JP2004139428A - Character recognition device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To reduce the frequency of misrecognition in the case of recognizing characters with accent marks in language systems such as German and French. <P>SOLUTION: Characters are recognized after previously removing accent marks which may cause misrecognition and words in the character string whose characters are recognized are collated by using a database from which accent marks are previously removed. <P>COPYRIGHT: (C)2004,JPO

Description

となっており、
ドイツ語によるウムラウト付き文字が含まれる単語は、
【数２】

となっており、
ドイツ語によるウムラウトなしの代用文字が含まれる単語は、
【数３】

となっている。
【００５８】
この第７の実施形態で用いる文字認識装置１０は、図２４に示すように、画像入力部１、前処理部２、文字行切り出し部３、文字切り出し部４、文字認識部５、後処理部６、答え出力部７、データベース８、アクセント記号識別部１１、判別部１３により構成されている。
【００５９】
ただし、文字認識部５には、ウムラウト付きの文字が含まれるドイツ語用の辞書３１、ウムラウト付きの文字を含まないドイツ語用の辞書３２が設けられている。データベース８には、ウムラウト付き文字が含まれる単語列の記憶部（単語辞書）３３と、ウムラウト付きの文字に代る代用文字が含まれる単語列の記憶部（単語辞書）３４とが設けられている。
アクセント記号識別部１１には、識別するアクセント記号ごとの基準画像が記憶されているアクセント記号（ウムラウト）用の辞書２５が設けられている。
【００６０】
アクセント記号識別部１１は、前処理部２のノイズ画像データテーブル２ｃに登録されているノイズ画像データとしての除去された黒画素の塊の形や大きさや二値画像全体における位置情報に基づいて、アクセント記号の画像を抽出し、画像サイズを正規化して、あらかじめアクセント記号用の辞書２５に用意されているアクセント記号画像とのマッチングをとることにより、アクセント記号の識別を行うものである。このアクセント記号識別部１１によるアクセント記号の識別結果は、後処理部６に出力される。
【００６１】
後処理部６は、アクセント記号識別部１１からウムラウトのアクセント記号の有りにより、データベース８のウムラウト付き文字が含まれる単語辞書３２により単語照合を行い、アクセント記号識別部１１からウムラウトのアクセント記号の無しにより、データベース８のウムラウト付き文字が含まれる単語をウムラウトつきの文字の代りに代用文字が含まれる単語に置き換えて登録されている単語辞書３３により単語照合を行うものである。
【００６２】
このような構成において、認識動作を図２５、図２６に示すフローチャートを参照しつつ説明する。
すなわち、ドイツの言語体系に見られるアクセント記号（ウムラウト）の付いた文字を用いて記載されている書類、あるいはアクセント記号の付いた文字の代りに代用文字を用いて記載されている書類等を、画像入力部１において、読取走査し、この読取画像を入力画像（図２参照）として前処理部２に出力する（ＳＴ５１）。
【００６３】
これにより、前処理部２は、入力された画像を二値画像に変換し、黒画素の連結成分を１つの塊とするようなラベリングを行う（ＳＴ５２）。この際、前処理部２は、上記したように、得られた黒画素の塊の外接矩形の両辺の長さが、あるしきい値よりも小さなものを「ノイズ」として除去する。
この際に除去された黒画素の塊の形や大きさや二値画像全体における位置情報をそのまま残したノイズ画像データが、ノイズ画像データテーブル２ｃに登録される（ＳＴ５３）。
【００６４】
これにより、アクセント記号識別部１１は、前処理部２のノイズ画像データテーブル２ｃに登録されているノイズ画像データとしての除去された黒画素の塊の形や大きさや二値画像全体における位置情報に基づいて、アクセント記号の画像を抽出し、画像サイズを正規化して、あらかじめアクセント記号用の辞書２５に用意されているアクセント記号（ウムラウト）画像とのマッチングをとることにより、アクセント記号の識別を行う（ＳＴ５４）。
【００６５】
このアクセント記号識別部１１によるアクセント記号の識別結果は、判別部１３に出力される。
これにより、判別部１３は、入力画像の全体に対する、アクセント記号識別部１１からのアクセント記号の識別結果としてウムラウトが記載されている場合、アクセント記号（ウムラウト）の付いた文字を用いて記載されているドイツ語文書と判断し、ウムラウトが記載されていない場合、アクセント記号（ウムラウト）の付いた文字の代りに代用文字を用いて記載されているドイツ語文書と判断する（ＳＴ５５）。
【００６６】
この判断結果は文字認識部５と後処理部６に出力される。
文字認識部５は、判別部１３からの判断結果がアクセント記号（ウムラウト）の付いた文字を用いて記載されているドイツ語文書の場合には、ウムラウト付きの文字が含まれる辞書３１を選択し、判別部１３からの判断結果がアクセント記号（ウムラウト）の付いた文字の代りに代用文字を用いて記載されているドイツ語文書の場合には、ウムラウト付きの文字を含まない辞書３２を選択する（ＳＴ５６）。
【００６７】
後処理部６は、判別部１３からの判断結果がアクセント記号（ウムラウト）の付いた文字を用いて記載されているドイツ語文書の場合には、ウムラウト付き文字が含まれる単語列の単語辞書３３を選択し、判別部１３からの判断結果がアクセント記号（ウムラウト）の付いた文字の代りに代用文字を用いて記載されているドイツ語文書の場合には、ウムラウト付きの文字の代りの代用文字が含まれる単語列の単語辞書３４を選択する（ＳＴ５７）。
