JP5217127B2 - Collective place name recognition program, collective place name recognition apparatus, and collective place name recognition method - Google Patents

Description

The present invention is Bulk place name recognition program relates collectively place name recognition apparatus and collectively place name recognition method, in particular, Bulk place name recognition program that recognizes reads the address from the handwritten characters without frame region, collectively place name recognition The present invention relates to an apparatus and a batch place name recognition method.

  Conventionally, two handwritten address recognition methods have been proposed.

  In the first conventional method, first, a plurality of single character areas (that is, areas including only one character) are extracted from the input address image. Next, key characters are extracted from the plurality of single character areas (in handwritten address recognition, key characters represent administrative districts such as provinces, districts, states, cities, towns, counties, dings, villages, etc. Refers to a single character that can be.) And after detecting the place name area | region divided by two continuous key characters, the character in the place name area is cut out one by one, and the cut out character is recognized one by one. Thereby, the place name in the place name area is recognized.

  However, the first conventional method has a problem that an error often occurs when handwritten address characters are separated one by one. In particular, when an address is written in a region without a frame with characters in contact with each other, there is a possibility that this method cannot be recognized. Such inconveniences often occur because handwritten characters are not shaped and have large variations.

  Also in the second conventional method, a cut-out character is first extracted, and a place name area delimited by two consecutive key characters is detected. The second conventional method is different from the first conventional method in that the place names in the place name area are collectively recognized.

、「路」など）であるか否かを決定する。当該単一文字領域の文字に対する認識候補文字（すなわち、予備的にその領域の文字である可能性があると判断された文字）として、複数のキー文字があった場合、照合距離が最小のキー文字のみを選択して当該単一文字領域を認識する。なお、照合距離とは、入力文字の画像から文字の特徴を抽出した特徴ベクトル（数値列）とあらかじめ特徴辞書に持つ各文字の特徴ベクトルとを比較して得られる距離値を言う。完全一致すれば０になるので、小さいほど両者が近い。照合距離が小さいほど、キー文字と当該文字領域内の画像との類似性が高い。 FIG. 1 is a block diagram showing in detail the processing of the second conventional method. As shown in FIG. 1, the character cutout unit 601 cuts out all possible single character regions from the input address image. The key character extraction unit 602 refers to the key character dictionary 603 to recognize each possible single character area, and the characters in the single character area are key characters (for example, “Ministry”, “City”,
[Character 1]

, “Road”, etc.). If there are multiple key characters as recognition candidate characters for the characters in the single character region (that is, characters that are preliminarily determined to be characters in the region), the key character with the smallest matching distance Select only to recognize the single character area. The collation distance is a distance value obtained by comparing a feature vector (numerical string) obtained by extracting a character feature from an input character image with a feature vector of each character previously stored in the feature dictionary. Since it becomes 0 if it completely coincides, the smaller the closer, the closer. The smaller the collation distance, the higher the similarity between the key character and the image in the character area.

はキー文字パスである。他方、
［文字１］

は「市」の上位行政単位ではなく、逆に、行政的な意味では「市」が
［文字１］

の上位にあるため、
［文字３］

はキー文字パスではない。地名語領域候補切り出し手段６０６は、キー文字パスにおける二つのキー文字によって区切られる地名語領域を抽出する。地名語一括認識手段６０７は、地名語認識辞書６０８を参照して、地名語を一括認識する。一文字認識手段６０９は、まず、一文字認識辞書６１０を参照して、地名でもキー文字でもない部分を入力画像から抽出する。例えば、
［文字４］

という住所からは、建物の名称と建物内での番号、つまり
［文字５］

を抽出する。そして、抽出した部分から複数の文字を切り出し、切り出した文字一つ一つを認識する。住所決定手段６１１は、住所知識部６１２内に入力された住所知識を使用し、認識結果と実際の住所とを比較して両者が一致するか否かを調べ、住所の全照合距離が所定の閾値よりも大きいか否かを調べる。所定の閾値よりも大きい場合、この住所の認識結果は信頼できないと判断され、システムはこの認識をリジェクトする。所定の閾値以下の場合、実際の住所と一致した認識結果が最終的に出力される。 Based on the address hierarchy structure knowledge unit 605, the key character set determination unit 604 combines all possible key character candidates as key character paths. In each key character path, the requirement that the preceding key character is an administrative unit that is legally higher than any following key character must be met. For example,
[Character 2]

Is the key character path. On the other hand
[Character 1]

Is not a higher-level administrative unit of "city", but conversely, in the administrative sense, "city" is [character 1]

Because it is at the top of
[Character 3]

Is not a key character path. The place name area candidate cutout unit 606 extracts a place name area divided by two key characters in the key character path. The place name word collective recognition unit 607 refers to the place name word recognition dictionary 608 and collectively recognizes place name words. First, the single character recognition unit 609 refers to the single character recognition dictionary 610 and extracts a portion that is neither a place name nor a key character from the input image. For example,
[Character 4]

From the address, the name of the building and the number in the building, that is, [Character 5]

To extract. Then, a plurality of characters are cut out from the extracted portion, and the cut out characters are recognized one by one. The address determination unit 611 uses the address knowledge input in the address knowledge unit 612 and compares the recognition result with the actual address to check whether or not they match, and the total collation distance of the address is a predetermined value. It is checked whether it is larger than the threshold value. If it is greater than the predetermined threshold, the address recognition result is determined to be unreliable and the system rejects this recognition. When the value is equal to or less than the predetermined threshold, a recognition result that matches the actual address is finally output.

であれば、キー文字パスは、「市−区−路」である。この場合、第１のキー文字「市」によって区切られる第１レベル地名領域は「北京」と認識される。次に、第１および第２のキー文字「市−区」によって区切られる地名領域が抽出される。そして、抽出された地名領域全体と、地名語認識辞書６０８中に記憶されている北京市内の全ての「区」とを照合することで地名領域が認識される。さらに、第３レベル地名照合処理１０７において、まず、（もし第２および第３のキー文字があれば）第２および第３のキー文字によって区切られる地名領域を切り出す。次に、第２レベルの地名照合結果と地名語認識辞書６０８とを参照して、切り出された地名領域を認識する。そして、最後のレベルになるまで同様の処理を繰り返す。第Ｎレベルの地名照合処理１０８においては、（もし第Ｎ−１のキー文字および第Ｎのキー文字があれば）第Ｎ−１のキー文字と第Ｎのキー文字とによって区切られる地名領域を抽出し、第Ｎ−１レベルの地名照合結果と地名語認識辞書６０８とを参照して、抽出した地名領域を認識する。 FIG. 2 shows a place word collective recognition process. In the first level place name matching process 105, first, place name areas delimited by the first key characters are cut out, and place names of the place name areas cut out using the place name word recognition dictionary 608 are collectively recognized. Next, in the second level place name matching process 106, the place name area delimited by the first key character and the second key character (if there is a second key character) is cut out, and the first level place name matching result is obtained. And the place name area dictionary 608 shown in FIG. For example, if the address actually entered is [Character 6]

If so, the key character path is “city-ku-road”. In this case, the first level place name area delimited by the first key character “city” is recognized as “Beijing”. Next, the place name area delimited by the first and second key characters “city-ku” is extracted. Then, the place name area is recognized by collating the entire extracted place name area with all “gu” in Beijing city stored in the place name word recognition dictionary 608. Further, in the third level place name matching process 107, first, the place name area delimited by the second and third key characters is cut out (if there are second and third key characters). Next, the extracted place name region is recognized by referring to the place name collation result of the second level and the place name word recognition dictionary 608. The same processing is repeated until the last level is reached. In the Nth level place name matching process 108, the place name area divided by the N-1th key character and the Nth key character (if there is an N-1th key character and an Nth key character) The extracted place name region is recognized by referring to the place name collation result of the (N-1) th level and the place name word recognition dictionary 608.

US Pat. No. 6,535,619 (B1) Bob Fisher, Simon Perkins, Ashley Walker and Erik Wolfart, "Hypermedia Image Processing Reference", [online], Department of Artificial Intelligence, University of Edinburgh, UK, Internet <URL: http://www.cee.hw.ac. uk / hipr / html / label.html> Qiuqi RUAN, "Mathematical ImageProcessing", the Publishing House of the Electronics Industry, page 325 E. Ukkonen, "On approximate atring matching", Proc. Int. Conf. On Foundations of Comp. Theory, Springer-Verlag, LNCS 158 p 487-495,1983 S.B. Needleman and C.D.Wunsch, "A general method applicable to the search for similarities in the aminoacid sequence of two proteins", Journal Molec. Biol. 48 p 443-453, 1970 Christopher J.C. Burges, "A Tutorialon Support Vector Machines for Pattern Recognition, Data Mining and KnowledgeDiscovery", volume 2, No 2, pages 121-167, 1998.

となるような状況下では、上記従来方法では、「市」ではなく、
［文字８］

の方が当該単一文字領域のキー文字であると認識される可能性が高いため、システムの精度を低下させる。 One of the problems with the conventional method described above is that the accuracy of the system is lowered because recognition fails if no correct key characters are extracted. Specifically, the key character extraction has two problems with the conventional method. First, (1) the one-character recognition means 609 in FIG. 1 collates all single character areas with all characters in the one-character recognition dictionary (as far as Chinese words are concerned, there are thousands of kanji used). Therefore, an error inevitably occurs in key character recognition, thereby reducing the recognition accuracy of the system. Next, (2) the key character extracting means 602 only recognizes a cut-out character that appears in the recognition candidate with the smallest collation distance as a key character. Will occur. For example, an image in a single character area actually corresponds to “city”, but when words are arranged according to a collation distance, a recognition candidate string is [character 7].

Under such circumstances, the above conventional method is not “city”,
[Character 8]

Since there is a high possibility that this is recognized as a key character of the single character area, the accuracy of the system is lowered.

という住所文字列が入力されたときに、「市」で区切られた第１レベルの地名が誤って
［文字９］

と照合されてしまうと、第１のキー文字「市」と第２のキー文字「区」とで区切られる実際の地名
［文字１０］

は、誤った第１レベルの地名
［文字９］

に含まれる行政単位ではないため、第２レベルの照合結果は常に誤りとなってしまう。さらに、住所が手書き入力された場合、地名の字体は様々に変形し、大きさもまちまちであるため、認識された単語が実際の住所と常に一致するとは限らない。このため、システムの精度が著しく低下してしまう。 Further, the conventional method has the following problems. The sequential processing structure as in the conventional method does not work when any preceding place name recognition fails. More specifically, subsequent modules are highly dependent on previous modules. If place name recognition at the preceding level fails, subsequent place name recognition will always be an error. For example, if the first level place name matching 105 does not output a correct result,
[Character 6]

When the address string is entered, the first level place name separated by "city" is mistakenly [character 9]

The actual place name delimited by the first key character “city” and the second key character “ku” [character 10].

Is the wrong first-level place name [character 9]

Because it is not an administrative unit included in, the second level collation result is always incorrect. Furthermore, when the address is input by handwriting, the place name fonts are variously deformed and the sizes thereof vary, so that the recognized word does not always match the actual address. For this reason, the accuracy of the system is significantly reduced.

