KR950001059B1 - Korean character address recognition method and apparatus - Google Patents

Abstract

The character recognition method and unit recognize hand-written Korean character adddress at high recognition speed and rate by comparing the input character and standard character using administrative district name dictionary constructed hierarchically. The recognition method comprises extracting the candidate character by using the district name dictionary; recognizing the character; processing the next part of the address based on the recognized district name. The recognition rate is above 99 percent.

Description

Korean address recognition method and device

1 is a structural diagram of the administrative region of Korea.

2 is a block diagram of an administrative district name dictionary.

3 is a method of extracting candidate characters from an administrative district name dictionary in the process of processing administrative district names corresponding to Tier 1 (Gyeonggi, Juksi, Do), Tier 2 (Old, County, and City), and Tier 3 (Exceptional District). Flowchart showing.

4 is a flowchart illustrating a method of extracting candidate characters from an administrative district name dictionary while processing administrative district names corresponding to hierarchy 4 (eup, myeon, dong) and hierarchy 5 (ri, dong).

5 is a flowchart showing a post-processing method of layer 1. FIG.

6 is a flowchart showing a post-processing method of Layer 2.

7 is a flowchart showing a post-processing method of layer 3.

8 is a flowchart showing a post-processing method of layer 4. FIG.

9 is a flowchart showing a post-processing method of layer 5. FIG.

10 is a block diagram of a system according to the present invention.

* Explanation of symbols for main parts of the drawings

1 scanner or tablet 2 address divider

3: candidate character extraction unit 4: administrative name dictionary

5: character recognition unit 6: post-processing unit

7: address output unit

[Industrial use]

The present invention relates to a method and apparatus for recognizing a Hangul address, and in particular, in the case of online or offline handwriting Hangul address recognition, a considerable confusion and error occurs in the character recognition step due to various handwriting types according to the handwriting. The present invention relates to a method of significantly reducing the number of comparisons in recognition, and also to a post-processing method for efficiently correcting misrecognition results resulting from recognition confusion and errors in the character recognizer. Here, online means that the user inputs data in real time using a tablet and a stylus pen. Offline means that a user inputs a medium such as paper, which is written by the user, through a scanner. it means. The Hangul address recognition method and apparatus of the present invention can achieve high-speed character recognition, as well as more accurate address recognition by applying post-processing for each administrative area based on the high recognition rate of the character recognizer. Of course, the present invention is not only limited to handwriting address recognition but also applies to the recognition of addresses described by typing, printing, and other mechanical media, but is particularly excellent in the case of handwriting address recognition involving many difficulties in recognition. will be.

[Traditional Technology and Problems]

Looking at the address recognition method researched in the prior art, as far as the present inventors have investigated, there are no related studies in Korea, and in the case of foreign countries, three research results on address recognition were published in Japan. As one, Izaki et al., Page 97-202 of the paper presented at the International Conference on Text Processing with a Large Character Set, held in October 1983. Post-processing of recognition "paper. In this paper, the character recognizer extracts the features for each character of the input address, compares them with the characteristics of each standard character, and outputs the smallest difference as candidate characters. This paper introduces a method to find the most similar candidate address by weighting the dictionary with the administrative name of the dictionary. In addition, in the publication addressed at the International Conference on Computer Processing of Chinese and Orinetal Languager in April 1990, Suzuki et al. This paper examines the accuracy of recognition by considering the difference between the values obtained by the letters of candidate 1st priority and the values obtained by the letters of candidate 1st and 2nd candidates in the process of outputting the candidate characters of the character recognizer. The letters in error give several candidate letters, suggesting a way to match these candidate letters with the names of administrative divisions in the dictionary.

In addition, the post-processing method that applies an error rule table and a connection matrix is introduced in address processing other than the administrative area name for which the administrative area name dictionary is not available. And correcting errors in handwriting address recognition by Marukawa et al. On pages 916-924 of the paper presented at the International Conference on Document Analysis and Recognition held in September 1991. Algorithm "introduces a method of segmenting administrative districts by introducing automata and processing character matching with dictionary of administrative district names for the processing of continuously written addresses.

As described in the above paper, the conventional address recognition method relates to a method of post-processing using the candidate characters when the character recognizer compares each character of the input address with all standard characters and outputs some candidate characters.

This method is inefficient because it involves a very wide and unnecessary comparison given that the recognition target is an address. When applying the conventional method for handwritten Hangul address recognition, the current KS C 5601 complete code system has as many as 2,350 character sets, so the number of comparisons is 2,350 times. In the case of Chinese characters, there are much more character sets, and the confusion range of character recognition is much larger.

In order to improve such inefficiency, in the present invention, only characters that can appear in the address, especially those that can appear in each position of the characters constituting the address, are used for comparison by using the administrative area name dictionary.