また、前処理部２は、ラベリングの結果を文字行切り出し部３に出力する。
【００６８】
次に、文字行切り出し部３において、文字行を切り出し（ＳＴ５８）、この切り出された行画像から、さらに文字切り出し部４において、文字を切り出し（ＳＴ５９）、この切出した文字を文字認識部５によって上記選択されているドイツ語用の辞書３１あるいはドイツ語用の辞書３２を用いて認識する（ＳＴ６０）。
これにより、後処理部６はこの文字認識結果としての複数文字列の認識結果と、上記選択されているデータベース８のウムラウト付き文字が含まれる単語列の単語辞書３３あるいはウムラウト付き文字の代りの代用文字が含まれる単語列の単語辞書３４との照合を行う（ＳＴ６１）。
【００６９】
この照合により一致する照合結果を答え出力部７において出力する（ＳＴ６２）。
したがって、第１の実施形態では、誤読の原因となっているアクセント記号をあらかじめ除去してから認識し、あらかじめアクセント記号を除去したデータベースを用いて単語照合を行うことで誤読を削減することができる。
【００７０】
また、第４の実施形態では、削除したアクセント記号の種類により言語の判別を行って、用いる辞書を選択することができる。
また、第５の実施形態では、アクセント記号の位置から文字を切り出す場所を特定することができる。
また、第６の実施形態では、アクセント記号の形状を記憶しておいて、以降のアクセント記号識別に利用する。
上記したように、アクセント記号を積極的に利用することで認識率を向上できる。
【００７１】
【発明の効果】
以上詳述したように、この発明によれば、ドイツ語やフランス語等の言語体系に見られるアクセント記号の付いた文字の認識を行うものにおいて、アクセント記号が存在する文字列であっても、誤読なく文字認識を行うことができる文字認識装置を提供できる。
【図面の簡単な説明】
【図１】この発明の実施形態を説明するための文字認識装置の概略構成を示すブロック図。
【図２】入力された画像を説明するための図。
【図３】ノイズ画像データを説明するための図。
【図４】ウムラウトが記載されている文字を説明するための図。
【図５】抽出される外接矩形を説明するための図。
【図６】アクセント記号の付いた文字を含む単語列と、アクセント記号を除いた文字からなる単語列と、比較する単語列とを説明するための図。
【図７】認識動作を説明するためのフローチャート。
【図８】一文字として分離された文字の外接矩形の上側を１画素行ずつ削った際の文字パターンを説明するための図。
【図９】一文字ずつの文字認識処理を説明するためのフローチャート。
【図１０】一文字ずつの文字幅を推定する際の処理を説明するための図。
【図１１】一文字ずつの文字幅が一番小さくなる位置で、外接矩形を上下２つに分離した状態を説明するための図。
【図１２】一文字ずつの文字認識処理を説明するためのフローチャート。
【図１３】ドイツ語文書の一例を示す図。
【図１４】フランス語文書の一例を示す図。
【図１５】ドイツ語文書のアクセント記号を示す図。
【図１６】フランス語文書のアクセント記号を示す図。
【図１７】文字認識装置の概略構成を示すブロック図。
【図１８】認識動作を説明するためのフローチャート。
【図１９】認識動作を説明するためのフローチャート。
【図２０】文字が連結されたカーシブ文字列を示す図。
【図２１】アクセント記号の記憶例を示す図。
【図２２】アクセント記号の記憶例を示す図。
【図２３】アクセント記号の記憶例を示す図。
【図２４】文字認識装置の概略構成を示すブロック図。
【図２５】認識動作を説明するためのフローチャート。
【図２６】認識動作を説明するためのフローチャート。
【符号の説明】
１…画像入力部、２…前処理部、３…文字行切り出し部、４…文字切り出し部、５…文字認識部、５ａ…辞書、６…後処理部、７…答え出力部、８…データベース、１０…文字認識装置、１１…アクセント記号識別部。[0001]
TECHNICAL FIELD OF THE INVENTION
The present invention relates to a character recognition device that recognizes accented characters found in language systems such as German and French.
[0002]
[Prior art]
In linguistic systems such as German and French, there are letters with accent marks (such as umlauts) above and below alphabetic letters in English. These accent marks are important information for character recognition. Conventionally, character recognition has been performed with accent marks attached.
[0003]
However, in the case of resolution problems or handwritten characters, accent marks are crushed or come into contact with the main body of the alphabet character, causing misreading.
[0004]
[Problems to be solved by the invention]
The present invention recognizes characters with accent marks found in language systems such as German and French, and is capable of performing character recognition without misreading even a character string having accent marks. It is intended to provide a recognition device.