  Further, the conventional method has the following problems. Due to the unique property that the quality of the handwritten input address varies, if the recognition result is rejected using a fixed threshold value, an address reading error occurs. Specifically, if a small value is set as a threshold value, the above-described conventional method rejects most of the recognition results for low-quality addresses, whereas if a large value is set as a threshold value, recognition results for high-quality addresses are rejected. Most of them are approved. For this reason, the reliability of the system is significantly impaired.

  A first object of the present invention is to provide an address recognition apparatus that can perform address recognition efficiently.

  The second object of the present invention is to provide an address recognition apparatus capable of performing address recognition with high accuracy.

  A third object of the present invention is to provide an address recognition device that can perform reliable address recognition.

  In order to achieve the above object, the present invention provides a key character extraction program that executes extraction of key characters from a plurality of single character regions for each of the plurality of single character regions, and an image corresponding to the single character region. A feature vector extraction procedure for extracting a feature vector from the feature vector, a feature vector extracted by the feature vector extraction procedure, and a feature vector of a key character in a key character dictionary storing all possible key characters; A collation procedure for retrieving all key character candidates and a key character for determining a valid key character path when recognizing a plurality of single character regions each having one or more key character candidates retrieved by the collation procedure And causing the computer to execute a path determination procedure.

  The present invention is also a key character extraction device for extracting a key character from a plurality of single character areas, and feature vector extraction means for extracting a feature vector from an image corresponding to each of the single character areas, and all possible The key character dictionary unit that stores only the key characters of the key, the feature vector extracted by the feature vector extraction means, and the feature vector of the possible key characters stored in the key character dictionary unit, and all possible A key character path for determining a valid key character path when recognizing a plurality of single character areas each having one or more key character candidates searched by the matching means And a determining means.

  The present invention is also a key character extraction method for extracting key characters from a plurality of single character areas for each of the plurality of single character areas, and extracting a feature vector from an image corresponding to the single character area. The feature vector extraction step, the feature vector extracted by the feature vector extraction step, and the feature vector of the key character in the key character dictionary storing all possible key characters are collated, and all possible key character candidates are obtained. A matching step for searching, and a key character path determining step for determining a valid key character path when recognizing a plurality of single character regions each having one or more key character candidates searched by the matching step. It is characterized by including.

  The present invention is also a collective place name recognition program for causing a computer to execute a collective place name recognition procedure for recognizing place names of place name images including place names at a certain administrative level, wherein the collective place name recognition procedure includes a feature vector from the place name image. The computer executes an extraction procedure for extracting a name, and a matching procedure for searching for all possible place name candidates by matching the feature vector extracted by the extraction procedure with the feature vector of the place name stored in the place name dictionary. A place name corresponding to a feature vector whose matching distance between the place name stored in the place name dictionary and the feature vector extracted by the extraction procedure is smaller than a predetermined threshold is determined as a possible place name candidate, Furthermore, an individual character recognition procedure for recognizing character candidates for characters included in each single character region of the place name image, and the individual character recognition procedure Based on more obtained recognition result, characterized in that to perform said a collective place name recognition procedure place name verification procedure for filtering the place name candidate determined by, to the computer.

  The collective place name recognition procedure preferably includes an individual character recognition procedure and a place name verification procedure.

  The present invention is also a collective place name recognition method including a collective place name recognition process for recognizing place names of place name images including place names at a certain administrative level, wherein the collective place name recognition step extracts a feature vector from the place name image. An extraction step; a collation step of collating the feature vector extracted in the extraction step with the feature vector of the place name stored in the place name dictionary and searching for all possible place name candidates; Among the stored place names, a place name corresponding to a feature vector whose collation distance with the feature vector extracted by the extraction step is smaller than a predetermined threshold is determined as a possible place name candidate. An individual character recognition step for recognizing character candidates for characters included in each character region, and the batch location based on the recognition result obtained by the individual character recognition step. A place name verification step of performing filtering of place name candidates determined by the recognition process, characterized in that it contains.

  The collective place name recognition method preferably includes an individual character recognition step and a place name verification step.

  Further, the present invention is a collective place name recognition device, wherein an extraction means for extracting a feature vector from an inputted place name image, a place name dictionary unit for storing a place name, and a place name feature vector stored in the place name dictionary unit Are compared with the feature vector extracted by the extraction means, and all possible place name candidates and collation distances are obtained, and among the place names stored in the place name dictionary, between the feature vectors extracted by the extraction means A collation means for determining a place name candidate corresponding to a feature vector whose collation distance is smaller than a predetermined threshold and a plurality of place name images including administrative place names arranged in descending order are input, and the place name image is recognized. And a collation evaluation means for evaluating and determining a plurality of address path candidates when a plurality of address paths including the administrative level place names in descending order are formed. And said that there were pictures.

  The present invention is also a collective place name recognition method including a collective place name recognition process for recognizing place names of place name images including place names at a certain administrative level, wherein the collective place name recognition step extracts a feature vector from the place name image. An extraction step; a collation step of collating the feature vector extracted in the extraction step with the feature vector of the place name stored in the place name dictionary and searching for all possible place name candidates; Among the stored place names, a place name corresponding to a feature vector whose collation distance with the feature vector extracted by the extraction step is smaller than a predetermined threshold is determined as a possible place name candidate. An individual character recognition step for recognizing character candidates for characters included in each character region, and the batch location based on the recognition result obtained by the individual character recognition step. A place name verification step of performing filtering of place name candidates determined by the recognition process, characterized in that it contains.

  Further, the present invention provides a character separation step for separating a single character region, a key character extraction step for extracting a key character for each single character region separated by the character separation step, and a key character by the key character extraction step. An address recognition method including a collective place name recognition step for collectively recognizing place names based on an extraction result of the address, an address determination step for determining an address based on a place name recognition result by the collective place name recognition step, and the key The character extraction step includes a feature vector extraction step of extracting a feature vector from an image corresponding to the single character region, a feature vector extracted by the feature vector extraction step, and a key character dictionary storing all possible key characters A collation step of collating with a feature vector of possible key characters and searching for all possible key character candidates; and Ri retrieved one or more key character candidates when recognizing a plurality of single-character area having respectively, characterized in that it includes a key character path determination step of determining a valid key characters path, the.

  According to one aspect of the present invention, in key character recognition, the present invention compares only with a key character dictionary (including only M key characters). This reduces the amount of processing and increases the recognition accuracy. Further, when there is only one key character in the candidate list of the combination pattern evaluation result at the time of key character recognition, the combination pattern is output as one of the possible key character candidates. Further, when the combination pattern includes L (L <= M) possible key character candidates, L possible key character candidates are output. Thus, for example, if two key character candidates ("Province" and "City") are included in the combined pattern, the two possible key characters "Province" and "City" are used for the combined pattern. Is output. This reduces the risk of losing the correct key characters and increases the accuracy of address recognition. In addition, when recognizing a plurality of single character areas each having one or more key character candidates, since a valid key character path is determined, it is not necessary to evaluate all the key character paths as in the conventional method. Since it is only necessary to select high-rank K key character paths having the smallest average collation distance, the efficiency and accuracy of address recognition can be improved.

  According to another aspect of the present invention, a place name verification method is provided. In this place name verification method, a place name region recognition result by batch place name recognition, and a result of recognizing each one by cutting out a single character from the place name. Are combined, and the result of collective place name recognition is filtered. In this method, by defining the edit distance between the individual character recognition result and the actual place name, the verification is performed by efficiently combining the collective place name recognition result and the individual character recognition result. When collective place name recognition cannot distinguish a large number of similar place names, normal recognition treats both separately, which increases the accuracy of address recognition.

  According to still another feature of the present invention, when performing place name recognition at a certain level, candidates from the first candidate to the Mth candidate can be generated. If the recognition result of the level does not match the actual place name, the place name recognition of the subsequent level also becomes an error, and M candidates having a large collation distance are output. Therefore, when a candidate with a large collation distance is generated by place name recognition, there is a very high possibility that the place name recognition at the previous level is incorrect. Therefore, feedback processing from the final level to the first level is performed. That is, the recognition result of the subsequent level is used to correct an error that has occurred at the previous level. This increases the accuracy of address recognition. According to still another feature of the present invention, place name recognition with feedback processing can be performed without evaluating all candidates, and the efficiency and accuracy of place name recognition are significantly increased. In addition, when a plurality of place name images including administrative level place names arranged in descending order are input, and a plurality of address paths including the administrative level place names in descending order are formed by recognition of the place name images, a plurality of address path candidates are formed. This makes it possible to improve the reliability of address recognition.

  Below, with reference to drawings, the address recognition method and address recognition device of the present invention are explained in detail. In the drawings, the same or similar components are denoted by the same reference numerals.

  FIG. 3 is a block diagram illustrating an address recognition apparatus according to an embodiment of the present invention. In FIG. 3, the single character area detecting means 1 cuts out single characters one by one from the input address image. Each image area of a single character is called a single character area. The key character extraction means 2 extracts key characters according to the result of the single character area detection means 1. Next, the place name area extracting unit 3 extracts a place name area based on the key characters extracted by the key character extracting unit 2. The collective place name recognition means 4 recursively recognizes the place name areas extracted by the place name area extraction means 3. The single character place name recognition means E1 recognizes the character output from the single character area detection means 1 in association with the place name in the place name area extracted by the place name area extraction means 3. The place name verification means 5 verifies and determines the place name recognition results output by the collective place name recognition means 4 and the single character place name recognition means E1. The reject means 6 determines whether to reject or approve the place name recognition result.

  Specifically, when an address image to be recognized is input, first, the single character area detecting means 1 cuts out a plurality of single character areas from the input address image. After cutting out the single character area, the key character extracting means 2 recognizes each single character area and extracts the key character. After extracting the key character, the place name area extracting means 3 separates the area delimited by the key character from the input image, and then the collective place name recognizing means 4 includes the feature vector of the area delimited by the key character, The place name feature vectors in the place name dictionary are collated, and place name candidates whose collation distance is smaller than a predetermined threshold are respectively obtained as place name candidate strings. Here, the collation distance indicates the degree of dissimilarity between an image divided by key characters and a place name candidate. The smaller the collation distance value, the smaller the difference between the image determined at that time and the place name candidate. That is, both are more similar. The calculation of the verification distance will be described later. At the same time, the ratio of the difference between the collation distance of each place name candidate in the place name candidate string and the minimum collation distance of the place name candidate to the minimum collation distance of the place name candidates is sequentially estimated. If the estimated ratio is greater than a predetermined threshold, it is determined that the place name candidate is significantly different from the image, and the place name candidate is excluded. Therefore, in the present invention, it is not necessary to evaluate all the place name candidates. That is, if the determined image and the place name candidate are significantly different from each other, it is not necessary to perform further comparison (refer to the description of FIGS. 5 and 6 for details), so that the recognition time is shortened. At this time, by using the single character place name recognition means E1 and the place name verification means 5 and combining the individual character recognition result and the collective place name recognition result, the collectively recognized place names can be filtered. Similarly, place name areas of different levels are recognized one by one. In many cases, the first place name candidates in place name recognition are usually incorrect, and other place name candidates have the correct result. Therefore, the matching distance of each corresponding path is calculated sequentially, the path with the smallest average matching distance is found, and the path is determined as the final place name path. Thus, the multi-candidate recursive recognition used in the present invention improves the recognition accuracy. In addition, when there are a plurality of paths close to the path with the smallest average collation distance, the single character place name recognition means E1 and the place name verification means 5 are used to combine the individual character recognition result and the collective place name recognition result. The place name candidates that are inconsistent with the plurality of paths among the place name candidates collectively recognized for the place name image can be filtered.

