selection of reliability / recognition rate / speed according to the limitations and demands of the visually impaired. In addition to the level of static recognition used, it is also necessary to develop a technique for correcting mobile-specific shaking, light, and three-dimensional tilt of a recognition object.

Recently, as the camera sensor and CPU performance have been partially improved, research on mobile OCR has been conducted beyond simple business card recognition. However, it is still approaching in a limited form due to problems such as camera sensor resolution and recognition speed of mobile phone CPU. It is true.

The code reader for a special two-dimensional bar code symbol called VoiceA, which is proposed for the visually impaired, stores two pages of information in a voice eye symbol of 1.5cm in width and length, and displays the information of the publication printed with the original. If you create a voice eye symbol for all print and publication documents without having to create a hardware, software configuration, and recording book designed for self-facing, you can use the device to output the voice, but it also prints the voice eye symbol. The existing documents are not universal and there are restrictions on their use.

Up to now, for general character recognition of Korean characters, the preprocessing process extracts the margins of characters and preprocesses image processing to form the input character image, and the thinning algorithm transformation technology is applied through histogram and binarization to extract the contour. Perform horizontal and vertical projection transformations, extract the minimum neighboring polygons and refine the background, perform boundary analysis extraction and diagonal-diagonal region projection transformation, and use eight-way component distribution feature extraction quantization to obtain endpoints, branch points, intersection points, It extracts structural feature points such as bend points, and then performs smoothing sampling to form a template feature point database, which is a complex process of extracting recognition characters through the character feature database and the circular and elastic joint algorithms.

Many products have been introduced using the Hangul character recognition technology, and the product called 'Mobile Reader' extracts text information by applying various algorithms to image processing and automatic recognition information operating in limited environments such as mobile phones and PDAs equipped with cameras. To judge. The commercialized product is currently installed in some mobile phones with business card recognition and electronic character recognition function.Diotec's mobile OCR is a document recognition solution that recognizes and reads printed document images and converts them into text form. It can be applied to a terminal equipped with a camera of 1M pixels or higher that supports autofocus.In addition to the advent of high resolution camera phones, its utilization is increasing. It is not evaluated as a necessary product to use.

Perceptcom's document recognition software 'Armi' is the first commercialized document recognition product in Korea. It reads the existing printed data in the form of an image through a scanner and analyzes the data and divides the data into picture and text areas. It was recognized as a possible recognition software and was used as a bundled product by most of the domestic scanner companies, and the card reader 'Armi' was also recognized as the best product in the domestic character recognition engine because it was proved to handle card recognition and management perfectly within a limited range. It is also used as a base solution by other companies.

In the conventional method, the character recognition process uses the labeling to extract the character, normalizes it, removes the noise through the morphology algorithm, refines it with the Zhang-suen algorithm, and extracts the feature points through the corner detection. Finding is done.

The present invention has been proposed to solve the above problems, and the number of characters that can be generated in Korean is 2,350 characters in KSC5601 code, unlike English, Japanese, and the like. Because of the large and relatively small number of basic phonemes that form a letter, the similarity of letters (eg, 'ㅁ' 'ㅂ' 'ッ', used as a consonant and a trailing consonant in the consonants, is large and is particularly recognized). In the case of complex bearings, there are many difficulties with the existing algorithms.

Therefore, existing character recognition algorithms using the structural characteristics of Hangul tend to rely heavily on postprocessing such as preprocessing and word dictionary search.

Hangul characters are composed of basic phonemes, and basic phonemes are composed by connecting basic elements with directions. Therefore, the efficient use of access information of these elements in Hangeul recognition is the key to recognition. In the paper, 'Korean Recognition by Hierarchical Stroke Extraction', published in 1990, the information of elements is hierarchically organized to make it more efficient. We proposed a method that can be used. After normalizing the input characters, it extracts the direction components by emphasizing the straight line segments in four directions (vertical direction, horizontal direction, diagonal direction and inverse direction direction), The length of the direction line is stored as information at the start position of the direction line, and the target area is divided into sub-regions of the same size, and the largest length information in the sub-area is determined as a representative value. Considering the four-way component for, determine the form of linkage of the line elements in the area. A feature representing the connection form of the line element of the proposed method has also been recognized that the extraction Hangul.