Another difference from the method presented in the above paper of the present invention is that after the character recognition for the address is completely completed, the post processing is not performed after the character recognition for the address is completed. It determines the name of the administrative district (hereinafter referred to as the candidate administrative district name) and processes the next hierarchy based on this candidate administrative district name.

[Purpose of invention]

An object of the present invention relates to a method and apparatus for recognizing handwritten Hangul addresses, in particular, that a character recognizer reduces a standard character to be compared using an administrative area name dictionary when comparing a character entered with a standard character. In addition, the present invention provides a method and apparatus for recognizing a Korean address using practical post-processing methods as well as improving recognition rate.

[details]

The present invention will be described in detail with reference to the accompanying drawings.

Before looking at the format of the administrative area name dictionary used in the present invention, it is understood that the administrative area of Korea has a hierarchical structure as shown in FIG. Tier 1 consists of Seoul Metropolitan City, Direct City, and Road, and Tier 2 consists of City (Geo, Direct City), City, and County (For Province), and Tier 4 corresponds to Eup, Myeon, and Dong, and Tier 5 is Terminal Administration. It is composed of two areas, Lee and Dong. As an exceptional case, there are several cities where the lower tier of Tier 2 is classified as wards (tier 2, exempt districts), not Eup, Myeon, or Dong (tier 4). Currently, there are five regions in Bucheon, Gyeongsangnam-do, and Masan, Gyeongsangnam-do. This is the case here. Thus, as shown in Figure 1, the administrative districts of Korea can be divided into five tiers.

Based on this structure, the administrative district name dictionary consists of five dictionaries as shown in FIG. Administrative District Name Dictionary 1 refers to the City, Direct City, and Province, Administrative District Name Dictionary 2 refers to City, County, and City, and Administrative District Name Dictionary 3 refers to the exception. In administrative district name dictionary 4, information about town, town, and town is recorded, and administrative district name dictionary 5 contains li, town. The dictionary name 4 used in the present invention includes not only towns and villages but also buildings, which is one post that corresponds to the level 4 for public institutions and relatively large buildings in order to facilitate mail separation in Korea. This is because the buildings are treated with arbitrary administrative names by assigning numbers. Therefore, in the present invention, not only the building having a postal code in the administrative area name dictionary for processing the building name, but also the name of the administrative area so that the user can add the building name without the postal code to the administrative area name dictionary according to the purpose of use. Reveal that a dictionary is constructed.

As such, being able to input a building name in an administrative area dictionary provides a very convenient method in consideration of the difference in the method of writing an address for each person. Furthermore, in light of the fact that the recent trend of housing is becoming apartment complexes, such apartments are also listed in the name of the administrative district in advance, so that convenience or maximum use can be achieved. For apartments, it may be possible to create a separate dictionary if the entry into the Administrative District Name Dictionary 4 becomes complicated in terms of the vastness of the data or the management of the Administrative District Name Dictionary. As such, when a separate dictionary is constructed for an apartment or a building without a postal code, the dictionary may be referred to as a special phrase. It can be seen that the treatment method of the special phrase can also be easily implemented by the treatment process of the present invention as the following description proceeds.

The address entered from the scriber is hierarchically listed with administrative district names as in the following example.

Hwangtan-ri, Gangnae-myeon, Cheongwon-gun, Chungcheongbuk-do

As mentioned above, it can be seen that the name of the administrative district that appears first in the construction of one address is for the tier 1 which represents the city, city, state or province. In addition, it can be seen that the administrative district names below are sequentially formed with the first layer as the starting point. Therefore, the present invention assumes that the address of the recognition target is written in the above format. In the following description, the individual parts (that is, any one of the layers 1 to 5) constituting one global address (that is, an address consisting of layers 1 to 5) may be referred to as "address words". will be. When layer 1 is called the uppermost layer, the address recognition process proceeds from the upper layer to the lower layer. Therefore, address recognition starts from the head address word. In the present invention, candidate characters that may appear at the position of the address word to be recognized are extracted through the character extracting unit before starting the recognition. That is, candidate characters that can be generated at the first position of the first address word are extracted using the administrative region name dictionary 1 containing the administrative region names for hierarchy 1. It should be noted that this is different from the comparison with the entire standard character as in the conventional method.

FIG. 3 is a flowchart illustrating a method of extracting candidate characters from an administrative area name dictionary in processing administrative area names corresponding to hierarchies 1 to 3, wherein each character forming one address word using the administrative area name dictionary is shown. Extract possible characters at the position of. First, the administrative zone name dictionary containing the administrative zone names corresponding to the address words currently being processed (which are the address words corresponding to the layer 1, the layer 2 or the layer 3) is accessed (step 10). Next, the number of characters constituting the address word (the address word forming the layer 1 or the layer 2 or the layer 3) to be processed is counted (step 12). This number is the number of characters to be recognized as shown in step 12, which is used to determine how many digits to extract for each administrative region name to be searched in the dictionary when extracting candidate characters. That is, when describing the above-mentioned address, the characters that may appear in each place of "Chungcheongbuk-do" are 15 such as "Gang (Gangwon-do), Gyeong (Gyeonggi-do, Gyeongsangbuk-do, Gyeongsangnam-do),…, Je (Jeju-do)". The number of letters appearing in the first digit of the name of the administrative district that represents the current city, district, and province is the number of letters) and the letters that may appear in the "Chung" are "Won (Gangwon-do), Sang (Gyeongsangnam-do, Gyeongsangbuk-do) ,. ..., Ju (Gwangju, Jeju Island) ".