  The single character place name recognition means E1 is present in the area delimited by the key characters, and extracts the characteristics of the single character area obtained by the single character area detection means 1, and extracts the extracted features and the single character dictionary. Recognize by sequentially comparing the character characteristics. After the place name is recognized, the place name verification means 5 combines the batch place name recognition result and the individual character recognition result, and finally checks the recognition result. As a result, the single character area recognition result and the recursive collective recognition result are combined to improve the recognition accuracy.

  Each means of FIG. 3 is demonstrated in order below.

  FIG. 4 is a block diagram showing the single character area detecting means 1 of FIG.

  In FIG. 4, the coupling element labeling means 34 labels all the coupling elements in the input image by a coupling element detection algorithm (see Non-Patent Document 1). A combination element is a set of pixel points of a character display color (usually black) in an image. Any two pixel points in the set can be combined by pixels in the set.

  Based on the size of the labeled coupling element, the position of the labeled coupling element, the distance from the adjacent element, etc., the single character region candidate determination means 35 integrates the corresponding coupling elements into a new coupling element. To do. The new combined element is a single character region candidate. For example, if the input address image is written sideways, if the coupling element A is positioned above the coupling element B, B and A should belong to the same character, so B and A And must be integrated into a new coupling element. For details of the integration of the coupling elements, refer to FIGS. 11A, 11B, and 11C of Patent Document 1 and the description of Patent Document 1.

The average character size calculating means 36 removes the apparently abnormal size of the combined elements obtained by the single character area candidate determining means 35 and calculates the average size of the remaining combined elements and characters. Here, CC _i (i = 1, 2,..., M _cc ) represents all m _cc combined elements obtained by the single character region candidate determining means 35, and Width (CC _i ) is If the width of the coupling element CC _i is represented and Height (CC _i ) represents its height, all CC _k satisfying any one of the following conditions (1) to (4) are removed.

(1) Width (CC _k ) <WCC _median − Th _{width_low}
(2) Width (CC _k )> WCC _median + Th _{width_low}
(3) Height (CC _k ) <HCC _median − Th _{height_low}
(4) Height (CC _k )> HCC _median + Th _{height_low}
Here, WCC _median is a _median value obtained from all Width (CC _i ) (i = 1, 2,..., M _cc ), and HCC _median is all Height (CC _i ) (i = 1, 2,..., M _cc ), and Th _{width_low} and Th _{height_low} are two predetermined positive thresholds. Refer to Non-Patent Document 2 for details of the median calculation.

  The pixel projecting unit 37 projects the input address image in the vertical direction (if the image is written in the horizontal direction), and converts the input address image (if the image is written in the vertical direction). Project in the horizontal direction. Vertical (or horizontal) image projection yields the number of black pixels at each point in the horizontal (or vertical) image.

  Temporary separation point candidate extraction means 38 and temporary separation point candidate determination means 39 are used to separate the new combined elements output from single character area candidate determination means 35 into actual character areas. That is, it is used to separate the combined characters into individual single characters.

  According to the number of black pixels obtained by the pixel projection unit 37, the temporary separation point candidate extraction unit 38 extracts position points that satisfy the following two conditions as temporary separation point candidates. These position points are points that can become separation points for combined characters.

(Condition 1) The position point is a point having a minimum value in the histogram of the number of black pixels obtained by projection.
(Condition 2) The number of black pixels on the position point is smaller than a predetermined threshold Th _seg .

  The extraction of the separation point candidates can be performed using, for example, the method described in Patent Document 1.

  Temporary separation point candidate determination means 39 determines whether or not the temporary separation point obtained from temporary separation point candidate extraction means 38 is a true separation point of the adjacent character area.

When the input character pattern is written horizontally, for each of the separated coupling elements, an area adjacent to a position corresponding to a multiple of one half of the average character width from the start position of the coupling element Temporary separation point candidates located within are determined as separation points. Specifically, when _Pi (0 <i <s: i is a positive integer) is a position point at a point corresponding to a multiple of one half of the average character width from the start position of the coupling element. ,
P _i -Th _delta <C <= Th _delta + P _i
If there are separation points C and i satisfying the above, it is determined that C is a separation point. The connecting element is then cut longitudinally into two new connecting elements at point C.

When the input character pattern is written vertically, any of the separated coupling elements is adjacent to a position corresponding to a multiple of one half of the average character width from the start position of the coupling element. Temporary separation point candidates located within the region are determined as separation points. Specifically, when _Pi (0 <i <s: i is a positive integer) is a position point at a point corresponding to a multiple of one half of the average character width from the start position of the coupling element. ,
P _i -Th _delta <C <= Th _delta + P _i
If there are separation points C and i satisfying the above, it is determined that C is a separation point. The connecting element is then cut transversely into two new connecting elements at point C.

  FIG. 5 is a block diagram showing the key character extraction means 2 according to the embodiment of the present invention.

  In FIG. 5, the key character candidate extraction unit 20 extracts key character candidates from the single character region (or combination element) input from the single character region detection unit 1. The key character path determining means 21 refers to the address hierarchy structure knowledge unit 24 and determines an appropriate key character path. Here, the address hierarchical structure knowledge unit 24 is a database or a program that can provide hierarchical knowledge of place names. The key character path filter means 22 stores a plurality of key character paths, and calculates the average collation distance of key characters in each path (that is, the ratio of the sum of collation distances of various levels of place names to the number of address levels). All paths are calculated in ascending order (from smallest to largest), and finally K key character paths with higher ranks are selected as final key character paths. Here, K is a predetermined positive integer.

はキー文字パスであるが、
［文字１］

は「市」の上位行政単位ではなく、「市」は行政上の意味で
［文字１］

を包含するものであるから、
［文字３］

はキー文字パスではない。次に、キー文字パスフィルタ手段２２が、認識結果中の各キー文字の平均照合距離を抽出する。各キー文字パスは、順次順位付けられる。地名領域抽出および認識においては、平均照合距離が最も小さいＫ個の文字パスのみについて評価を行う。ここで、Ｋは、あらかじめ定められた正の整数である。例えば、入力された住所が
［文字１１］

である場合、”京”および”市”に対応する画像領域について、「市」がキー文字と認識され、かつ、”区”に対応する画像領域について、「区」がキー文字と認識された場合、キー文字パスは、「市」（”京”）−「区」（”区”）、「市」（”市”）−「区」（”区”）となる。この場合、「市」（”京”）−「区」（”区”）において”京”および”区”が「市」および「区」として認識された場合の平均照合距離と、「市」（”市”）−「区」（”区”）において”市”および”区”が「市」および「区」として認識された場合の平均照合距離とが計算される。キー文字の平均照合距離が大きいほど、それが正しいキー文字パスである可能性は低いため、可能な限り除去せねばならない。このように、本発明の地名領域抽出・認識においては、従来の方法のように全てのキー文字パスを評価する必要はなく、平均照合距離が最小である高順位のＫ個のキー文字パスを選択するのみでよいため、住所認識の効率および正確性を高めることができる（ただし、Ｋはあらかじめ定められた正の整数である）。 Specifically, when an address image is input for recognition, first, a combination element is extracted from the input address image, and then a key character candidate extraction unit 20 extracts a key character from the combination element ( For details, see the description of FIG. 6 below). After extracting the key characters, the key character path determining means 21 determines a key character path that matches the place name hierarchy knowledge. The determined key character path describes the order of the key characters, and the preceding key character is an administrative unit higher than the succeeding key character. For example,
[Character 2]

Is the key character path,
[Character 1]

Is not a higher-level administrative unit of “city”, but “city” is an administrative meaning [letter 1]

Because it includes
[Character 3]

Is not a key character path. Next, the key character path filter means 22 extracts the average collation distance of each key character in the recognition result. Each key character path is ranked sequentially. In place name area extraction and recognition, only K character paths with the smallest average collation distance are evaluated. Here, K is a predetermined positive integer. For example, if the input address is [Character 11]

, “City” is recognized as a key character for the image areas corresponding to “Kyo” and “City”, and “Ku” is recognized as a key character for the image area corresponding to “Ku”. In this case, the key character path is “city” (“Kyo”) − “ku” (“ku”), “city” (“city”) − “ku” (“ku”). In this case, the average collation distance when “Kyo” and “Ku” are recognized as “City” and “Ku” in “City” (“Kyo”) − “Ku” (“Ku”), and “City” The average collation distance when “city” and “city” are recognized as “city” and “city” in (“city”) − “city” (“city”) is calculated. The greater the average collation distance of key characters, the less likely it is to be a correct key character path, so it must be removed as much as possible. As described above, in the place name area extraction / recognition of the present invention, it is not necessary to evaluate all the key character paths as in the conventional method, and K key character paths with high ranks having the smallest average collation distance are used. Since only selection is required, the efficiency and accuracy of address recognition can be improved (where K is a predetermined positive integer).

  FIG. 6 is a block diagram of key character candidate extraction means 20 according to the embodiment of the present invention.

  The feature extraction unit 42 extracts a feature vector from the image corresponding to the coupling element (that is, the single character region) input from the single character region detection unit 1. The dictionary collating unit 43 refers to the key character dictionary unit 41, collates the feature vector output from the feature extracting unit 42 with the feature vector of the key character in the key character dictionary unit 41, and has a possible key Output characters as candidates. The key character candidate determining means 44 determines whether or not these selected key character candidates are true key characters.

である。結合要素との照合距離Ｄｉｓ（ＣＣ_i，Ｋｃ）が所定の閾値Ｔｈ_keyよりも小さいキー文字Ｋｃが存在する場合、当該結合要素は、キー文字の可能性がある候補Ｋｃであると決定される。照合距離Ｄｉｓ（ＣＣ_i，Ｋｃ）は、ｆＣＣ_iとｆｋｅｙ_Kcとの間のユークリッド距離と定義される。すなわち、

である。 Specifically, when a coupling element is input from the single character area detection unit 1, the feature extraction unit 42 first extracts the feature of the coupling element. Next, the dictionary collation means 43 collates the feature vectors extracted from the images of these coupling elements (labeled as fCC _i ) with the feature vectors of all the key characters in the key character dictionary unit 41. The key characters in the key character dictionary section 41 are labeled fkey _k (where k = 1, 2, 3,..., 22), and correspond to all 22 key characters in a Chinese address. That is,
[Character 12]

It is. When there is a key character Kc whose collation distance Dis (CC _i , Kc) with the combined element is smaller than a predetermined threshold Th _key , the combined element is determined to be a candidate Kc that may be a key character. . The collation distance Dis (CC _i , Kc) is defined as the Euclidean distance between fCC _i and fkey _Kc . That is,

It is.

As described above, the dimensions of fCC _i and fkey _Kc are m.