[0005]
[Means for Solving the Problems]
The character recognition device of the present invention performs character recognition by removing accent marks from accented characters in advance and recognizing characters with accent marks in a language system such as German or French. The word collation is performed by collating with a database composed of word strings created using characters in which accent characters have been removed from the characters.
[0006]
A character recognition device according to the present invention is a document created by a word including a character with an accent mark found in a language system such as German, or a word including a character in which a character with an accent mark is represented by an alternative character. A first word dictionary consisting of a word string containing a word containing an accented character, and an accented character represented by an alternative character in a document that recognizes characters from any of the documents created by A second word dictionary consisting of a word string containing the recognized characters, a judging means for judging the presence or absence of accent marks in the document, a recognizing means for recognizing the characters of the document, and a word string from the characters recognized by the recognizing means. Extraction means for extracting the word string extracted by the extraction means when the presence of accent marks in the document is determined by the determination means. The word string extracted by the extracting means and the second word are output when the determining means determines that there is no accent mark in the document by outputting the matching character string as a word by comparison with the first word dictionary. Output means for outputting a character string that matches by comparison with the dictionary as a word.
[0007]
BEST MODE FOR CARRYING OUT THE INVENTION
Hereinafter, a character recognition device according to an embodiment of the present invention will be described with reference to the drawings.
As shown in FIG. 1, the character recognition device 10 includes an image input unit 1, a preprocessing unit 2, a character line cutout unit 3, a character cutout unit 4, a character recognition unit 5, a post-processing unit 6, an answer output unit 7, It is composed of a database 8.
[0008]
First, the image input unit 1 inputs an image of a portion considered to be necessary information from an object in which necessary information is written. As the image input unit 1, a CCD camera, a scanner, or the like is used.
Next, in the pre-processing unit 2, the input image is converted into a binary image (bitmap data in the form of a matrix in pixel units), and labeling is performed such that the connected components of black pixels are formed into one block.
[0009]
The maximum X coordinate, Y coordinate, minimum X coordinate, and Y coordinate (coordinates of the circumscribed rectangular area) of the connected component (area) for each label image are measured and registered in the circumscribed rectangular area table 2a.
Also, the start point coordinates and end point coordinates of all the runs obtained in the raster direction scanning order for the connected components of each label image are registered in the run start point and end point coordinate table 2b.
[0010]
If both sides of the circumscribed rectangle of the obtained black pixel block are smaller than a certain threshold on both sides, they are removed as "noise". At this time, the accent mark may be removed. In addition, noise image data that retains the shape and size of the lump of black pixels removed at this time and position information in the entire binary image as they are is registered in the noise image data table 2c.
[0011]
For example, in the case of a German document as shown in FIG. 2, data for a point portion as shown in FIG. 3 is registered.
[0012]
Next, the character line cutout unit 3 extracts a character line. In the binary image, black pixels are projected in each of the horizontal direction (first direction) and the vertical direction (second direction), and a direction having a more regular peak is determined as a character line direction. An image is separated by judging a position where the value between the peaks of the projection values is small as a line boundary. What is separated here is a character line.
[0013]
From the line image cut out by the character line cutout unit 3, the line cutout unit 4 further separates the line image for each black pixel block. Compare the width of the preceding and succeeding black pixels with the width (the length of the circumscribed rectangle of the black pixels in the row direction), and if the width is clearly larger than the preceding and following black pixels, further increase the pixels in the direction perpendicular to the row direction. To separate. As a method of separation, a row image is projected with black pixels in a direction perpendicular to the row direction, and a portion having a small projection value is determined as a boundary of a character and separated. Alternatively, in the line image, the image is scanned from the upper and lower ends in the direction perpendicular to the line direction, the character width is obtained from the position where the image first meets the black pixel, and the portion having the small character width is determined and separated as the boundary of the character.
[0014]
The character obtained from the character cutout unit 4 is recognized by the character recognition unit 5. The recognition method may be an existing method (such as a composite similarity method).
Although the character recognition result may be used as it is, depending on the user of the present invention, identifiable information (for example, a place name) can be used. Perform collation.
Finally, the result of character recognition output from the character recognition unit 5 and the post-processing unit 6 is output by the answer output unit 7 as necessary.
[0015]
[First Embodiment]
The first embodiment performs character recognition by removing accent marks from accented characters (alphabet characters) in advance, and creates a word dictionary created using characters in which accent marks have been removed from accented characters in advance. Is performed by collating with the database 8 of the above.
[0016]
That is, in the pre-processing unit 2 of FIG. 1, the input image as shown in FIG. 2 is converted into a binary image (bitmap data), and the connected components of the black pixels of the binary image are converted into one block. Is performed. If the length of both sides of the circumscribed rectangle of the obtained black pixel block is smaller than a certain threshold value, it is removed as "noise". At this time, accent marks are removed.
[0017]
For example, when an umlaut is described as shown in FIG. 4, a circumscribed rectangle formed by a cluster of black pixels is as shown in FIG. 5, and two circumscribed rectangles formed by an umlaut and a circumscribed rectangle formed by “a” are extracted. At this time, since the lengths of both sides of the two circumscribed rectangles by the umlaut are smaller than a certain threshold value, black pixels by the umlaut are removed as noise from the binary image.