The candidate characters that can appear in each of the four digits of the address word "Chungcheongbuk-do" are extracted by using the administrative district name dictionary. As can be seen from the above, the character recognizer does not compare the current KS C 5601 complete code system with 2,350 standard characters in recognition of the first character, "Cong," which is the address word of layer 1, but by comparing only 15 standard characters. It can be seen that this is reduced to about 15 / 2,350. In steps 14 and 16, the DO loop is repeated to extract candidate characters from the administrative district name dictionary, and the set of candidate characters thus obtained is stored in an array A [j] in a form (step 18). Next, if the array (A) containing the candidate characters is passed to the character recognition unit (5 in FIG. 10) (step 20), the character recognizer extracts from each character and the candidate character extracting unit forming the input address word. The most similar characters are output by rank in comparison with the characters (step 22). In this case, in order to avoid confusion in representation with candidate characters extracted by the candidate character extracting unit, the term "character candidates by rank" will be referred to.

Table 1 shows an example of candidate characters by rank. According to the example of Table 1, the recognition result of the first order of the character recognizer for the input character "chong" is that "chong" was selected as the most similar, and then "kyeong" to the second rank, thus up to n ranks. Consider ranking candidate characters. As for the second input character "blue", it can be seen that "before" is recognized as the most similar in the first rank, and the candidate letters for each rank are output like the recognition of "chong" for the next rank.

TABLE 1

Ranked candidate characters for any address word recognized by the character recognizer

The result recognized as the first priority with respect to the address word “Chungcheongbuk-do” entered in Table 1 above is “charging specialty”. Since this is not a result desired by the user, the candidate character up to n rank output by the character recognizer is used to enter post-processing. A flowchart showing the post-processing method for the administrative district name of layer 1 is shown in FIG. Post-processing matches the candidate characters shown in Table 1 with the administrative district names in the administrative district name dictionary. The postprocessing of the address word at the head of the address entered, namely "Chungcheongbuk-do", is applied to layer 1 As it is applicable, letter matching is done using Administrative District Name Dictionary1.

Referring to the process of FIG. 5, in step 60, characters that may appear at the position of each letter of the address word (layer 1) to be processed are read from the administrative region name dictionary and passed to the character recognizer (3 in FIG. 10). In the next step 962), the candidate character for each rank is received up to n ranks from the character recognizer that performs character recognition based on the candidate characters for each position extracted by the candidate character extractor. Here, the question of how to extract the candidate characters for each rank and how to receive them is determined in consideration of the cumulative recognition rate of the character recognizer. In the case of the character recognizer used in the present invention, the cumulative expression rate of 98% or more was shown up to 5th rank. This means that the character recognizer outputs the character that the user wants within 5 ranks of the character recognizer. As described above, the determination of n value is determined by the performance of the character recognizer. When the performance is low, the determination of n value is determined by the performance of the character recognizer. By applying it, the low performance of the character recognizer can be overcome. In step 64, characters are matched with the cities, cities, provinces, and provinces listed in the dictionary of administrative district names.

Here, only the name of the administrative district with the same number as the input characters is used for matching. The reason for this is that people write their address in Seoul, Seoul, Seoul, Chungcheongbuk-do, and Chungbuk. This is because it is inefficient to match letters based on the official administrative name (Seoul, Chungcheongbuk-do, etc.) in recognizing addresses written in various forms. In order to solve such a problem, it can be easily solved by tying various types of descriptions of administrative district names of the first hierarchy and reflecting them when constructing the administrative district name dictionary 1. Subsequently, in step 66, as a result of the matching process, the names of candidate administrative districts up to k ranks selected from the administrative district name dictionary and the differences obtained from the matching process are remembered. This process is to introduce the concept of backtracking. This concept will be discussed later as necessary. And the question of how to consider the k value used here, that is, the candidate administrative district name, is to match the administrative district name with the administrative zone name in the dictionary based on the candidate characters by rank that the character recognizer outputs. It will be decided according to the rank among the selected candidate administrative district names. As described above, the determination of the k value is determined according to the performance of the character recognizer. When the performance is low, the k value is increased. In the present invention, k is set to 3, so that candidate administrative district names up to 3 ranks are considered. By the above process, the post-processing of the first layer is completed.