  In this way, the key character candidate extracting means 20 collates a single character area with only a limited number (22 in the case of Chinese) of key characters, and all thousands of Chinese characters as in the conventional method. There is no need to check against. This increases the efficiency of the system and at the same time increases the accuracy of key character recognition.

In addition, in the recognition of the coupling element, when the collation distance with L (L <= 22) candidates is smaller than a predetermined threshold Th _key , the coupling element may be L key characters. It is recognized that there is. Therefore, for example, if the key characters that may be the combination elements are “Province” and “City”, the combination element is recognized as two key characters “Province” and “City”. This reduces the risk of losing correct characters and increases the accuracy of address recognition.

Furthermore, the key character candidates coupling element CC _i are L key character _{(K 1, K 2, ···} , K L) ( _{but, K 1, K 2, ···} , K L is the matching distance Therefore, the key character candidate determination unit 44 recognizes the combination element as including only key characters satisfying the following expression.
[Dis (CC _i , K _n ) − Dis (CC _i , K ₁ )] / Dis (CC _i , K ₁ ) <Th _{dis_k}
Here, Dis (CC _i , K _n ) represents a collation distance between the key character K _n and the coupling element CC _i, and Th _{dis_k} is a positive constant and represents a predetermined threshold value. By using such a key character removal method, a key character close to the first character candidate is held, and evaluation is performed in subsequent key character path determination and place name verification. If there is a key character that does not satisfy the above conditions, the collation distance of the key character indicates that it is far from the collation distance of the first candidate, so it is unlikely that the key character is a correct key character, Such key characters are removed. Thereby, after the key character extraction, there is no need to extract and recognize the place name area in order to evaluate all the possible key characters, and the efficiency and accuracy of address recognition are improved.

  FIG. 7 is a block diagram showing the collective place name recognition means 4 according to the embodiment of the present invention. As shown in FIG. 7, the place name collating means 27 extracts features from images corresponding to place names at various levels inputted from the place name area extracting means 3, and stores place name dictionary (place name database, place name dictionary, etc.). The place name feature stored in the unit 54 is collated, thereby outputting the first to Ncth candidates. The place name candidate evaluation means 28 evaluates the first to Ncth candidates, and thereby determines several candidates as higher administrative units for lower addresses. The recursive matching evaluation means 29 calculates an average matching distance of all address candidates, and selects an address having the smallest average matching distance therefrom. The place name collating unit 27 may collate the extracted feature vector with the feature vector of the place name in the place name dictionary, and a feature extracting unit (not shown) for extracting a feature vector in the place name image. Collation means for finding all place name candidates.

  In place name recognition, since the first candidate is usually not a correct place name, the multi-candidate recursive recognition method of the present invention significantly improves the recognition accuracy of the system. Furthermore, in the evaluation of place names, it is not necessary to examine all candidates, and the recognition efficiency of the system is improved.

  FIG. 8 shows as an example a flowchart of processing performed by the collective place name recognition unit 4 according to the embodiment of the present invention.

Assume that there are N-level key characters in the input address image. In step SF1, address recognition starts from level 1 (i = 1). In step SF2, i-level place name matching is performed. In the place name matching, if a higher-order address exists, the features extracted from the place-name image corresponding to the i-level place name are separated by the key characters of the i-level place name and Matched with the place name included in the address. First to Nc-th place name candidates are generated and arranged in ascending order based on the collation distance. In step SF3, the candidate serial number counter k is set so that counting starts from 2. In step SF4, the ratio of the difference between the collation distance of the kth place name candidate and the collation distance of the first candidate to the collation distance of the first candidate is calculated. In step SF5, it is determined whether or not the ratio is greater than a predetermined threshold value Th _dis or whether or not the candidate is the last candidate. In particular,
[Dis (cand1 _k ) − Dis (cand1 ₁ )] / Dis (cand1 ₁ ) <Th _dis
Or if it is determined that the current candidate is the last candidate, the process proceeds to step SF7, and the 1st to kth candidates are ranked higher for use in the next level place name recognition. As an administrative unit. Here, Dis (Cand1 _k ) represents the collation distance of the kth candidate in the first-level place name recognition, and the collation distance represents the dissimilarity between the image and the candidate divided by the key characters. The smaller the value of the collation distance, the smaller the difference between the image and the candidate determined at that time. That is, both are closer. The calculation of the verification distance will be described later. Th _dis is a positive constant and is a predetermined threshold value. If the value of the collation distance exceeds this threshold, it means that all the images determined at that time are significantly different from the candidates, and no further comparison is necessary. Furthermore, since the place name candidates are arranged based on the collation distance, if the ratio calculated for the current candidate exceeds Th _dis , the ratio of the subsequent candidates will necessarily exceed the threshold value. There is no need to evaluate above. On the other hand, if it is determined in step SF5 that the ratio does not exceed the predetermined threshold Th _dis and it is determined that the candidate is not the last candidate, the value of k is incremented in step SF6, and the process returns to step SF4.

  After all the place name candidates have been evaluated, the 1st to kth place name candidates are held as upper addresses matching lower place names. It can be seen that the number k of place name candidates held in the address verification at each level changes dynamically and can be different at each level. If the collation distance of the place name candidate is significantly different from the collation distance of the first candidate, it is unlikely that the place name candidate is a correct place name, and the final k is a relatively small number. In other cases, the number is relatively large.

であり、第１のキー文字が「市」、第２のキー文字が「区」である場合、第１のキー文字および第２のキー文字の間の画像と、「北京」市に包含される全ての名前とが可能性のある住所の一つとして照合される。さらに、他の可能性のある住所として、第１のキー文字および第２のキー文字の間の画像と、
［文字９］

市に包含される全ての名前とが照合される。そして、１からＮｃ番目までの可能性のある候補が出力される。 In step SF8, it is determined whether or not all N place name levels have been recognized. If all levels are not recognized, the value of i is incremented in step SF9, and then the next level place name verification is performed. For example, in the second level place name verification, the pattern between the first and second key characters is included by the upper place name within the range of the higher administrative unit output from the first level place name candidates, and It is necessary to collate with the place name of the administrative unit delimited by the second key character. For example, if the first level place name candidates are “Beijing” and [Character 9]

And the first key character is “city” and the second key character is “gu”, and the image between the first key character and the second key character is included in the “Beijing” city. All names are matched as one possible address. In addition, as other possible addresses, images between the first key character and the second key character,
[Character 9]

All names included in the city are matched. Then, the 1st to Ncth possible candidates are output.

  When place name collation at various levels is completed, that is, when it is determined to be “Yes” in SF8 (the place name of the last level has no lower place names, and therefore place name candidates are evaluated. There is no need, and there is no need to output the place name candidates as upper place names including other place names.) In SF10, the collation distances of all place names in each address candidate are cumulatively added. In SF11, the address candidates are ranked according to the cumulatively added collation distance obtained in SF10, and the address having the smallest cumulatively added collation distance is output as a recognition result.

  The process described above is a recursive process. That is, i-level place name matching must be performed for all k candidates obtained as higher addresses from the i-1 level. Similarly, all k place name candidates obtained from the i-level place name matching after step SF7 must be processed as higher-order addresses in the next level place name matching.

上記のとおり、特徴ベクトルｆ_iおよび特徴ベクトルｖ_iの次元はｍである。 The collation distance Dis (Cand _i ) is defined as the Euclidean distance between the feature vector f _i of the i-th place name candidate and the corresponding feature vector v _i extracted from the place name image. That is, the following equation is established.

As described above, the dimension of the feature vector f _i and the feature vector v _i is m.

  FIG. 8 only shows an example of an operation flow for collective place name recognition. In the collective place name recognition process, the matching result in step SF2 is further combined with the single character place name recognition result, and the obtained address candidates can be filtered using place name verification means. Then, the accuracy of place name recognition is improved by performing the processing after step SF3. Similarly, the recognition result of step SF7 can be combined with the single character place name authentication result, and the obtained address candidates can be filtered using place name verification means. The place name verification means 5 and the single character place name authentication means E1 will be described later.

  FIG. 9 exemplarily shows a recursive collective address recognition method according to an embodiment of the present invention.

であると仮定する。この場合、キー文字パスは、「市−区」となる。そして、キー文字抽出手段２を用いて、入力された画像から「市−区」パスに含まれるキー文字を抽出する。 In FIG. 9, the handwritten character image input first is [character 11].

Assume that In this case, the key character path is “city-ku”. Then, the key character extraction means 2 is used to extract key characters included in the “city-ward” path from the input image.

、「北京」、
［文字１３］

のそれぞれと比較され、それぞれ照合距離は、２３０、２４０、３１０となる。したがって、Ｒ１は、「北京」および
［文字１３］

よりも
［文字９］

である可能性が高く、かつ、
［文字１３］

よりも「北京」である可能性が高いことになる。 Next, the place name areas R1 and R2 are cut out by extracting an image delimited by the first key characters. As shown in FIG. 8, the place name region R1 is collectively recognized in step SF2, and Nc (three in this example) candidate SR1 are output. Further, the collation distance is output for each candidate. The collation distance indicates the dissimilarity between the place name region R1 and the place name in the place name dictionary. The place name area R1 is
[Character 9]

,"Beijing",
[Character 13]

The comparison distances are 230, 240, and 310, respectively. Therefore, R1 is "Beijing" and [letter 13]

Than [character 9]

Is likely, and
[Character 13]

More likely to be “Beijing”.

と
［文字９］

との照合距離の間の差の
［文字９］

の照合距離に対する比は、（３１０−２３０）／２３０であり、比較的大きい。他方、「北京」と
［文字９］

との照合距離の間の差の
［文字９］

の照合距離に対する比は、（２４０−２３０）／２３０であり、比較的小さい。この場合、
［文字１３］

が、Ｒ１の実際の地名である可能性は低いため、ＳＲ２で除外される。具体的には、あらかじめ定められた閾値が０．２５であるとすれば、（３１０−２３０）／２３０＞０．２５、かつ、（２４０−２３０）／２３０＜０．２５であるため、
［文字１３］

は除外され、
［文字９］

および「北京」はそのまま上位行政単位として使用される。 Next, it is evaluated whether or not all Nc candidates can be regarded as higher administrative units with respect to the next level place name (for example, the second level place name).
[Character 13]

And [character 9]

[Character 9] of the difference between the reference distance and

Is a relatively large ratio of (310-230) / 230. On the other hand, "Beijing" and [letter 9]

[Character 9] of the difference between the reference distance and

The ratio to the collation distance is (240−230) / 230, which is relatively small. in this case,
[Character 13]

However, since it is unlikely that it is the actual place name of R1, it is excluded by SR2. Specifically, if the predetermined threshold is 0.25, (310-230) / 230> 0.25 and (240-230) / 230 <0.25,
[Character 13]

Are excluded,
[Character 9]

And “Beijing” is used as a higher administrative unit as it is.