Unlike the method proposed in the present invention, vertical lines, horizontal lines, diagonal lines, and circles are divided into three, horizontally and vertically, and divided into single, middle, and long lines according to overlapping lengths. Most of the research has been conducted in that it is possible to distinguish Korean codes by applying the complementary recognition method to the characters with the same code.

 In Hangeul, most consonants are composed of vertical line and horizontal line, but ㅅ, ㅇ, ㅈ have diagonal line and original line. Hangul phonemes are composed of straight line elements in four directions, and consonants are more complicated than vowels.

The concatenation of the elements in the consonant is what form of connection of the elements does play an important role in distinguishing the elements. In vowels, there is a relatively simple connection between vertical and horizontal lines, and the connection location, as well as the connection form, has important information.

In other words, consonants and vowels are composed of lines, which are differentiated into directions (horizontal, vertical, oblique, historical, and circle) and length (long, middle, and short). It is very fast compared to extracting complex shapes or locations.

In other words, if you code the number of stroke direction and length in one step, it is recognized immediately if each character has its own code value (36% of the total characters), and if it has overlapping code values, the position direction of the vowel (for example If you apply the algorithm to easily discriminate between ㅏ and ㅓ, ㅗ and TT, it can distinguish up to 64% of the entire Hangul characters, and the remainder consonants (for example, if analyzing duplicate cases, ㅁ, ㅂ, プ, division and 'cloth') And 'sing', such as 'Song'), make 100% Korean characters distinguish.

That is, in the case of consonants, most of the consonants are composed of vertical lines and horizontal lines, such as ㄱ ㄴ ㄹ ㅁ ㅂ ㅌ ㅅ, ㅅ ㅅ ㅋ ㅋ including the diagonal and history lines and ㅇ ㅎ circle.

That is, a and b are one horizontal and one vertical (easily divided into horizontal positions), and c is 2 horizontal and 1 vertical, ㅌ is 3 horizontal and 1 vertical, and d is 3 and 1 vertical.

ㅁ ㅂ ッ is all horizontal and vertical 2, so to distinguish it, ㅁ if the starting point on the top and the left side of the horizontal line correspond to the left vertical line ㅁ, and 있으면 if the horizontal starting point is below, and if the horizontal line is above the vertical line. .

As mentioned above, the consonants and vowels of Hangeul are basically composed of vertical, horizontal, diagonal, and circle by the principle of Hunminjeongeum, and if the number and size of these elements are coded together, an independent pattern is created.

The calculations confirmed that 64% is distinguished when the code is inserted, as in the 2,350 Korean alphabets according to the KS standard.

The advantage of this method is that 'Bob', 'Pop', 'Madam', 'Mam', 'Night', and 'Palm' represent the most identical codes, depending on the specific code, for example to determine if the consonant is 'ㅁ'. If you are one of the two letters 'mom', thank you for the finality 'ㅁ' If the 'mom', or 'thank' most of the calculation is terminated in the maximum three steps.

In addition, the post-processing algorithm can be corrected by checking a word.

Recently, as digital cameras, mobile phones, and PDA cameras have become commonplace, image acquisition from portable devices has been simplified.

If the Korean characters included in the acquired image are quickly and accurately extracted and recognized, they are useful for real life such as automatic entry of bank slips, delivery slips, order forms, applications, etc. Can be utilized.

However, the existing algorithms are affected by the tilt, roughness, and camera characteristics of documents and undergo complicated transformations, extractions, and quantizations. Therefore, they are difficult to apply to PDAs, especially mobile phones, which have a long calculation time and are less powerful than computers. In the early days, the recognition rate was lowered, and some mobile phones were initially introduced for business card recognition.

In order to overcome this problem, the present invention has been made insensitive to a document environment, and when applied to a PDA, a reading aid for the visually impaired person can capture and recognize a document and convert it into a TTS.

These reading aids can be used by the elderly and dyslexia patients, as well as visually impaired, to solve the inconvenience of life as well as to escape from the alienation of information caused by the disability.