If the candidate administrative district name has been selected by recognition and post-processing of layer 1, the next address word is processed based on the candidate administrative district name. That is, if "Chungcheongbuk-do" is selected as the first priority administrative district name for Tier 1, the next address word is considered to be the administrative district of Tier 2 belonging to "Chungcheongbuk-do," and the sub-layer of "Chungcheongbuk-do" in the administrative zone name dictionary 2 Approach the place where the relevant administrative district names are located This approach can be solved by pre-composing the administrative district name with a pointer that connects the upper hierarchy and its lower hierarchy to the administrative district name preconfiguration. Hierarchy means "Cheongju", "Chungju", "Cheongwon-gun", etc.

The processing of the address word corresponding to layer 2 is the same as the method of processing the address word of layer 1. After extracting candidate characters that can appear in place of each character constituting the address word by the method described in FIG. 3 using the administrative district name dictionary 2, recognition and post-processing are performed based on the extracted characters. As a result, if the candidate administrative district name selected as "Priority" is "Pleasant Army", the next address word is considered to be Tier 4, and accesses to Tier 4 by accessing the location containing the sub-administrative district names of "Pleasant Army" in the Administrative Territory Dictionary. Process the corresponding address word.

If the result of applying recognition and post-processing to the address word corresponding to layer 2 is that an exceptional region such as "Bukchon" or "Masan" is selected as the candidate administrative district name, the next address word to be considered as layer 3 Then, access the location where the sub-administrative district names of the area are stored in the administrative district name dictionary 3, and process layer 3. Thus, proceeding from layer 2 to layer 3 and then performing the processing of layer 4 (that is, the name of the administrative district corresponding to the exception phrase) can be easily solved by using a pointer to the name of the administrative district having the exception phrase. That is, it maintains a separate pointer value that connects layer 2 and layer 3. In this way, candidate character extraction for layer 2 and layer 3 is performed, and the process is shown in FIG.

Now, look at the post-processing process for layer 2 is shown in Figure 6, first to initialize the number of backtracking to check the number of backtracking applied later (step 80), and in step 82 the address to be processed Characters that may appear at each character position of the word (layer 2) are read from the administrative name dictionary and passed to the character recognizer. Next, in step 84, the candidate characters for each rank are received up to n ranks from the character recognizer. In step 86, a text match is made with the city, county, and city names contained in the administrative name dictionary. At this time, only the number of characters equal to the input characters or more than one administrative name is matched. In step 88, as a result of the matching process, the candidate administrative area names up to the selected k rank in the administrative area name dictionary and the differences obtained in the matching process are memorized. As mentioned in FIG. 5, in order to introduce the concept of backtracking It is necessary. In step 90, as a result of the matching process, the difference obtained from the first priority candidate administrative area name selected from the administrative area name dictionary is compared with a threshold. If the difference obtained by the first candidate administrative district name is larger than the threshold, then step 92 checks how many times backtracking is performed in processing layer two.

In the present invention, backtracking is performed up to k times, which means that the maximum number of backtrackings is performed in order to access the candidate administrative area names up to the selected k rank by matching characters with the administrative area name of the administrative area name dictionary based on the candidate characters for each rank. It is decided. As such, the number of backtrackings is based on the k value used in step 88. When the character recognizer is low in performance, the k value is increased, that is, the candidate administrative area name is considered higher than k ranks, and accordingly, the k rank is higher than the k ranks. By applying back tracking, more accurate address recognition can be achieved. In step 92, it is checked whether the backtracking has been performed k times, and if yes, the process proceeds to step 98. In step 90, it is determined whether or not the first priority zone name is an exception phrase, and proceeds to the respective place.

If backtracking has not been performed k times in step 92, the number of backtrackings is accumulated (step 94), and backtracking is attempted (step 96). Backtracking is first stored in step 66, taking into account the next higher priority candidate administrative district name for the higher tier (here, tier 1), and accessing the administrative zonal name dictionary containing its lower tier, and steps 982) through Step 90 proceeds. The necessity or unity of backtracking with respect to the threshold will be described later. If backtracking has been performed k times in step 92, the process proceeds to step 98. In step 98, k + 1 upper layers (layer 1) and lower layers (layer 2) obtained by k backtracking are performed. Among the candidate administrative district names for the candidates, candidate administrative district names corresponding to each of the upper and lower hierarchies (layer 1 first candidate administrative district name, Tier 1 candidate administrative district name), and (tier 1 candidates Administrative District Name, candidate administrative district name for Tier 2 created by first backtracking, ..., (K-ranked candidate Administrative District name for Tier 1, Tier 1 created for kth backtracking) The candidate administrative district name in the ranking))) sums the differences obtained during the matching process, and selects the candidate administrative district names for the two hierarchies whose values are the smallest. This is to make corrections in consideration of large cases, which will be mentioned again with respect to backtracking.