[0018]
In addition, the database 8 stores a word string (FIG. 6A) composed of characters obtained by removing accent marks from a word string (FIG. 6A) composed of accented characters found in a language system such as German or French. (B)) (word dictionary).
[0019]
In such a configuration, the recognition operation will be described with reference to the flowchart shown in FIG.
That is, a document or the like described using characters with accent marks found in a language system such as German or French is read and scanned by the image input unit 1, and the read image is input to the input image (see FIG. 2). ) Is output to the preprocessing unit 2 (ST1).
[0020]
As a result, the preprocessing unit 2 converts the input image into a binary image, and performs labeling such that the connected components of the black pixels become one block (ST2). At this time, as described above, the preprocessing unit 2 removes, as described above, those in which the lengths of both sides of the circumscribed rectangle of the obtained black pixel block are smaller than a certain threshold value as “noise”, and also removes accent marks. Remove.
The preprocessing unit 2 outputs the result of the labeling to the character line cutout unit 3.
[0021]
Next, the character line cutout unit 3 cuts out a character line (ST3), and further cuts out characters from the cutout line image in a character cutout unit 4 (ST4). Recognize (ST5).
As a result, the post-processing unit 6 recognizes a plurality of character strings as the character recognition result ((c) in FIG. 6) and a word string ((b) in FIG. )) (ST6).
A matching result that is matched by this matching is output from the answer output unit 7 (ST7).
[0022]
[Second embodiment]
In the second embodiment, as in the first embodiment, when the pre-processing unit 2 removes accent marks together with noise, in the case of an image having a low resolution, for example, the accent marks touch the alphabetic characters in the case of a low-resolution image. In some cases, such measures may be left, and this measure is taken. This is a countermeasure when noise cannot be removed.
[0023]
In this case, the circumscribed rectangle of each character cut out by the character cutout unit 4 is checked, and the height (side in the line direction and the vertical direction) of the preceding and succeeding characters is clearly larger (one higher than a predetermined pixel). ) Could be an accented character (which would include uppercase letters). In this case, as shown in FIGS. 8A, 8B, and 8C, the upper side or the lower side of the circumscribed rectangle of a character separated as one character is set one pixel row at a time (a width corresponding to the width of one pixel). The character recognition is repeated while shaving (one line at a time), and when the matching degree (similarity or the like) with the reference character included in the dictionary 5a is high, the answer is correct.
[0024]
In step 5 of the flowchart of the first embodiment, when the extracted character is recognized by the character recognition unit 5, the following processing is executed.
That is, in FIG. 9, a circumscribed rectangle for each character cut out by the character cutout unit 4 is checked (ST11), and whether or not the character is an accented character is determined based on whether or not the height of the preceding and following characters is clearly large. A decision is made (ST12).
[0025]
If the result of this determination is an accented character, the character recognizing unit 5 removes each reading pattern from the upper (uppermost line) to the lower predetermined pixel line of the circumscribed rectangle of the character separated as one character. Is calculated (ST13), and a character having the highest matching degree is output as a recognition result (ST14).
[0026]
In the case of a normal character that is not an accented character in step 12, the character recognizing unit 5 determines a matching degree (similarity degree) between a pattern obtained by normalizing a read pattern of a character separated as one character and each reference character included in the dictionary 5a. , Etc.) (ST15), and outputs the character with the highest matching degree as a recognition result (ST16).
[0027]
[Third Embodiment]
The third embodiment is a countermeasure in a case where the pre-processing unit 2 cannot remove the accent marks together with the noise as in the first embodiment.
In this case, the circumscribed rectangle of each character cut out by the character cutout unit 4 is checked, and the height (side in the line direction and the vertical direction) of the preceding and succeeding characters is clearly larger (one higher than a predetermined pixel). ) Could be an accented character (which would include uppercase letters). In this case, as shown in FIG. 10, the character widths of a plurality of positions of the character separated as one character are estimated, and the character is cut at the smallest portion of the estimated character width, as shown in FIG. Next, a circumscribed rectangle is generated for each of the upper and lower cut characters, and the character recognition unit 5 recognizes a character pattern corresponding to the larger circumscribed rectangle as a cut character.
[0028]
As a method of estimating the character width, pixels (one pixel unit or a plurality of pixels) are scanned from opposite left and right ends of the character circumscribed rectangle of the character extracted as a possibility of the character with the accent mark toward the opposite side. (See FIG. 10), the character width is estimated by finding the position where the black pixel is hit.
[0029]
In the case of FIG. 10, a gap is opened between the lower side of the accent mark and the upper side of the alphabet “a”, so that the character width therebetween becomes the smallest. At this position, as shown in FIG. It is separated into upper and lower parts, the larger circumscribed rectangle is determined, and character recognition is performed based on the character pattern in the determined circumscribed rectangle.
[0030]
In step 5 of the flowchart of the first embodiment, when the extracted character is recognized by the character recognition unit 5, the following processing is executed.
That is, in FIG. 12, a circumscribed rectangle for each character cut out by the character cutout unit 4 is checked (ST21), and whether or not the character is an accented character is determined based on whether or not the height of the preceding and following characters is clearly large. A decision is made (ST22).