Most of the text extraction studies on images have been conducted on document images inputted by scanners or moving image images containing subtitles. However, the recent widespread use of digital cameras has led to the extraction of text areas from images with various lighting and complex backgrounds. As research is expanding, the present invention is expected to provide an opportunity for further activation of such research.

The above objects, features and advantages will become more apparent from the following detailed description taken in conjunction with the accompanying drawings. Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings.

      1 is a flow chart of the Hangul character recognition algorithm and recognition process that is widely used at present. Apply horizontal and vertical projection transformations through contour extraction, perform boundary analysis extraction and diagonal-diagonal region projection transformation through minimum neighboring polygon extraction and background thinning, and extract 8-way component distribution feature quantization. Through the process of extracting structural feature points such as end points, branch points, intersection points, and bend points, smoothing and sampling, forming a template feature point database, and extracting the recognition characters through the character feature database and circular joint and elastic joint algorithm. In a way that does It is sex.

      Thinning is to peel off the surface of an object little by little and finally have a thickness of 1, and in the early stages of document recognition, many researchers indicated that an object is represented by a skeletal line to compress the amount of information as much as possible and then it was studied. First, the skeletal line width should be 1, second, the location of the skeletal line should be located at the center of the line shape. Third, the skeletal line should maintain the connectivity in the original figure. Fifth, the shape of the miscellaneous lines due to the small irregularities of the pattern outline should not be added to the skeletal line.

      Thinning can be basically divided into sequential processing and parallel processing. In the sequential processing method, the current processing result is influenced not only by the past processing value but also by the future processing result. In this method, it is possible to process only the target image.

      The parallel processing method requires a buffer equal to the size of the target image because the current processing result is not influenced or affected at all in the past or in the future, and buffers the processing result for the current pixel of interest. In this manner, after the one-time processing for the entire image is finished, the processing result stored in the buffer is overwritten on the original target image.

      Zhang Suen's thinning algorithm, which is widely used, was first published in 1984. It is one of the parallel processing methods. The basic algorithm is as follows.

      The black pixel I (i, j) to be processed is deleted if the following conditions are satisfied.

      ① for pixel I (i, j) the connectivity of pixels around it is 1,

      ② For pixel I (i, j), there should be at least 2 to 6 black pixels among the pixels around

      At least one of the pixels I (i, j + 1), I (i-1, j), I (i, j-1) must be a background pixel, i.

      At least one of the pixels I (i-1, j), I (i + 1, j), I (i, j-1) must be a background pixel, i.

      When the first subiteration is satisfied, the next condition is applied again.

      ① for pixel I (i, j) the connectivity of pixels around it is 1,

      ② For pixel I (i, j), there should be at least 2 to 6 black pixels among the pixels around

      At least one of pixels I (i-1, j), I (i, j + 1), and I (i + 1, j) must be a background pixel, or 0,

      At least one of the pixels I (i, j + 1), I (i + 1, j), I (i, j-1) must be a background pixel, i.

      Continue with the two subiterations above until there are no more pixels left to erase.

      2 is a step-by-step flowchart for recognizing Hangul characters according to the principle of Hunminjeongeum. Divide 32x32 or 62x62 dots, etc. into 3 parts and extract strokes for 4 directions (vertical, horizontal, diagonal, and circle), and the length of these elements (long, medium, and short) is 1 cell out of 9 cells. If you enter, however, if you enter into two cells, if you enter into three cells, the information is classified and classified, and 385 out of 1076 characters have a unique code.

      Through this, the number of single + medium + long (horizontal line) + short + medium + small (vertical line) + short + middle + long (diagonal) + circle type will create a one-step code value for each letter.

      In this case, no distinction is made about vowels. (Those ㅓ and 코드, ㅗ and TT, ㅑ and ㅕ, ㅛ and ㅠ are recognized the same.) However, since 385 characters represent unique code values, they are classified by simply calculating the number of four-way line elements.

      By calculating 4-way line segment, the number of horizontal lines is masked from the top row, and if the horizontal lines of more than the minimum number of dots (more than 4 dots in the case of 32x32 dots) are checked, the line number and the number of dots are recorded. When it comes down and checks, if it is almost the same size (1 ~ 2 dot difference occurs in Myungjo), it is judged as the same horizontal line (if not thinned).