FIG. 7 is a flowchart showing a post-processing procedure for the exception phrase of the layer 3, first, initializing the number of backtracking (step 112). In step 114, characters that may appear at the position of each character of the address word (layer 3) to be processed are read from the administrative district name and passed to the character recognizer, and in step 116, candidate characters for the rank outputted from the character recognizer are received. Get up to n ranks. Next, in step 118, the text is matched with the exception name included in the administrative name dictionary. At this time, only the number of characters equal to the number of input characters or more than one administrative region name is matched. Hereinafter, steps 120 to 130 are the same as those described in the post-process of FIG. However, the difference is that only the upper layer in backtracking differs from that of FIG.

Now, a description will be given of the process of processing the administrative division names of layers 4 and 5. First, there are some things to keep in mind in Tiers 4 and 5, where administrative names below Tier 4 may contain numbers from 0 to 9, in the form of one number or two consecutive numbers. For example, it refers to an administrative district such as "Sirim 2 dong" or "Sirim 12 dong." Since Hangul is represented by 2 bytes but the number is represented by 1 byte, administrative area names of level 4 and below cannot obtain the desired candidate character using a simple candidate character extraction method such as FIG. Because Figure 3 does not check whether it is a letter or a number, it is unconditionally two bytes long, so if a number appears, the candidate character cannot be obtained correctly. Therefore, the process of checking whether the character is a number or a number in FIG. 4 is added to read two bytes in case of letters and one byte in case of numbers. Although the method described in FIG. 4 can be applied to all layers, the present invention applies the method of FIG. 3 to address words corresponding to layers 1 to 3, and applies to the address words corresponding to layers 4 and 5 in the present invention. The reason for applying the method of 4 degrees is that the number does not occur in the administrative area name above the third layer, so it is unnecessary to check whether it is a letter or a number. Referring again to FIG. 4, in step 30, the administrative area name dictionary containing the administrative area names that may appear in the address word (layer 4 or layer 5) to be processed is approached. Hereinafter, the number of administrative district names to be searched in the dictionary will be referred to as the number of administrative district names. In the next step 32, the number of characters constituting the address word (layer 4 or layer 5) to be processed (hereinafter referred to as the number of characters to be recognized) is counted. In steps 34 and 38, the (DO) loop is repeated to extract candidate characters from the administrative district name dictionary. As described above, the characters must be distinguished from the case where a number appears in the administrative district name and a character appears. (Steps 40 to 48). Then, in step 50, the candidate character array A, which stores characters that can appear in each process of the address word to be processed, is passed to the character recognizer, and the post-processing process is started.

After processing layer 4, if there is another address word to be processed, the address word corresponds to layer 5, and thus, candidate characters for layer 5 are extracted according to FIG. FIG. 8 is a flowchart of the post-processing process of layer 4 and FIG. 9 is a flowchart of the post-processing process of layer 5, and these post-processing processes are fundamental to FIG. I will avoid it. However, in the process of FIG. 8, there is a determination step 160 of whether there is a target character to be recognized and a step 162 of outputting the determined administrative region for each hierarchy. This is to prepare for the case where it could not be filled out due to lack of space.

Now, the character matching method and the method of backtracking proposed in the present invention will be described in detail with reference to Table 2 below.

TABLE 2

Ranked candidate characters and weights used to match address words

Where C ₁ C ₂ ... Cm is an address word consisting of m characters, C ₁ (1), C ₁ (2) ..., C ₁ (n) are the nth candidate characters for the input character C ₁ , and W ( 1), W (2), ..., W (n) represent weights for each rank. When referred to advance one of the address word contained in S ₁ S _{2 ... Sm, C 1 C 2 ...} Cm and S ₁ S ₂ ... Sm of the difference is defined as the sum of the differences of each character. For example, if candidate character C ₁ (4) matches S ₁ (ie, S ₁ = C ₁ (4)), the difference between candidate character C ₁ and dictionary character S ₁ is as much as W (4). The difference (D) between one administrative district name entered and the previous administrative district name can be defined by the following equation.

Where a = 1, 2, ..., n

The weight W (α) used in the present invention is defined as W (α) = α-1 when α represents the rank of the candidate character. If S ₁ does not match any of C ₁ (1), C ₁ (2), ... C (n), the weight is W (n + 1).

In this way, the names of candidate administrative districts of the dictionary having the smallest difference from the candidate letters of the ranking are determined up to the rank k. After gaining knowledge of the lower hierarchy of administrative area names selected as the first priority, the candidate characters are extracted for the next address word that appears next, and the characters are recognized based on the candidate characters, and the ranking is output by the character recognizer. Address recognition proceeds in the same manner as above, which applies postprocessing using candidate characters.