[0031]
If the result of this determination is a character with an accent, the character recognition unit 5 estimates the character widths of a plurality of characters separated as one character (ST23), and cuts the character at the smallest portion of the estimated character width. As shown in FIG. 11, a circumscribed rectangle is generated for each of the characters cut vertically and vertically (ST24), and a pattern obtained by normalizing the character pattern for the larger circumscribed rectangle and each reference character included in the dictionary 5a is generated. The degree of matching (similarity etc.) is calculated (ST25), and the character with the highest degree of matching is output as a recognition result (ST26).
[0032]
In the case of a normal character that is not an accented character in step 22, the character recognition unit 5 matches a pattern (normalized read pattern of a character separated as one character) with each reference character included in the dictionary 5 a (similarity degree). , Etc.) (ST27), and outputs the character with the highest matching degree as a recognition result (ST28).
[0033]
[Fourth embodiment]
In the fourth embodiment, a language is determined based on accent symbol discrimination, utilizing the fact that accent marks used differ depending on the language, and recognition processing is performed based on the determined language-specific dictionary. is there.
For example, many countries in continental Europe are close together, so their documents are often written in multiple languages.
[0034]
In this case, for example, looking at the German document shown in FIG. 13 and the French document shown in FIG. 14, it can be seen that each language has a unique accent mark.
That is, in the case of a German document, umlauts are described as accent marks as shown in FIG. Further, in the case of a French document, as shown in FIG. 16, Axant-Tegu, Axant-Grave, Axant-Sylconflex, Trema and Cedilla are described as accent marks.
[0035]
Thus, if only umlaut is described as an accent mark, it is determined to be a German document, and if Axant-Tegu, Axant-Grave, Axant-Silconflex, Trema, Cedille, etc. are listed, French is used. Judge as a document.
In this way, by identifying the accent mark, it is possible to estimate in what language it is written.
[0036]
Also, accent symbols are extracted from the noise image data and the smaller circumscribed rectangle of the character cut in the third embodiment, and the extracted accent symbols are identified. Is estimated, a dictionary of the language is selected, and character recognition and word matching are performed.
[0037]
As a method of identifying accent symbols, there is a method of extracting an accent symbol image, normalizing the image size, and matching with an accent symbol image prepared in advance as a dictionary for accent symbols.
As shown in FIG. 17, a character recognition device 10 used in the fourth embodiment includes an image input unit 1, a preprocessing unit 2, a character line cutout unit 3, a character cutout unit 4, a character recognition unit 5, a post-processing unit. 6, an answer output unit 7, a database 8, an accent symbol identification unit 11, and a language determination unit 12.
[0038]
However, the character recognition unit 5 is provided with a dictionary 21 for French and a dictionary 22 for German. The database 8 is provided with a storage unit 23 for word strings for French and a storage unit 24 for word strings for German.
The accent symbol identification section 11 is provided with an accent symbol dictionary 25 in which a reference image for each accent symbol to be identified is stored.
[0039]
The accent symbol identification unit 11 determines the shape and size of a block of black pixels removed as noise image data registered in the noise image data table 2c of the preprocessing unit 2 and position information in the entire binary image. Accent symbol identification is performed by extracting an accent symbol image, normalizing the image size, and matching the accent symbol image prepared in the accent symbol dictionary 25 in advance. The accent symbol identification result by the accent symbol identification unit 11 is output to the language identification unit 12.
[0040]
The language discriminating unit 12 discriminates the language based on the result of accent symbol identification from the accent symbol identifying unit 11 for the entire input image. The result of language determination by the language determination unit 12 is output to the character recognition unit 5 and the post-processing unit 6.
[0041]
For example, when only umlaut is described as an accent mark, the language determination unit 12 determines that the document is a German document, and describes at least one of Axan-Tegu, Axan-Grave, Axan-Sylconflex, Cedilla, and the like. If so, it is determined to be a French document.
It should be noted that a French document may include a mixture of trema with the same notation as umlauts as French accent marks.
[0042]
In such a configuration, the recognition operation will be described with reference to the flowcharts shown in FIGS.
That is, a document or the like described using characters with accent marks found in a language system such as German or French is read and scanned by the image input unit 1, and the read image is input to the input image (see FIG. 2). ) To the preprocessing unit 2 (ST31).
[0043]
Thus, the preprocessing unit 2 converts the input image into a binary image, and performs labeling such that the connected components of the black pixels are formed into one block (ST32). At this time, as described above, the pre-processing unit 2 removes, as the “noise”, those in which the lengths of both sides of the circumscribed rectangle of the obtained lump of black pixels are smaller than a certain threshold.
At this time, noise image data that retains the shape and size of the lump of black pixels removed and the position information in the entire binary image is registered in the noise image data table 2c (ST33).
[0044]
With this, the accent symbol identification unit 11 stores the shape and size of the removed black pixels as the noise image data registered in the noise image data table 2c of the preprocessing unit 2 and the position information in the entire binary image. Accent symbol images are extracted, the image size is normalized, and the accent symbols are identified by matching them with the accent symbol images prepared in advance in the accent symbol dictionary 25 (ST34). .