      If horizontal lines appear more than 4 dots apart on the same line, this is considered a separate horizontal line. Horizontal lines that have been counted during horizontal check are removed from the image. Therefore, the image after the horizontal search has no horizontal section and only vertical lines, diagonal lines and circles remain.

      The number of vertical lines in the next process is masked from the leftmost column, and if the vertical lines of more than the minimum number of dots (more than 4 dots for 32x32 dots) are checked, record the column number and the number of dots, and move them one more column to the right. If it is almost the same size (1 ~ 2 dot difference occurs in case of light body, if not thinned), it is judged as the same vertical line.

      If vertical lines appear more than 4 dots apart in the same column, this is considered a separate vertical line. Vertical lines that have been counted during vertical line check are deleted from the image. Therefore, after the vertical line search, the vertical line section disappears and only diagonal lines and circles remain.

      As such, in the Hangeul character image of 32x32 dots, there are also line elements having a line width of 3 pixels or more, especially in the case of Myungjo, and in this case, it can be determined that two or more line elements exist in one line element.

      However, since it is actually a single ship, it is necessary to make the above judgment to compensate for this error. The number of diagonal lines is masked at intervals of 45 degrees from the left column of the top row, and if the diagonal lines of more than the minimum number of dots (more than 4 dots in the case of 32x32 dots) are checked, the line and column numbers and the number of dots are written diagonally. If one more dot comes down and checked, it is judged to be the same diagonal line if the size is almost the same (1 ~ 2 dot difference occurs in case of Myungjo body, if not diagonalized). If diagonal lines appearing more than four dots apart on the same line, this is considered a separate diagonal line.

      The diagonal lines that have been counted during the oblique check are deleted from the image. Therefore, after the diagonal search is completed, the diagonal section disappears and only the reverse diagonal line and circle remain. The number of history lines is masked at 45-degree intervals from the right column of the top row, and when the history lines more than the minimum number of dots (more than 4 dots for 32x32 dots) are checked, the line, column number, and number of dots are recorded. If you check one dot diagonally down and check it, if it is almost the same size (1 ~ 2 dot difference in case of Myungjo), it will be judged as the same history line.

      When history lines appear more than 4 dots apart on the same line, it is considered a separate history line. History lines that have been counted during the history line check are removed from the image. Therefore, the image that finishes searching the history line is missing the history line section and only the circle remains. We need a way to distinguish horizontal, vertical, oblique, and historical lines from circles when searching.

      In this sense, if the thinning is performed by changing the line width to the width of one dot before searching the 4-way line in the letter image, it is easy to distinguish the circle when discriminating horizontal, vertical, and diagonal lines.

      In other words, when the horizontal line is changed while the horizontal line is being searched (in the case of a wide circle instead of a round circle), the distance between the connecting dots increases and decreases, and the center and the radius (vertical and horizontal) are recorded.

      In this step, judging the position of the vowel (in the first step, 인식 and ㅓ, ㅗ and TT, ㅑ and ㅕ, ㅛ and ㅠ are recognized the same) and judge it as direction + end + middle + long (horizontal line) + direction + end If you sort by sorting + medium + small (vertical line) + short + medium + long (diagonal) + circle, 692 out of 1076 characters have unique code values and are automatically classified.

      That is, in the first step, when the code is a duplicate character (for example, '아' and '어'), the horizontal line and the vertical line forming the vowel are distinguished and discriminated. Even after this step, if the Korean character types with duplicate code values are distinguished, the length, middle, and end of the horizontal and vertical lines of a, b (horizontal 1, vertical 1) and ㅁ, ㅂ, ㄴ (horizontal 2, vertical 2) In most cases, the highest number of duplicates appears as four.

      In three steps, a, b (horizontal 1, vertical 1) and ㅁ, ㅂ, プ (horizontal 2, vertical 2) are distinguished, and only the initial and final positions are determined, and the direction + end + middle + long (horizontal line) + direction + When the code value is defined as single + middle + small (vertical line) + single + middle + long (diagonal) + one + (a b) + (ㅁ ㅁ ッ division), 1,066 of 1,076 characters are recognized. Is done.