であれば、地名辞書にもとづいて
［文字９］

に包含されるすべての地区の実際の地名パターンを地名領域Ｒ２中の地名と比較し、Ｎｃ個の候補を出力する。これらの候補およびその照合距離をＳＲ３に示す。前のレベルが「北京」であれば、地名辞書にもとづいて「北京」に包含されるすべての地区の実際の地名パターンを地名領域Ｒ２中の地名と比較し、Ｎｃ個の候補を出力する。これらの候補およびその照合距離をＳＲ４に示す。 Next, the above steps are repeated to perform second level place name recognition. The previous level is [Character 9]

Then, based on the place name dictionary [Character 9]

Are compared with the place names in the place name region R2, and Nc candidates are output. These candidates and their collation distances are shown in SR3. If the previous level is “Beijing”, the actual place name patterns of all the areas included in “Beijing” are compared with the place names in the place name region R2 based on the place name dictionary, and Nc candidates are output. These candidates and their collation distances are shown in SR4.

＋
［文字１４］

）をそれ以外の他の候補パスと比較した場合、平均照合距離の平均値が最短となるのは「北京」＋
［文字１０］

であり、その平均照合距離は（２４０＋１６０）／２＝２００となる。したがって、
［文字１１］

が認識結果として出力される。 As a result, all candidate paths (eg,
[Character 9]

+
[Character 14]

) Is compared to other candidate paths, the average average collation distance is “Beijing” +
[Character 10]

The average collation distance is (240 + 160) / 2 = 200. Therefore,
[Character 11]

Is output as a recognition result.

  FIG. 10 is a block diagram showing single character place name recognition means according to the embodiment of the present invention.

  As described above, the single character place name recognition unit and the place name verification unit can filter a plurality of possible place name candidates acquired by batch place name recognition. In addition, the single character place name recognition means and place name verification means determine the result (if any) closer to the shortest average collation distance in the evaluation result (ie, place name path) of the collation evaluation means of the collective place name recognition means 4 can do.

  As shown in FIG. 10, the single character place name recognition unit E1 includes a feature extraction unit 61, a dictionary collation unit 62, and a single character dictionary unit 59.

  The feature extracting means 61 extracts the features of the single character area included in the place name area extracted by the place name area extracting means 3 and outputted by the single character area detecting means 1 one by one. Next, the dictionary collating unit 62 collates each feature of the single character area obtained by the feature extracting unit 61 with the feature of the single-character Chinese word stored in the single character dictionary unit 59. Then, the first to Nth candidates for each single character area are output.

  FIG. 11 is a block diagram showing the place name verification means 5 according to the embodiment of the present invention.

  Based on the place name recognition result of the single character place name recognition means E1 and the place name recognition result obtained by the collective place name recognition means 4, the place name verification means 5 performs these verifications and combinations. FIG. 11 is a block diagram showing the place name verification means 5 according to the embodiment of the present invention. As shown in FIG. 11, the place name verification unit 5 includes an edit distance calculation unit 55 and a batch recognition / individual character recognition combination unit 58.

  The edit distance calculation means 55 edits the place name result obtained by the single character place name recognition means E1 and all place names that match the definition of the key character and are included in the address recognized from the place name of the previous level. Calculate the distance. The batch recognition / individual character recognition combination means 58 uses the edit distance calculation means 55 to combine the result obtained by the batch place name recognition means 4 and the result obtained by the single character place name recognition means E1. In accordance with the candidate sequence or candidate string obtained by the collective place name recognizing means 4, the collective recognition / individual character recognition combination means 58 calculates the overall rank of each place name and finally outputs the first place name. Note that the edit distance is a numerical value indicating how different two character strings are different. For example, the edit distance between “Tokyo” and “Higashizukato” is the difference between “Kyo” and “Bunch”. Because it is only, it becomes “1”. The edit distance is also called Levenshtein distance. In this case as well, the smaller they are, the closer they are.

  The specific processing of place name verification is as follows.

  First, the edit distance calculating means 55 uses the dynamic programming (see Non-Patent Documents 3 and 4) to obtain a place name area that is a reasonable place name separated by key characters and obtained at a higher level. The edit distance defined by the equation (1) (defined below) between the included place name and the result W obtained by the single character place name recognition means E1 is calculated. The edit distance means conversion from a recognition result of a single character part to an operand of a reasonable place name by an operation such as insertion, replacement, or deletion (the operand will be described later). Therefore, when the rational place name editing distance is small, the difference between the place name and the individual character recognition result is very small. On the other hand, when the rational place name editing distance is large, the difference between the place name and the individual character recognition result is large. .

  In the present invention, the edit distance is defined as follows.

式（１）において、ｋは入力された地名画像中の単一文字領域の総数であり、ｗは当該地名画像中の各単一文字領域の認識候補の数である。 S is an n-dimensional vector, indicating a place name including n characters, W is an N × m matrix, and W _ij is the i-th single character region or j-th connected element in the inputted place name image. Let W _i denote the recognition result of the i-th single character region or coupling element, that is, the m-dimensional recognition candidate list. The edit distance between S and W is the smallest cost when W is converted to S, considering that candidates are adjusted by insertion, deletion, replacement, or the like. The cost between S _k and W _i is calculated as follows:

In Equation (1), k is the total number of single character regions in the input place name image, and w is the number of recognition candidates for each single character region in the place name image.

Therefore, when the character S _{k of the} place name S to be collated exists in the recognition candidate W _i of the coupling element, the higher the rank in which the character is positioned in the candidate column W _i (that is, the collation distance is W _ij If there is a small j satisfying = S _k ), the matching cost is reduced and the possibility that W _i is S _k increases. Conversely, the lower the order in which the character is positioned in the candidate string W _i (that is, if there is a large j that satisfies the collation distance W _ij = S _k ), the collation cost increases and W _i becomes smaller. The possibility of being S _k is reduced.

  The collective recognition / individual character recognition combination means 58 uses the edit distance calculation means 55 to combine the collective recognition result and the single character place name recognition result in the following steps.

(1) The weight value of each place name A _i included in the upper level place name and separated by the current key character is calculated. The weight value is calculated by the following equation (2).
Weight (A _i ) = (1 − t1) × Rank_ED (A _i ) + t1 × i (2)
In equation (2), A _i is the i-th place name candidate acquired by the collective place name recognition means 4, ED (A _i ) is the edit distance of A _i , and Rank_ED (A _i ) is individually independent of A _i It is the rank of the edit distance with the character recognition candidate matrix, t1 is equal to round (ED (A _i )) / k, and round (ED (A _i )) indicates that ED (A _i ) is rounded off. .

(2) is included in the higher level of place names, and place names ranks weighting values of all place names A _i, delimited by the key character at the current time in ascending order, the weighting value is recognized at present the smallest A _i And

When the edit distance of A _i is extremely small, the reliability of the individual character recognition result is high, and a larger weighting value (1-t1) is assigned to Rank_ED (A _i ). When the editing distance of A _i is remarkably large, t1 becomes remarkably large, and the candidate rank i obtained from the collective recognition unit plays an important role in the weighting calculation. That is, if the editing distance of A _i is sufficiently small, the result of individual character recognition is reliable, while if the editing distance of all reasonable addresses is sufficiently large, the result of individual character recognition is not very reliable. In this case, it is necessary to output the recognition result obtained from the collective place name recognition means as a combined place name. Thus, since the recognition result obtained from the collective place name recognition means and the result of the individual character recognition are combined, the accuracy of the system is remarkably improved.

  FIG. 12 shows an example of address verification according to the present invention.

である。ここで、第１レベルの地名
［文字１６］

がすでに認識され、キー文字
［文字１７］

もすでに認識されていると仮定する。（Ａ）に、「二盛一」の画像領域から単一文字領域検出手段１により得られた単一文字領域の個別文字認識結果の例を示す。Ｗ₁、Ｗ₂、およびＷ₃は、それぞれ、単一文字領域「二」、「盛」、「一」に対応する認識候補列を示す。候補の数は３と仮定する。（Ｂ）に、「二盛一」の画像領域を一括認識することによって得られた地名候補の例を示す。ここで、
［文字１６］

市には、
［文字１７］

がつく地名は３つしかなく、それらはそれぞれ「二盛」、「二盛一」、「二旺一」であると仮定する。 In FIG. 12, the actual address of the input address image shown in (A) and (B) is
[Character 15]

It is. Here, the first level place name [character 16]

Is already recognized and the key character [character 17]

Is also already recognized. (A) shows an example of an individual character recognition result of a single character area obtained by the single character area detection means 1 from an image area of “Niimoriichi”. W ₁ , W ₂ , and W ₃ indicate recognition candidate strings corresponding to single character areas “two”, “high”, and “one”, respectively. Assume that the number of candidates is three. (B) shows an example of a place name candidate obtained by collectively recognizing the image area of “Niimoriichi”. here,
[Character 16]

The city has
[Character 17]

It is assumed that there are only three place names that are ticked, and that they are “Nisemori”, “Nishimori”, and “Ninoichi”, respectively.

市の領域に包含され、かつ、
［文字１７］

によって区切られるすべての地名（すなわち、「二盛」、「二盛一」、「二旺一」）との間の編集距離を計算する。たとえば、「二盛」とＷとの間の最小編集距離は、ダイナミック・プログラミングによって、「二」をＷ₁と照合し、「盛」をＷ₂と照合し、無効文字（ＮＵＬＬとラベリングする）を追加してこれをＷ₃と照合することで計算される。この場合の照合コストは以下のようになる。 By the place name verification, the matrix W = {W ₁ , W ₂ , W ₃ } in (A) and
[Character 16]

Included in the city territory, and
[Character 17]

The edit distance between all place names (ie, “Niimori”, “Niimoriichi”, “Ninoichi”) delimited by is calculated. For example, the minimum edit distance between the W and the "two-Sheng", where the dynamic programming, to match the "two" W ₁ and, the "Sheng" against the W _2, (to NULL and labeling) invalid character This is calculated by matching the W ₃ by adding. The verification cost in this case is as follows.

Cost ("Two", W ₁ ) = (2 − 1) / (3 × 3) = 1/9;
Cost ("Sai", W ₂ ) = (1 − 1) / (3 × 3) = 0;
Cost (NULL, W ₃ ) = 1.

  Therefore, the edit distance between W and “Niimori” is 1/9 + 0 + 1 = 10/9.

Similarly, the edit distance between W and “Nisseiichi” is calculated as follows.
Cost ("Two", W ₁ ) = (2 − 1) / (3 × 3) = 1/9;
Cost ("Sai", W ₂ ) = (1 − 1) / (3 × 3) = 0;
Cost ("One", W ₃ ) = (2 − 1) / (3 × 3) = 1/9.
Therefore, the edit distance between W and “Futakuichi” is 1/9 + 0 + 1/9 = 2/9.

Similarly, the edit distance between W and “Two-one” is calculated as follows.
Cost ("Two", W ₁ ) = (2 − 1) / (3 × 3) = 1/9;
Cost ("旺", W ₂ ) = 1;
Cost ("One", W ₃ ) = (2 − 1) / (3 × 3) = 1/9.
Therefore, the edit distance between W and “Two-by-one” is 1/9 + 1 + 1/9 = 11/9.

市に包含されるすべての地名
［文字１７］

の編集距離の順位付けシーケンスは、「二盛一」、「二盛」、「二旺一」となる。 Therefore,
[Character 16]

All place names included in the city [Character 17]

The editing distance ranking sequence is “Niimoriichi”, “Niimori”, “Ninoichi”.