The accent symbol identification result by the accent symbol identification unit 11 is output to the language identification unit 12.
[0045]
Thereby, when only the umlaut is described as the accent symbol identification result from the accent symbol identifying unit 11 with respect to the entire input image, the language discriminating unit 12 determines that the document is a German document. If at least one of Grave, Aksan-Silconflex, Cedilla, etc. is described, it is determined that the document is a French document (ST35).
This determination result is output to the character recognition unit 5 and the post-processing unit 6.
[0046]
Character recognition unit 5 selects dictionary 22 for German when the result of determination from language discrimination unit 12 is a German document, and selects dictionary 21 for French if it is a French document (ST36). ).
The post-processing unit 6 selects the German word string storage unit 24 of the database 8 when the determination result from the language determination unit 12 is a German document. The storage unit 23 of the word string for French is selected (ST37).
Further, the preprocessing unit 2 outputs the result of the labeling to the character line cutout unit 3.
[0047]
Next, the character line cutout unit 3 cuts out a character line (ST38), and further cuts out characters from the cutout line image in the character cutout unit 4 (ST39). Recognition is performed using the selected dictionary 21 for French or the dictionary 22 for German (ST40).
[0048]
Thereby, the post-processing unit 6 checks the recognition result of the plurality of character strings as the character recognition result against the selected word dictionary 23 for French or the word dictionary 24 for German in the selected database 8 ( ST41).
The matching result obtained by this matching is output from the answer output unit 7 (ST42).
[0049]
[Fifth Embodiment]
In the fifth embodiment, when an accent mark is added to a cursive character string in which characters are connected, which is common in handwritten characters, the position of the connected character can be estimated from that position.
In the character cutout unit 4 of the first embodiment, as shown in FIG. 20, when an accent mark is extracted, the character width is checked in both left and right directions from the accent mark position S based on the accent mark position S. A portion where the character width fluctuates rapidly is determined as a character and cut out.
[0050]
As a method for estimating the character width, a pixel (one pixel unit or a plurality of pixels) may be used from the upper and lower ends of the circumscribed rectangle of the character string extracted as the possibility that the character string is concatenated. ) Is scanned (see FIG. 20), and the position where the pixel collides with the black pixel is obtained to estimate the character width.
In the case of FIG. 20, the dotted line portion is a portion in which the character width fluctuates based on the position of the accent mark, and by dividing the dotted line portion, the handwritten characters “n”, “i”, “c”, and “e” are Each character is detected and cut out.
[0051]
[Sixth Embodiment]
In the sixth embodiment, a handwritten accent symbol image is registered and used for determining accent symbols in the next and subsequent times. Thereby, the recognition rate of the handwritten accent symbol written by the same person can be improved. This is particularly effective when there are few people using the present apparatus. This is because various writing habits of a specific person can be registered.
[0052]
That is, when the same person writes the handwritten characters, since the writing tool is also the same writing tool, the habit of the person and the writing tool appears in the way of writing the accent mark. As shown in FIG. 21 and FIG. 22, when umlaut or axan-tegu as an accent symbol could be extracted, the accent symbol image was stored in the storage device 50, and the accent symbol was extracted from the next time. In this case, the accuracy of the correct answer is increased by checking the identity by matching the accent symbol image.
[0053]
As for the matching method, the method of pattern matching by normalizing the image size, and the method of accent pattern with many straight lines, the length and angle of the contour, and the length and angle of the core line There is also a method of storing the value of and comparing the values.
For example, in the case of Axan-Silcon Flex, as shown in FIG. 23, the size and shape such as the length L of the contour and the inclination θ of the contour are stored.
[0054]
[Seventh embodiment]
In the seventh embodiment, in consideration of the fact that umlauts are used as accent marks in German, and that the umlauts can be expressed by substitution rules instead of the umlauts, a sentence created using umlauts is considered. The word dictionary to be used is changed between the case of performing recognition and the case of performing recognition of a sentence that is formed by a description (substitute character) according to a substitute rule instead of an umlaut.
[0055]
In other words, when recognizing a sentence created using umlauts, word matching is performed using a word dictionary in which words containing characters with umlauts are registered, and words are created using notation based on substitution rules instead of umlauts. When recognizing a sentence, a word including characters with an umlaut is replaced with a word including a substitute character without an umlaut, and word matching is performed using a registered word dictionary.
[0056]
Which of the two types of dictionaries to use can be determined based on the result of determining whether or not there is an accent symbol based on the noise image data. (Refer to the fourth embodiment)
As a result, the matching accuracy of word matching can be improved.
[0057]
Substitution rules without German umlauts are:
(Equation 1)

It is,
Words containing umlauts in German are:
(Equation 2)

It is,
Words that contain German umlauts without umlauts
[Equation 3]

It has become.
[0058]
As shown in FIG. 24, a character recognition device 10 used in the seventh embodiment includes an image input unit 1, a preprocessing unit 2, a character line cutout unit 3, a character cutout unit 4, a character recognition unit 5, a post-processing unit. 6, an answer output unit 7, a database 8, an accent symbol identification unit 11, and a determination unit 13.