      The ten letters that are not categorized are 'Dewa and Kitchen', 'Doo and Ku', 'Dowa Ko', 'Song and Clothes', and '뤽 and Rest'. 'Song and clothes' and '륏 and rest' can be divided into whether the position of the diagonal line is above or below, and 'Dowako', 'Duwawaku', and 'Deuwa Kitchen' are the vertical lines that make up consonants. (C and ㅋ are the vertical lines on the left and ㅋ on the right). This makes 100% classification.

      3 is a part of the Korean character classification table coded by the Hunminjeongeum Changje principle to which the present invention is applied, and in this format, a classification table database according to each Korean font is constructed, and 99% of Korean characters are easily distinguished through a three-step comparison process. Finally, the 10 characters are divided with the position of the oblique line and the position of the vertical line.

      The present invention described above is capable of various substitutions, modifications, and changes without departing from the technical spirit of the present invention for those skilled in the art to which the present invention pertains. It is not limited by the drawings.

1 is a flow chart of a conventional Hangul character recognition algorithm and recognition process

      2 is a step-by-step flowchart for the recognition of Hangul characters by the Hunminjeongeum Changje principle to which the present invention is applied

      Figure 3 is an example Hangul character classification coded by the Hunminjeongeum Changje principle to which the present invention is applied

      * Explanation of symbols for the main parts of the drawings

       1: Character line, stroke separation 2: Vowel connection location direction division

3: initial, final

Claims (9)

      In the method of recognizing a character,       Form a letter image of a letter,      The horizontal lines, vertical lines, and diagonal lines are extracted from the corresponding text image, and each characteristic is recorded.       Each number is coded to form a unique code value that represents the letter A method of recognizing characters by counting the number of lines, which is characterized by finding and comparing the values calculated based on the number of line elements in the recorded character database and matching them. According to claim 1       After forming a letter image of a letter A method of recognizing a character by calculating the number of line elements characterized by proceeding after skeletonizing the image of the character through a thinning process Compare the number of horizontal lines, vertical lines, and diagonal lines in one step as shown in paragraph 1,       If you have duplicate code values       By constructing additional components by coding the direction components of the vertical and horizontal lines representing the vowel A method of recognizing a character by calculating the number and direction of a line element characterized by recognizing and comparing the value calculated based on the number of line elements in the recorded character database and matching and matching. According to claim 3,       You can also code the direction components of the vertical and horizontal lines that represent the vowel,       If you see duplicate code values,       You can distinguish between the base and the       Additional components are coded to distinguish the taste from the rice cakes. A method of recognizing a character by distinguishing the number of consonants and the consonants, which are based on the number of the elements, and compares and compares the calculated value in the recorded character database. The method of claim 4,       You can also code the direction components of the vertical and horizontal lines that represent the vowel,       If you see duplicate code values,       To distinguish between dishes and keyboards,       The component that separates the position of the oblique line is additionally coded. A method of recognizing a character by distinguishing the number of consonants and the consonants, which are based on the number of the elements, and compares and compares the calculated value in the recorded character database. The method of claim 5,       To distinguish between Dijon and Kiss      Dividing and teasing method characterized by dividing by the position of the vertical line forming the consonant The method of claim 4,       To distinguish between the base and the      How to distinguish the air force and knee is characterized by the location of the horizontal line constituting the consonant The method of claim 4,       To distinguish between the heart and the mouth and blood      How to distinguish between Myeonge and Byeongeup and Peeup, characterized by the horizontal line and the vertical line of the consonant (horizontal position with respect to the left vertical line; According to claim 1,       To distinguish between       When moving rows and columns       Basic target is composed of 1 or more dots (2 ~ 3 dot width, change depending on the resolution of the text image. If the dot width is set to 2 dots for 32x32 dots, it moves to the 16-step area and checks for the presence of line elements. )       Characterized in that it is configured to absorb the noise portion of the straight line generated in the thinning process Line Segment Inspection Method in an Area of 1 Dot or More in a Character Image

Images