  At this time, since the edit distance 1/9 of “Nisseiichi” is the shortest, it is verified by the verification rule (1) that “Nisseiichi” is a correct place name.

  FIG. 13 is a block diagram showing reject means according to the embodiment of the present invention.

  As shown in FIG. 13, the rejection unit 6 includes a feature extraction unit 63, a rejection determination unit 64, and a rejection learning unit 65.

Specifically, the feature extraction unit 63 extracts features {v ₁ , v ₂ , v ₃ ,..., V _N } from various levels of the place name image region. Here, v _i (0 <i <N + 1) is a feature extracted from an i-level place name, and N is the total number of place name levels. The reject determining means 64 substitutes the feature obtained by the feature extracting means 63 into the mapping function f obtained by the reject learning means 65, and obtains the function f (v ₁ , v ₂ , v ₃ ,..., V _N ). Calculate the function value. If this value is greater than zero, the recognition result is a correct address. If this value is less than or equal to zero, the recognition result is rejected.

  The reject learning means 65 performs learning using a learning set collected from actual address images. The collation distance for each level of place name recognition for each address sample is collected as a feature. If the address sample is recognized correctly, the address image is labeled for approval or +1. In other cases, the address image is rejected or -1 labeled.

  The above processing is based on Non-Patent Document 5. The reason why the mapping function (hereinafter referred to as a classifier) is constructed is to automatically perform classification based on features in the learning data. This classifier is used as a reject rule to evaluate a new address and decide whether to reject or accept the recognition result depending on whether the function output is positive (> 0) or negative (<0). . This function can automatically build reject rules that adaptively adapt to the quality of the address, increasing the reliability of address recognition.

  FIG. 14 is a detailed block diagram showing the address recognition apparatus according to the embodiment of the present invention.

  As shown in FIG. 14, the address recognition apparatus according to the present invention includes a single character area detecting means 1, a key character candidate extracting means 20, a key character path selecting means 15, a place name area extracting means 3, a single character place name. Recognizing means E1, collective place name recognizing means 4, place name verifying means 5, individual character recognizing means 69, and rejecting means 6 are provided. The single character area detecting means 1 includes a combination element labeling means 34, a single character area candidate determining means 35, an average character size calculating means 36, a pixel projecting means 37, a temporary separation point candidate extracting means 38, Separation point candidate determination means 39 is provided.

  The coupling element labeling means 34 labels the coupling element from the input binary image. The single character area candidate determining means 35 integrates each combination element based on the size of the labeled element, the position of the labeled element, the distance from the adjacent element, etc., and finally determines the single character area candidate. .

  The average character size calculation means 36 removes a pattern whose size does not match that of other coupling elements, and calculates the average character size.

  The pixel projecting unit 37 projects the input address image vertically (when the image is written horizontally) or horizontally (when the image is written vertically). With vertical (or horizontal) image projection, the number of black pixels at each point of the image in the horizontal (or vertical) direction is obtained.

  Temporary separation point candidate extraction means 38 and temporary separation point candidate determination means 39 are used to separate the new combined elements output by single character area determination means 35 into actual character areas. That is, both are used to separate the combined characters into individual single characters.

  Temporary separation point candidate determination means 39 determines whether or not the temporary separation point obtained by temporary separation point candidate extraction means 38 is a true separation point of adjacent character regions.

  The key character dictionary 41 records and stores the characteristics of all key characters.

  The key character candidate extraction unit 20 includes a feature extraction unit 42, a dictionary collation unit 43, and a key character candidate determination unit 44.

  The feature extraction means 42 extracts the features of the areas separated by the single character area detection means 1. The dictionary collating unit 43 collates the feature extracted by the feature extracting unit 42 with the feature stored in the key character dictionary 41 and outputs the 1st to Nth character candidates. Next, the key character candidate determination means 44 determines whether or not the key character candidate is a true key character. When the ratio of the collation distance of the key character candidate to the collation distance of the first character candidate is smaller than a predetermined threshold, the key character candidate is a true key character.

  The address hierarchy structure knowledge unit 24 stores knowledge of the address hierarchy structure. For example, the key character “Province” may include the key character “City”, but the key character “City” cannot include the key character “Province”.

  The key character path selection unit 15 includes a key character path determination unit 21 and a key character path filter unit 22.

  The key character path determining means 21 refers to the address hierarchical structure knowledge unit 24 to determine a reasonable key character path. The key character path filter means 22 stores a plurality of key character paths, calculates an average collation distance between key characters in each path, and finally outputs K main paths. Here, K is a predetermined positive constant.

  The place name area extraction means 3 extracts place name areas delimited by key characters in a specific key character path.

  The place name dictionary unit 54 stores a feature vector of each name in the address.

  The collective place name recognition unit 4 sequentially recognizes the place names extracted by the place name extraction unit 3. The collective place name recognition means 4 includes place name collating means 27, place name candidate evaluating means 28, and recursive collating and evaluating means 29.

  The place name collation means 27 extracts features from images corresponding to place names at various levels, collates with the place name features stored in the place name dictionary unit 54, and thereby outputs the 1st to Ncth candidates. The place name candidate evaluation means 28 evaluates the 1st to Ncth candidates, and determines some candidates that are upper administrative units for lower addresses. The recursive matching evaluation means 29 calculates an average distance of all address candidates, and selects an address having the smallest average matching distance as a final recognition result.

  The single character dictionary unit 59 stores a feature vector of each character in the address.

  The single character place name recognition unit E1 sequentially recognizes the single character regions output from the single character region detection unit 1. These single character areas are included in the place name area extracted by the place name area extracting means 3. The single character place name recognition unit E1 includes a feature extraction unit 61 and a dictionary collation unit 62.

  The feature extraction means 61 extracts the feature vector of the single character area obtained by the single character area detection means 1. These single character areas are included in the place name area extracted by the place name area extracting means 3. Next, the dictionary collating unit 62 collates the feature vector obtained by the feature extracting unit 61 with the feature vector stored in the single character dictionary unit 59, and outputs the 1st to Nth candidates of each single character region. .

  The place name verification means 5 verifies the place name recognition result obtained by the single character place name recognition means E1 and the place name recognition result acquired by the collective place name recognition means 4 and combines them. The place name verification means 68 includes an edit distance calculation means 55 and a collective recognition / individual character recognition combination means 58.

  The edit distance calculation means 55 calculates the place name result obtained by the single character place name recognition means E1 and all place names that match the definition of the key character and are included in the address recognized from the place name of the previous level. Calculate the edit distance between. The batch recognition / individual character recognition combination means 58 combines the result obtained by the batch place name recognition means 4 and the result obtained by the single character place name recognition means E1. The collective recognition / individual character recognition combination means 58 ranks the calculated names of places and finally outputs the first place name.

  The individual character recognition means 69 recognizes a pattern that is neither a place name nor a key character. The individual character recognition unit 69 includes a feature extraction unit 61 and a dictionary collation unit 62.

  The reject learning means 65 learns a support vector machine classifier based on a group of addresses. Note that the support vector machine is an excellent method for classification into two classes, and when there are four levels (province-city-district-town) of the address, there are four words in total. Learn in advance the sequence of distance values for sequences and those that are not. Next, a sequence of distances of four words is calculated from the input character string of the four layers, and it is determined whether it belongs to the correct side or not to be learned.

  The reject unit 6 determines whether or not to approve the recognition result, and includes a feature extraction unit 63 and a reject determination unit 64. The feature extraction unit 63 extracts feature vectors. Based on the support vector machine classifier output from the reject learning unit 65, the reject determining unit 64 determines whether to recognize the recognition result and outputs it.

  Finally, if the result is not rejected, the reject means 6 outputs the recognition result. If the result is rejected, the reject means 6 outputs an error.

  The foregoing descriptions of the embodiments of the present invention have been given by way of illustration, and are not intended to preclude other examples, but are to clarify the present application. It is not limited. It will be apparent to those skilled in the art that many modifications and variations are possible. The selection and description of the embodiments are intended to better illustrate the actual principles and applications of the present invention, so that others having skill in this field can implement various implementations of the present invention. It is intended to be able to understand the modifications of the present invention for example applications and special applications. For example, the present invention can be applied to recognition of handwritten place names written in Japanese or Korean. Furthermore, although the present invention uses the collation distance to determine the similarity, the similarity can be determined using a method other than the collation distance. For example, a cosine gradient between feature vectors, a block distance between feature vectors, or the like can be used. Therefore, the collation distance in the present invention must be broadly understood as an expression of a definite amount of similarity that can be recalled by those skilled in the art.

(Supplementary note 1) A key character extraction program for executing extraction of a key character from a plurality of single character regions for each of the plurality of single character regions,
A feature vector extraction procedure for extracting a feature vector from an image corresponding to the single character region;
A collation procedure for collating the feature vector extracted by the feature vector extraction procedure with a key character feature vector in a key character dictionary storing all possible key characters, and searching for all possible key character candidates;
A key character path determination procedure for determining a valid key character path when recognizing a plurality of single character regions each having one or more key character candidates searched by the matching procedure;
Key character extraction program characterized by causing a computer to execute

(Supplementary note 2) The key character extraction program according to supplementary note 1, wherein the key character dictionary stores a set of characters representing an administrative district level.

(Additional remark 3) The said collation procedure is a key character which can be included in the said key character dictionary, Comprising: The feature vector with which the collation distance between the feature vectors extracted by the said feature vector extraction procedure is smaller than a predetermined threshold value The key character extraction program according to appendix 1, wherein a possible key character corresponding to is determined as a key character candidate.

(Supplementary note 4) A key character extraction method for executing extraction of key characters from a plurality of single character areas for each of the plurality of single character areas,
A feature vector extraction step of extracting a feature vector from an image corresponding to the single character region;
Collating the feature vector extracted by the feature vector extracting step with the feature vector of the key character in the key character dictionary storing all possible key characters, and searching for all possible key character candidates;
A key character path determining step for determining a valid key character path when recognizing a plurality of single character regions each having one or more key character candidates searched by the matching step;
Key character extraction method characterized by including

(Additional remark 5) The said key character dictionary memorize | stored the group of the character showing the administrative district level, The key character extraction method of Additional remark 4 characterized by the above-mentioned.

(Additional remark 6) The said collation process is a key character which can be contained in the said key character dictionary, Comprising: The collation distance between the feature vectors extracted by the said feature vector extraction process is smaller than a predetermined threshold value The key character extraction method according to appendix 4, wherein a possible key character corresponding to is determined as a key character candidate.

(Additional remark 7) The key character candidate with respect to a single character area is determined by the said collation process, and the key character candidate which became the 1st place by ranking a key character candidate based on collation distance is 1st. When the key character candidate is selected, a ratio of the difference between the collation distance of the key character candidate and the collation distance of the first key character candidate to the collation distance of the first key character is calculated. The key character extraction method according to appendix 6, further comprising a key character candidate determination step of determining a key character candidate having a collation distance that becomes smaller as a key character candidate result.

(Supplementary Note 8) A key character path for filtering a part of valid key character paths from a plurality of valid key character paths according to a predetermined rule when a plurality of valid key character paths are determined in the key character path determining step. The key character extraction method according to any one of supplementary notes 4 to 7, further comprising a filtering step.