[0059]
However, the character recognition unit 5 is provided with a dictionary 31 for German containing characters with umlaut and a dictionary 32 for German not containing characters with umlaut. The database 8 is provided with a storage unit (word dictionary) 33 for a word string including characters with umlaut and a storage unit (word dictionary) 34 for a word string including substitute characters for characters with umlaut. I have.
The accent symbol identification unit 11 is provided with an accent symbol (umlaut) dictionary 25 in which a reference image for each accent symbol to be identified is stored.
[0060]
The accent symbol identification unit 11 determines the shape and size of a block of black pixels removed as noise image data registered in the noise image data table 2c of the preprocessing unit 2 and position information in the entire binary image. Accent symbol identification is performed by extracting an accent symbol image, normalizing the image size, and matching the accent symbol image prepared in the accent symbol dictionary 25 in advance. The result of accent symbol identification by the accent symbol identification unit 11 is output to the post-processing unit 6.
[0061]
The post-processing unit 6 performs word matching with the word dictionary 32 including characters with umlaut in the database 8 when there is an umlaut accent symbol from the accent symbol identification unit 11, and there is no umlaut accent symbol from the accent symbol identification unit 11. Thus, the words containing the umlauted characters in the database 8 are replaced with the words containing the substitute characters instead of the umlauted characters, and word matching is performed by the registered word dictionary 33.
[0062]
In such a configuration, the recognition operation will be described with reference to the flowcharts shown in FIGS.
That is, documents written using accented characters (umlaut) found in the German language system, or documents written using substitute characters instead of accented characters, etc. The image input unit 1 performs read scanning and outputs the read image to the preprocessing unit 2 as an input image (see FIG. 2) (ST51).
[0063]
As a result, the preprocessing unit 2 converts the input image into a binary image, and performs labeling such that the connected components of black pixels are formed into one block (ST52). At this time, as described above, the pre-processing unit 2 removes, as the “noise”, those in which the lengths of both sides of the circumscribed rectangle of the obtained lump of black pixels are smaller than a certain threshold.
At this time, noise image data that retains the shape and size of the lump of black pixels removed and the position information in the entire binary image as they are is registered in the noise image data table 2c (ST53).
[0064]
With this, the accent symbol identification unit 11 stores the shape and size of the removed black pixels as the noise image data registered in the noise image data table 2c of the preprocessing unit 2 and the position information in the entire binary image. Then, the accent symbol image is extracted, the image size is normalized, and the accent symbol is identified by matching with an accent symbol (umlaut) image prepared in advance in the accent symbol dictionary 25. (ST54).
[0065]
The result of the accent symbol identification by the accent symbol identification unit 11 is output to the determination unit 13.
Accordingly, when umlaut is described as the accent symbol identification result from the accent symbol identifying unit 11 with respect to the entire input image, the discriminating unit 13 is described using a character with an accent symbol (umlaut). If the document is determined to be a German document and no umlaut is described, it is determined that the document is described using a substitute character instead of a character with an accent mark (umlaut) (ST55).
[0066]
This determination result is output to the character recognition unit 5 and the post-processing unit 6.
In the case of a German document in which the judgment result from the discriminating unit 13 is described using a character with an accent mark (umlaut), the character recognition unit 5 selects the dictionary 31 including the character with the umlaut. In the case of a German document in which the determination result from the determination unit 13 is described using a substitute character instead of a character with an accent mark (umlaut), the dictionary 32 that does not include a character with an umlaut is selected. (ST56).
[0067]
In the case of a German document in which the determination result from the determination unit 13 is described using a character with an accent mark (umlaut), the post-processing unit 6 includes a word dictionary 33 of a word string including a character with an umlaut. Is selected, and in the case of a German document in which the determination result from the determination unit 13 is described using a substitute character instead of a character with an accent mark (umlaut), a substitute character instead of a character with an umlaut Is selected (ST57).
Further, the preprocessing unit 2 outputs the result of the labeling to the character line cutout unit 3.
[0068]
Next, the character line cutout unit 3 cuts out a character line (ST58), and further cuts out characters from the cutout line image in the character cutout unit 4 (ST59). Recognition is performed using the selected German dictionary 31 or German dictionary 32 (ST60).
As a result, the post-processing unit 6 substitutes the recognition result of the plurality of character strings as the character recognition result with the word dictionary 33 of the word string including the umlauted character of the selected database 8 or the substitute for the umlauted character. The word string including the characters is collated with the word dictionary 34 (ST61).
[0069]
The matching result obtained by this matching is output from the answer output unit 7 (ST62).
Therefore, in the first embodiment, it is possible to reduce misreading by performing recognition by removing accent marks that cause misreading in advance and then performing word matching using a database from which accent marks have been removed. .
[0070]
Further, in the fourth embodiment, it is possible to determine the language based on the type of the deleted accent mark and select a dictionary to be used.
Further, in the fifth embodiment, it is possible to specify a place where a character is cut out from the position of an accent mark.
In the sixth embodiment, the shape of the accent symbol is stored and used for subsequent accent symbol identification.