However, the treatment of lower administrative districts by determining the administrative district selected as the first priority may lead to erroneous results. The reason for this is that the difference obtained by each administrative division name in the matching process is the same for two or more administrative division names, or the candidate character obtained by the recognizer is more closely matched with other administrative division names than the corresponding administrative division name. . To solve this problem, we introduce threshold and backtracking for the difference. The determination of the threshold depends on the performance of the character recognizer and the number of characters in the address word being processed. That is, the threshold value is set to be inversely proportional to the recognition rate of the character recognizer and proportional to the number of characters of the address word to be processed. As mentioned earlier, the system's character recognizer compares a single character entered only within a limited character set, resulting in very high performance. Therefore, by selecting the threshold value very small, the accuracy of the post-processing is further increased. In the present invention, it is preferable that the threshold value is 2, but the basis of the present invention is that the candidate characters are first extracted and matched based on this, so that the recognition rate of the character recognizer is naturally increased, and the average number of characters in the Korean administrative district name is 3 Is due to being -4 characters.

When the layer currently being processed is called layer L, if the difference between the candidate characters and the name of the candidate administrative area selected as the first priority by matching the administrative administrative name with the dictionary is less than or equal to the threshold, the stratified layer Lt1 is progressed, otherwise the stratification is performed. At 1L-1, layer L is reprocessed taking into account the next rank. This method of re-searching the upper hierarchy is called backtracking. The thresholds and backtracking introduced above were discussed in October 1983 by International Conference on Text Processing with Izaki et al. Although mentioned in the paper "Post-processing of Handwritten Chinese Character Recognition" published by a Large Charcater Set), this is somewhat different from the method applied to the present invention. In other words, Izaki et al., When the address word entered does not exist in the dictionary or the letters constituting the address word input by the handwriting with severe handwriting deformation do not exist among the candidate letters for each rank output by the recognizer. Since the result of address recognition can be wrongly derived, the threshold value is introduced to process the next layer if the threshold value is lower than the threshold value. Otherwise, candidate characters of the first rank are given as candidate administrative region names.

In addition, when there is confusion in the selection of candidate administrative district names due to the existence of two or more administrative district names having the same difference for a certain layer L, backtracking is introduced as a method to solve this problem.

In other words, when the problem layer is called layer L, the name of the administrative area whose difference is the same in the layer L based on the candidate administrative area name selected first in this layer by first processing the layer L + 1, which is the lower layer of the layer L Among these, the administrative district name corresponding to the upper tier of the candidate administrative district name selected in the hierarchy L + 1 is determined as the candidate administrative district name for the hierarchy L. Therefore, the threshold value used in the present invention was used as a criterion to check that the upper administrative region name may have been incorrectly selected when the difference obtained by the first candidate administrative district name for the layer L becomes greater than or equal to the threshold. Applying backtracking, the backtracking in the present invention is different from the backtracking mentioned in the above paper, in that the processing of the layer L is performed in consideration of the next rank of the upper layer rather than the lower layer.

Even if the above-described threshold check and backtracking are repeatedly performed, if the difference obtained by the candidate administrative district name selected as the first priority for the layer L in the character matching process continues beyond the threshold value, the upper layer of the layer L Consider together the differences obtained for candidate administrative district names for the population L-1. If you attempt k backtracking k times, you get k + 1 candidate addresses for the two tiers, and the sum of the differences of each candidate administrative names in the character matching process is used to determine the candidate administrative names of the two tiers. The names of the administrative districts of the two smaller hierarchies are determined as the candidate administrative district names of the tier L and the tier L-1. In this way, if the administrative area name of the layer L is selected, the layer L + 1 is processed based on this.

By considering the two hierarchies together, if there is more than one difference in the name of each administrative division in any hierarchy, or because of the error of the character recognizer in that hierarchy, the wrong administrative district name is determined as the candidate administrative district name. Even if the next hierarchy is handled based on the wrong candidate administrative district name, correct corrective results can be derived.

In applying the above word matching method, only administrative district names with the same number of input characters are candidate addresses for administrative district names for Tier 1, and administrative subdivisions with the same number of characters or more than one character are candidate candidates. Selected by address. The reason for this is that it can be solved by outputting only the number of input characters for the habits of those who omit the special characters (do, city, city ..) representing the administrative divisions by hierarchy, without expanding the dictionary. However, for Tier 1, dictionaries were expanded because of the large scope of omission and very few administrative district names. Specifically, for "Seoul," "Seoul" or "Seoul" was added to the dictionary.

Through this dictionary search we find that the characters that can appear at any particular location in the address are extremely limited. For example, the characters that can appear in the third digit or more of the name of the administrative district in Tier 3 are "eup", "myeon", "dong", "ga", and "1-9". It may even be that the third letter of Tier 3 belonging to an administrative district is "east." In this case, the character recognition step is meaningless, so in the system according to the present invention, when there is only one character extracted from the administrative division name dictionary, the character recognition unit is immediately inserted into the candidate character array without entering the character recognition unit. Such a method makes it possible to know that the construction of a dictionary of administrative district names containing all the data on the administrative district names in Korea is more important than anything else.