(Supplementary Note 9) The predetermined rule calculates an average collation distance of key character candidates included in a valid key character path, ranks all key character paths so that the average collation distance is in ascending order, The key character extraction method according to appendix 8, wherein the rule is to select only a predetermined number of key character paths.

(Supplementary note 10) A key character extraction device for extracting a key character from a plurality of single character regions,
Feature vector extraction means for extracting a feature vector from an image corresponding to each of the single character regions;
A key character dictionary that stores only all possible key characters;
Collating means for collating the feature vector extracted by the feature vector extracting means with the feature vector of possible key characters stored in the key character dictionary unit, and searching for all possible key character candidates;
A key character path determining unit that determines a valid key character path when recognizing a plurality of single character regions each having one or more key character candidates searched by the matching unit;
A key character extraction device comprising:

(Additional remark 11) The key character extraction device of Additional remark 10 characterized by the possible key character memorize | stored in the said key character dictionary part is a character showing an administrative district level.

(Additional remark 12) The said collation means is a possible key character memorize | stored in the said key character dictionary part, Comprising: The collation distance between the feature vectors extracted by the said feature vector extraction means is smaller than a predetermined threshold value The key character extraction device according to appendix 10, wherein a possible key character corresponding to the feature vector is determined as a key character candidate.

(Additional remark 13) The key character candidate with respect to a single character area is determined by the said collation means, and the key character candidate which became the 1st place by ranking a key character candidate based on collation distance is 1st. When a key character candidate is selected, a ratio of a difference between the collation distance of the key character candidate and the first key character candidate to the collation distance of the first key character is calculated, and the ratio becomes smaller than a predetermined threshold value. 13. The key character extraction device according to appendix 12, further comprising key character candidate determination means for determining a key character candidate that is a collation distance as a key character candidate result.

(Supplementary Note 14) A key character path for filtering a part of valid key character paths from a plurality of valid key character paths according to a predetermined rule when a plurality of valid key character paths are determined by the key character path determining means. 14. The key character extraction device according to any one of appendices 10 to 13, further comprising filtering means.

(Supplementary Note 15) The predetermined rule calculates an average collation distance of key character candidates included in a valid key character path, ranks all the key character paths so that the average collation distance is in ascending order, 15. The key character extraction device according to appendix 14, wherein the rule selects only a predetermined number of key character paths.

(Supplementary Note 16) A collective place name recognition program for causing a computer to execute a collective place name recognition procedure for recognizing place names of place name images including place names at a certain administrative level,
The collective place name recognition procedure includes:
An extraction procedure for extracting a feature vector from the place name image;
A collation procedure for collating the feature vectors extracted by the extraction procedure with the feature vectors of the place names stored in the place name dictionary and searching for all possible place name candidates;
To the computer,
Of the place names stored in the place name dictionary, a place name corresponding to a feature vector whose collation distance with the feature vector extracted by the extraction procedure is smaller than a predetermined threshold is determined as a possible place name candidate,
Further, an individual character recognition procedure for recognizing character candidates for characters included in each single character region of the place name image,
Based on the recognition result obtained by the individual character recognition procedure, a place name verification procedure for filtering place name candidates determined by the collective place name recognition procedure;
Collective place name recognition program characterized by causing computer to execute.

(Supplementary Note 17) The batch place name recognition procedure further causes the computer to execute a place name candidate evaluation procedure for filtering a plurality of possible place name candidates according to a predetermined rule when a plurality of possible place name candidates are obtained by the matching procedure. The collective place name recognition program according to supplementary note 16, characterized by:

(Supplementary Note 18) The predetermined rule is that a plurality of possible place name candidates are ranked based on the collation distance, and a place name candidate that is ranked first by the ranking is set as the first place name candidate. Calculate the ratio of the difference between the collation distance of the candidate and the collation distance of the first place name candidate to the collation distance of the first place name candidate, and select only the place name candidates whose collation distance is smaller than a predetermined threshold. 18. The collective place name recognition program according to appendix 17, wherein

(Supplementary note 19) A collective place name recognition method including a collective place name recognition process for recognizing place names of place name images including place names at a certain administrative level,
The collective place name recognition process includes:
An extraction step of extracting a feature vector from the place name image;
A collation step of collating the feature vector extracted by the extraction step with the feature vector of the place name stored in the place name dictionary and searching for all possible place name candidates;
Including
Of the place names stored in the place name dictionary, a place name corresponding to a feature vector whose collation distance with the feature vector extracted by the extraction step is smaller than a predetermined threshold is determined as a possible place name candidate,
Further, an individual character recognition step for recognizing character candidates for characters included in each single character region of the place name image,
Based on the recognition result obtained by the individual character recognition step, a place name verification step for filtering place name candidates determined by the collective place name recognition step;
Collective place name recognition method characterized by including.

(Supplementary note 20) The collective place name recognition step includes a place name candidate evaluation step of filtering a plurality of possible place name candidates according to a predetermined rule when a plurality of possible place name candidates are obtained by the matching step. The collective place name recognition method according to appendix 19.

(Supplementary Note 21) The predetermined rule is that a plurality of possible place name candidates are ranked based on a collation distance, and a place name candidate that is ranked first by the ranking is set as a first place name candidate. Calculate the ratio of the difference between the collation distance of the candidate and the collation distance of the first place name candidate to the collation distance of the first place name candidate, and select only the place name candidates whose collation distance is smaller than a predetermined threshold. The collective place name recognition method according to appendix 20, wherein the rule is a rule for

(Supplementary note 22) The place name verification step includes a step of calculating an edit distance between each of the place name candidates determined by the collective place name recognition step and a recognition result obtained by the individual character recognition step, and each edit A step of calculating a rank of the distance; a step of calculating a weight based on the rank of the edit distance; and a step of filtering the place name candidates obtained by the collective place name recognition step based on the calculated weight. The collective place name recognition method according to appendix 19, 20 or 21, characterized in that:

(Supplementary Note 23) In the collective place name recognition step, when one or more place name candidates are recognized for each of a plurality of place name images in which the place name administrative level is in descending order, addresses from the highest level to the lowest level Additional description 19, 20 or 21 further comprising a collation evaluation step of calculating an average collation distance of each path formed by the candidates and selecting only a path having the smallest average collation distance as an address path candidate. Collective place name recognition method of description.

(Supplementary Note 24) In the collective place name recognition process, when one or more place name candidates are recognized for each of a plurality of place name images whose place names are in descending order, addresses from the highest level to the lowest level The collective place name according to appendix 19, further comprising a collation evaluation step of calculating an average collation distance of each of the paths formed by the candidates and selecting only a path having the smallest average collation distance as an address path candidate. Recognition method.

(Supplementary Note 25) When there are a plurality of paths having an average collation distance close to the minimum average collation distance, a path individual character recognition step of executing individual character recognition for a place name image having a plurality of possible place name candidates in the plurality of paths; A path place name verification step of filtering possible place name candidates of the place name image based on the recognition result obtained in the path individual character recognition step, and the place name verification step includes a place name candidate of the place name image. A step of calculating an edit distance between each and a recognition result obtained by the individual character recognition step, a step of calculating a rank of each edit distance, and a weight based on the rank of the edit distance 25. The collective place name recognition method according to appendix 24, comprising a step and a step of determining a place name based on the calculated weight.

(Supplementary note 26) Further includes a place name rejecting step for approving or rejecting the address path candidate selected in the collation evaluation step, wherein the place name rejecting step includes a step of extracting features from the place name areas of each level, and a place name of each level Appendix 25 characterized by including a step of analyzing a feature extracted from a region using a mapping function obtained by reject learning, and a step of approving or rejecting an address path candidate based on a result of the analysis Collective place name recognition method of description.

(Supplementary note 27) A collective place name recognition device,
Extracting means for extracting feature vectors from the input place name image;
A place name dictionary section for storing place names;
The place name feature vector stored in the place name dictionary unit is matched with the feature vector extracted by the extracting means to obtain all possible place name candidates and matching distances, and among the place names stored in the place name dictionary, Collating means for determining a place name corresponding to a feature vector whose matching distance between the feature vector extracted by the extracting means is smaller than a predetermined threshold as possible place name candidates;
Evaluation of multiple address path candidates when multiple place name images containing administrative level place names arranged in descending order are input and multiple place name paths containing administrative level place names are formed by recognition of the place name images. And a collation evaluation means for making a decision,
A collective place name recognition apparatus characterized by further comprising:

(Supplementary note 28) The supplementary note 27, further comprising place name candidate evaluation means for filtering a plurality of possible place name candidates according to a predetermined rule when a plurality of possible place name candidates are obtained by the collating means. Collective place name recognition device.

(Supplementary note 29) When the predetermined rule ranks a plurality of possible place name candidates based on the collation distance and sets the first place name candidate as the first place name candidate as the first place name candidate by the ranking, Calculate the ratio of the difference between the collation distance of the candidate and the collation distance of the first place name candidate to the collation distance of the first place name candidate, and select only the place name candidates whose collation distance is smaller than a predetermined threshold. 29. The collective place name recognition apparatus according to appendix 28, wherein

(Additional remark 30) The said collation evaluation means calculates an average collation distance about each address path, and determines an address path with the smallest average collation distance as an address path candidate, The additional remark 27, 28 or 29 characterized by the above-mentioned Collective place name recognition device.

(Supplementary note 31) The collective place name recognition apparatus according to supplementary note 30, further comprising reject means for approving or rejecting the address path candidate determined by the collation evaluation means.

(Additional remark 32) The rejection means uses the feature extraction means for extracting features from each name area, reject learning means for executing a rejection learning to obtain a mapping function, and extraction of place name areas at each level using the mapping function. 32. The collective place name recognition apparatus according to supplementary note 31, further comprising: a rejection determination unit that determines whether to approve or reject the address path candidate based on the made feature.

(Supplementary Note 33) When there are a plurality of address paths having an average collation distance close to the minimum average collation distance, individual character recognition is performed on a place name image having a plurality of possible place name candidates in the plurality of address paths, Alternatively, when the collating unit determines a plurality of possible place name candidates for the place name image, individual character recognition means for performing individual character recognition on the place name image having a plurality of possible place name candidates; and the individual character And a place name verification means for filtering place name candidates of place name images having a plurality of possible place name candidates based on the recognition result obtained by the recognition means and the determination result by the collation evaluation means. The collective place name recognition apparatus according to any one of Supplementary notes 27 to 32.

(Supplementary Note 34) The place name verification means includes an edit distance calculation means for calculating an edit distance between each place name candidate of a place name image having a plurality of possible place name candidates and a recognition result by the individual character recognition means, (Supplementary note 33) characterized by comprising batch recognition / individual character recognition combination means for determining addresses of place name images having a plurality of possible place name candidates based on the edit distance calculated by the edit distance calculation means. Collective place name recognition device of description.