10 is a block diagram schematically illustrating an address recognition apparatus according to the present invention. In the case of an online address recognition apparatus, Hangul address data is read through the tablet 1 as an input apparatus, and in the case of an offline address recognition apparatus, a scanner The Korean address data is read by (1). The Hangul address passed through the input device is just binary data, so the character recognition algorithm can be applied only by dividing this data into character units. Therefore, the input binary data is divided by character units in the address divider 2, and a representative character division method mainly used in the case of an offline address recognition apparatus using a scanner as an input device includes all the pixels stored in the scanner on the x-axis or Projection on the y-axis to use the distribution of black points. The conventional address recognition apparatus starts character recognition immediately after the address splitting step is completed. However, as shown in FIG. 10, it is noteworthy that the candidate characters that may appear at each position of the address before entering the character recognition unit are extracted from the administrative region name dictionary 4 and passed to the character recognition unit. to be. As such, by using the candidate character extracting unit, the range of characters that can be generated can be reduced to enable fast and accurate address recognition. Characters that can appear at each position of the entered word are regarded as a single administrative district until the spaces appear, and are extracted into the character recognition unit by the method described in the flowcharts of FIG. 3 and FIG. The character recognition unit 4 outputs the most similar characters by rank in comparison with the characteristics of the input characters and the characteristics of the respective characters extracted by the candidate character extraction unit 3. The post-processing section 6 matches characters with the administrative section name dictionary 4 using the candidate characters output by the character recognition section 5, and retains some candidate administrative section names for the hierarchy.

When one layer is processed, the lower layer is approached based on the candidate administrative region name generated by the post-processing unit 6. In this post-processing section, the introduction of the threshold and backtracking are applied. Character matching and threshold usage and backtracking methods are described in detail above, and the differences between the layers in applying post-processing to each layer are detailed in the flowcharts of FIGS. In other words, since Tier 1 does not exist in the upper tier, it retains only the names of administrative districts by rank without introducing thresholds, and goes to the lower tier, while Tier 2 determines candidate administrative district names, and these administrative districts are named Bucheon-si, Gyeonggi-do, and Ulsan-si, Gyeongsangnam-do. Check for an exceptional region of the. If it is an exceptional region, it will handle layer 3. Otherwise, it will handle layer 4. In Layer 4, even in the region where Layer 5, which is a lower layer of Layer 4, exists, the writer does not know the region name for Layer 5 or is divided into sections that are too small, resulting in the habit of the note-taker skipping Layer 5, or lack of space. It checks whether there is an address word to be recognized for the processing such as the case where it cannot be entered. Processing the last word of the input address in the manner described above ends with the result output unit 7.

The present invention can be applied to a method and apparatus for recognizing an online or offline Korean address, or to speech address recognition, which is one of the fields of speech recognition that is attracting attention these days. The method can be solved by constructing a dictionary in a format as shown in FIG. 10 and then performing speech recognition for each syllable of the voice address data input through the speaker.

The effect of the present invention is to extract the range of characters to be compared in the Hangul address recognition first, and then to match the characters by using the administrative area name dictionary to create the candidate administrative area name, and then post-process it again to recognize the speed of one character In addition, the recognition rate can be greatly improved and a relatively high recognition rate can be induced. In addition, the misrecognition result generated by the character recognizer can be corrected through post-processing by introducing a threshold value and a backtracking method, thereby making it possible to recognize an address of 99% or more with respect to an input address.

Claims (17)

In the Hangul address recognition method, a candidate character extraction process of extracting candidate characters from administrative region name dictionaries by a candidate character extractor, and the character recognizer recognizes and outputs the candidate characters based on the candidate characters extracted by the candidate character extractor. A post-processing step of character matching candidate characters for each rank with the administrative-zone names listed in the corresponding administrative-zone name dictionaries, recognition of an address word of an input administrative-zone name to be recognized by a character recognition unit, and the post-processing process. And a candidate administrative district name selection process for selecting a candidate administrative district name by means of, and fixing the next address word based on the candidate administrative district name. The method of claim 1, wherein the candidate character extracting process extracts from the administrative region name stored in the administrative region name dictionary corresponding to the address word to be recognized according to the position of each character constituting the address word to be recognized. And a character matching process using character candidates for each rank up to n ranks outputted by the character recognition unit based on the candidate characters extracted in the candidate character extraction process. The method of claim 2, wherein the value n is determined in consideration of a cumulative recognition rate of the character recognizer used. 4. A dictionary according to claim 1 or 2 or 3, wherein the administrative district name dictionaries. District 1 administrative division name dictionary 1, which consists of direct city and road. group. Administrative division name dictionary 2 of hierarchy 2 comprised of city, administrative division name dictionary 3 of hierarchy 3 comprised of exception phrase, administrative division name dictionary 4 of hierarchy 4 comprised of town, town, town and administrative of hierarchy 5 comprised of town Hangul address recognition method comprising a zone name dictionary 5. The Korean address recognition method according to claim 4, wherein the administrative district name dictionary 4 includes a building having a postal code and also includes a building name having no postal code according to a user's purpose of use. 6. The character of the address word to be recognized according to claim 5, wherein the administrative district name dictionary 1 of the tier 1 provides a modified form such as Seoul or Seoul in advance, so that the post-processing process for the tier 1 corresponds to the tier 1 A Hangul address recognition method in which only administrative district names composed of the same number of characters are matched. 7. The method of claim 6, wherein the post-processing process for the hierarchical layer 1 includes: a candidate administrative district name up to k rank selected from an administrative district name dictionary 1 corresponding to the hierarchical layer 1 as a result of the matching process; And storing a difference between administrative district names, wherein k is selected from the selected administrative district names by letter matching with the administrative district name of the administrative district name dictionary 1 based on the candidate letter for each rank output by the character recognizer. Hangul address recognition method characterized in that it is determined according to whether included in the ranking. 8. The method of claim 7, wherein the post-processing process for the layers 2 to 5 matches only the administrative number of the names equal to or more than one letter of the address words to be recognized respectively corresponding to the layers 2 to 5. How to recognize Hangul address. 10. The Korean address recognition method of claim 8, wherein if the result of processing the address word of the layer 2 corresponds to the exception phrase, layer 3 corresponding to the exception phrase is processed; otherwise, the layer 4 address is processed. Way. 10. The method of claim 9, wherein each of the post-processing steps for the layer 2 to layer 5 is a candidate administrative area name up to a k rank selected from an administrative area name dictionary corresponding to each of the layer 2 to layer 5 as a result of the matching process. Storing a difference between the candidate characters for each rank and the candidate administrative area names, wherein the k value corresponds to each of the hierarchical regions 2 to 5 based on the candidate characters for each rank output by the character recognizer; Hangul address recognition method characterized in that it is determined according to the number of rank among the candidate administrative district names selected by character matching with the administrative district name of the dictionary. 11. The method of claim 10, wherein each of the post-processing processes for the layers 2 to 5 corresponds to the rank candidate characters output by the character recognizer in processing any layer L except for the layer 1. If the difference between the administrative name of the administrative name and the candidate administrative name selected as the first priority in the matching process is less than the threshold, proceed to layer L + 1. Otherwise, the layer L is reprocessed considering the next rank. Korean address recognition method using back tracking. 12. The method of claim 11, wherein each of the post-processing steps for the layer 2 to layer 5 ranks first in a process of matching the candidate letter for each rank with the administrative area name of the corresponding administrative area name dictionary even if the backtracking attempts repeatedly. If the difference obtained by the candidate administrative district name continues to be greater than or equal to the threshold value, and the number of times the backtracking is performed k times, considering the candidate administrative district names for the K + 1 upper and lower layers respectively, The names of the candidate administrative districts for the two hierarchies with the smallest value are determined as the names of the candidate administrative districts for the layer L and the layer L-1 by adding the differences obtained from the matching process. In this way, the Hangul address recognition method for processing layer L + 1 based on the selected administrative region name of layer L. In the Hangul address recognition apparatus, an address divider for dividing an address word input through an input device into character units, administrative zone name dictionaries configured according to hierarchical administrative zones, and candidate characters for extracting candidate characters from the administrative zone name dictionaries An extractor, a character recognizer that recognizes only the characters to be recognized in the address word and the candidate characters extracted by the candidate character extractor as matching objects, and candidate characters for each rank in the corresponding administrative section name dictionaries; And a post-processing unit for selecting candidate administrative district names by letter matching with administrative district names and a result output unit for outputting a recognized address. 15. The candidate character extraction of claim 13, wherein the candidate character extraction is extracted from the administrative region name contained in the administrative region name dictionary corresponding to the address word to be recognized according to the position of each character constituting the address word to be recognized, and the post-processing is the extracted candidate. Character matching apparatus using a character candidate for each rank up to n ranks output by the character recognition unit based on the characters. 15. The administrative district name dictionary according to claim 14, wherein the administrative district name dictionaries comprise a dictionary of administrative district name 1 of tier 1 composed of cities, districts, and roads, and an administrative district name dictionary 2 of tier 2 composed of districts, cities, and cities. A Hangul address recognizing apparatus comprising: dictionary 3, administrative division name dictionary 4 of hierarchy 4 composed of eup.myeon, dong, and administrative division name dictionary 5 of hierarchy 5 composed of li. The Korean address recognition apparatus according to claim 15, wherein the administrative district name dictionary 4 includes a building having a postal code, and also includes a building name having no postal code according to a user's purpose of use. 17. The apparatus of claim 16, wherein the administrative district name dictionary 1 of the layer 1 includes a modified form such as Seoul and Seoul.