(Supplementary Note 35) Character separation step for separating single character regions, key character extraction step for extracting key characters for each single character region separated by the character separation step, and extraction of key characters by the key character extraction step An address recognition method including a collective place name recognition step for collectively recognizing place names based on a result, an address determination step for determining an address based on a place name recognition result by the collective place name recognition step,
The key character extraction step includes:
A feature vector extraction step of extracting a feature vector from an image corresponding to the single character region;
A collation step of collating the feature vector extracted by the feature vector extraction step with a feature vector of possible key characters in a key character dictionary storing all possible key characters, and searching for all possible key character candidates. When,
A key character path determining step for determining a valid key character path when recognizing a plurality of single character regions each having one or more key character candidates searched by the matching step;
A method for recognizing an address, comprising:

(Additional remark 36) The said key character extraction process is a possible key character contained in the said key character dictionary, Comprising: The collation distance between the feature vectors extracted by the said feature vector extraction process is smaller than a predetermined threshold value 36. The address recognition method according to appendix 35, wherein a possible key character corresponding to the feature vector is determined as a key character candidate.

(Additional remark 37) The key character candidate with respect to a single character area is determined by the collation process, and the key character candidate which became the first place by ranking the key character candidates based on the collation distance is the first. When the key character candidate is selected, a ratio of the difference between the collation distance of the key character candidate and the collation distance of the first key character candidate to the collation distance of the first key character is calculated. 36. The address recognition method according to supplementary note 35, further comprising a key character candidate determination step of determining a key character candidate having a matching distance that is smaller as a key character candidate result.

(Supplementary Note 38) In the key character extraction step, when a plurality of valid key character paths are determined in the key character path determination step, some valid key characters are extracted from a plurality of valid key character paths according to a predetermined rule. 38. The address recognition method according to appendix 35, 36 or 37, further comprising a key character path filtering step of filtering the path.

(Supplementary Note 39) The collective place name recognition step includes a collective recognition step of recognizing place names of place name images including place names at a certain administrative level, and the collective recognition step includes an extraction step of extracting a feature vector from the place name images. A collation step of collating the feature vector extracted by the extraction step with the feature vector of the place name stored in the place name dictionary and searching for all possible place name candidates, and among the place names stored in the place name dictionary The address recognition method according to appendix 38, wherein a place name corresponding to a feature vector whose collation distance with the feature vector extracted by the extraction step is smaller than a predetermined threshold is determined as a possible place name candidate. .

(Supplementary Note 40) The batch recognition step further includes a place name candidate evaluation step of filtering a plurality of possible place name candidates according to a predetermined rule when a plurality of possible place name candidates are obtained by the matching step, The rule ranks a plurality of possible place name candidates based on the collation distance, and sets the first place name candidate as the first place name candidate as the first place name candidate as a result of the ranking. The rule is to calculate a ratio of the difference between the collation distance of the place name candidates and the collation distance of the first place name candidates, and to select only the place name candidates whose collation distance is smaller than a predetermined threshold. The address recognition method according to appendix 39.

(Additional remark 41) Based on the recognition result obtained by the individual character recognition process which recognizes the character candidate of the character contained in each of the single character area | region of the said place name image, and the said individual character recognition process, by the said batch place name recognition process A place name verification step for filtering the determined place name candidates, wherein the place name verification step includes each of the place name candidates determined by the batch place name recognition step, and a recognition result obtained by the individual character recognition step; A step of calculating the edit distance between the steps, a step of calculating the rank of each edit distance, and the rank of the edit distance,
Weight (A _i ) = (1 − t1) × Rank_ED (A _i ) + t1 × i
And a step of filtering the place name candidates obtained by the collective place name recognition process based on the calculated weight, and A _i is the i obtained by the collective place name recognition process. a th place name candidate, ED (a _i) is the edit distance of a _i, a Rank_ED (a _i) is the edit distance of rank, t1 is round (ED (a _i)) equal to / k, round english (us) (ED (A _i )) means a rounding operation of ED (A _i ), k is the total number of single character areas in the place name image, and weight (A _i ) means the weight of the i-th place name candidate. 41. The address recognition method according to appendix 40, wherein

(Additional remark 42) When one or more place name candidates are recognized for each of a plurality of place name images in which the place name administrative level is descending in the collective place name recognition step, addresses from the highest level to the lowest level 42. The address recognition method according to appendix 41, further comprising a collation evaluation step of calculating an average collation distance of each of the paths formed by the candidates and selecting only a path having the smallest average collation distance as an address path candidate. .

(Additional remark 43) It further includes the place name rejection process which approves or rejects the address path candidate selected by the said collation evaluation process, The said place name rejection process extracts the characteristic from the place name area | region of each level, The place name of each level Appendix 42 characterized by including a step of analyzing a feature extracted from a region using a mapping function obtained by reject learning, and a step of approving or rejecting an address path candidate based on a result of the analysis The address recognition method described.

  As described above, the present invention relates to a key character extraction program, a key character extraction device, a key character extraction method, a collective place name recognition program, a collective place name recognition apparatus, and a collective place name recognition method, and is particularly handwritten in an area without a frame. The present invention is useful for a key character extraction program, a key character extraction device, a key character extraction method, a collective place name recognition program, a collective place name recognition device, and a collective place name recognition method that recognize and read an address from a character.

FIG. 1 is a block diagram showing a conventional address recognition device. FIG. 2 is a schematic diagram showing the processing of a collective place name recognition unit that is provided in the conventional address recognition apparatus of FIG. 1 and performs collective place name recognition. FIG. 3 is a block diagram illustrating an address recognition apparatus according to an embodiment of the present invention. FIG. 4 is a block diagram showing an example of the single character area detecting means. FIG. 5 is a block diagram showing key character extraction means according to an embodiment of the present invention. FIG. 6 is a block diagram showing key character candidate extraction means according to the embodiment of the present invention. FIG. 7 is a block diagram showing collective address recognition means according to an embodiment of the present invention. FIG. 8 is a flowchart showing the collective place name recognition means according to the embodiment of the present invention. FIG. 9 is a diagram exemplarily illustrating a collective address recognition method according to an embodiment of the present invention. FIG. 10 is a block diagram showing single character place name recognition means according to an embodiment of the present invention. FIG. 11 is a block diagram showing place name verification means according to the embodiment of the present invention. FIG. 12 is a view exemplarily showing a place name verification method according to the present invention. FIG. 13 is a block diagram showing reject means according to the embodiment of the present invention. FIG. 14 is a detailed block diagram illustrating an address recognition apparatus according to an embodiment of the present invention.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Single character area detection means 2 Key character extraction means 3 Place name area extraction means 4 Collective place name recognition means 5 Place name verification means 6 Reject means 16 Key character path selection means 20 Key character candidate extraction means 21 Key character path determination means 22 Key character path Filter means 24 Address hierarchy structure knowledge part 27 Place name matching means 28 Place name candidate evaluation means 29 Recursive matching evaluation means 34 Joining element labeling means 35 Single character area candidate determination means 36 Average character size calculation means 37 Pixel projection means 38 Temporary separation point Candidate extraction means 39 Temporary separation point candidate determination means 41 Key character dictionary part 42 Feature extraction means 43 Dictionary collation means 44 Key character candidate determination means 54 Place name dictionary part 55 Edit distance calculation means 58 Collective recognition / individual character recognition combination means 59 One-character dictionary part 61 Feature extraction Stage 62 Dictionary collation means 63 feature extracting means 64 reject determining unit 65 reject learning means 69 individual character recognition means E1 single character place name recognition means

Claims (5)

A collective place name recognition program that causes a computer to execute a collective place name recognition procedure for recognizing place names of place name images including place names at a certain administrative level,
The collective place name recognition procedure includes:
An extraction procedure for extracting a feature vector from the place name image;
A collation procedure for collating the feature vectors extracted by the extraction procedure with the feature vectors of the place names stored in the place name dictionary and searching for all possible place name candidates;
To the computer,
Of the place names stored in the place name dictionary, a place name corresponding to a feature vector whose collation distance with the feature vector extracted by the extraction procedure is smaller than a predetermined threshold is determined as a possible place name candidate,
Further, an individual character recognition procedure for recognizing character candidates for characters included in each single character region of the place name image,
Wherein the recognition result obtained by the individual character recognition procedure, the batch place name recognition procedure whether the different extent is between strings between the determined place name candidates calculated by, unlike the individual characters if small A place name verification procedure for performing filtering so that the recognition result obtained by the recognition procedure is superior, or if the difference is large, the place name candidates determined by the collective place name recognition procedure are superior ;
Collective place name recognition program characterized by causing computer to execute.   The collective place name recognition procedure further causes the computer to execute a place name candidate evaluation procedure for filtering a plurality of possible place name candidates according to a predetermined rule when a plurality of possible place name candidates are obtained by the collating procedure. The collective place name recognition program according to claim 1.   The predetermined rule is that when a plurality of possible place name candidates are ranked based on the collation distance, and the possible place name candidates ranked first by the ranking are set as the first place name candidates, the collation distance of the place name candidates This is a rule that calculates a ratio of the difference between the collation distance of the first place name candidate and the collation distance of the first place name candidate to the collation distance of the first place name candidate, and selects only the place name candidates whose collation distance is smaller than a predetermined threshold. The collective place name recognition program according to claim 2 characterized by things.   A collective place name recognition device,
  Extracting means for extracting feature vectors from the input place name image;
  A place name dictionary section for storing place names;
  The place name feature vector stored in the place name dictionary unit is matched with the feature vector extracted by the extracting means to obtain all possible place name candidates and matching distances, and among the place names stored in the place name dictionary, Collating means for determining a place name corresponding to a feature vector whose matching distance between the feature vector extracted by the extracting means is smaller than a predetermined threshold as possible place name candidates;
  Evaluation of multiple address path candidates when multiple place name images containing administrative level place names arranged in descending order are input and multiple place name paths containing administrative level place names are formed by recognition of the place name images. And a collation evaluation means for making a decision,
  When there are a plurality of address paths having an average collation distance close to the minimum average collation distance, individual character recognition is performed on a place name image having a plurality of possible place name candidates in the plurality of address paths, or the collation Individual character recognition means for performing individual character recognition on a place name image having a plurality of possible place name candidates when the means determines a plurality of possible place name candidates for the place name image;
  Calculate how much the character string differs between the recognition result obtained by the individual character recognition means and the determination result by the collation evaluation means, and if the difference is small, the individual character recognition means A place name verification unit that performs filtering so that the obtained recognition result is superior, or when the difference is large, the determination result by the collation evaluation unit is superior;
  A collective place name recognition device characterized by comprising: A collective place name recognition method including a collective place name recognition process for recognizing place names of place name images including place names at a certain administrative level,
The collective place name recognition process includes:
An extraction step of extracting a feature vector from the place name image;
A collation step of collating the feature vector extracted by the extraction step with the feature vector of the place name stored in the place name dictionary and searching for all possible place name candidates;
Including
Of the place names stored in the place name dictionary, a place name corresponding to a feature vector whose collation distance with the feature vector extracted by the extraction step is smaller than a predetermined threshold is determined as a possible place name candidate,
Further, an individual character recognition step for recognizing character candidates for characters included in each single character region of the place name image,
Wherein the recognition result obtained by the individual character recognition process, the batch place names recognized whether the different extent is between strings between the determined place name candidates calculated by step, unlike the individual characters if small A place name verification step for performing filtering so that the recognition result obtained by the recognition step is dominant, or if the difference is large, the place name candidates determined by the collective place name recognition step are dominant ;
Collective place name recognition method characterized by including.

Images