As described above, the recognition rate can be improved by actively using accent marks.
[0071]
【The invention's effect】
As described in detail above, according to the present invention, in recognition of accented characters found in a language system such as German or French, even if the character string contains an accent mark, it may be misread. And a character recognition device capable of performing character recognition without the need.
[Brief description of the drawings]
FIG. 1 is a block diagram showing a schematic configuration of a character recognition device for describing an embodiment of the present invention.
FIG. 2 is a diagram illustrating an input image.
FIG. 3 is a diagram for explaining noise image data.
FIG. 4 is a diagram for explaining characters in which umlauts are described.
FIG. 5 is a diagram for explaining a circumscribed rectangle to be extracted;
FIG. 6 is a diagram for explaining a word string including a character with an accent mark, a word string including characters without an accent mark, and a word string to be compared.
FIG. 7 is a flowchart illustrating a recognition operation.
FIG. 8 is a view for explaining a character pattern when the upper side of a circumscribed rectangle of a character separated as one character is cut by one pixel row.
FIG. 9 is a flowchart for explaining character recognition processing for each character.
FIG. 10 is a view for explaining processing when estimating the character width of each character.
FIG. 11 is a view for explaining a state where a circumscribed rectangle is divided into upper and lower parts at a position where the character width of each character is the smallest.
FIG. 12 is a flowchart for explaining a character recognition process for each character.
FIG. 13 is a diagram showing an example of a German document.
FIG. 14 is a diagram showing an example of a French document.
FIG. 15 is a diagram showing accent marks in a German document.
FIG. 16 is a diagram showing accent marks in a French document.
FIG. 17 is a block diagram illustrating a schematic configuration of a character recognition device.
FIG. 18 is a flowchart illustrating a recognition operation.
FIG. 19 is a flowchart illustrating a recognition operation.
FIG. 20 is a diagram showing a cursive character string in which characters are connected.
FIG. 21 is a diagram showing a storage example of accent marks.
FIG. 22 is a diagram showing a storage example of accent symbols.
FIG. 23 is a diagram showing a storage example of accent symbols.
FIG. 24 is a block diagram illustrating a schematic configuration of a character recognition device.
FIG. 25 is a flowchart illustrating a recognition operation.
FIG. 26 is a flowchart illustrating a recognition operation.
[Explanation of symbols]
DESCRIPTION OF SYMBOLS 1 ... Image input part, 2 ... Preprocessing part, 3 ... Character line cutout part, 4 ... Character cutout part, 5 ... Character recognition part, 5a ... Dictionary, 6 ... Post-processing part, 7 ... Answer output part, 8 ... Database Reference numeral 10: character recognition device; 11: accent symbol identification unit.

Claims (12)

In recognizing accented characters found in language systems such as German and French,
Accent marks are removed from accented characters in advance and character recognition is performed.
A character recognition device, which performs word matching by collating with a database consisting of a word string created using characters in which accent symbols have been removed from accented characters in advance. In recognizing accented characters found in language systems such as German and French,
When character recognition is performed in a language system having characters with accent marks, character recognition is performed by cutting lines formed with a width corresponding to one pixel above or below a circumscribed rectangle of a character separated as one character one by one. A character recognition device characterized by repeating. When performing character recognition in a language system having accented characters, the pixel position that is considered to be a character part is obtained by scanning pixels in a direction away from the left and right ends of the circumscribed rectangle of the character separated as one character A character recognition device that estimates a character width, cuts a character at a portion having a small width in the obtained character width, and performs character recognition on a larger cut character element. By extracting accents from accented characters and identifying accents, the language of the sentence in which the accented characters are written is estimated, and a character dictionary of the estimated language is selected to recognize characters. A character recognition device characterized by performing the following. A character recognition apparatus characterized in that accent characters are extracted from a cursive character string to which characters used in handwritten characters are connected, and characters are cut out from the cursive character string based on the positions of the accent marks to perform character recognition. In the handwritten character recognition, a character recognition device characterized by storing the shape of an accent mark of a previously detected accented character and using the shape of the accent mark detected thereafter. A character recognition apparatus characterized by using a word dictionary converted into two vowels corresponding to a vowel with an umlaut when it is determined in the German-speaking character recognition that there is no umlaut in the document. In a character recognition device for recognizing a word including a character to which an accent mark is added,
Conversion means for converting the supplied image into a binary image;
A character line cutout unit that cuts out a character line by projecting black pixels in the first and second directions orthogonal to each other from the binary image converted by the conversion unit;
A black pixel is projected in a second direction perpendicular to the first direction from the image in the first direction cut out by the character line cutout means, and a small portion of the obtained projection value is defined as a boundary between characters. Character separating means for determining and separating a line image for each black pixel block,
Recognizing means for recognizing characters separated by the separating means;
Extracting means for extracting a word string from the character recognized by the recognition means;
Output means for outputting a character string that matches as a word by comparing the word string extracted by the extraction means with a database of character strings from which accent marks have been removed,
A character recognition device comprising: In recognizing accented characters found in language systems such as German and French,
By extracting accent symbols from the accented characters, identifying the extracted accent symbols, estimating the language of the sentence in which the accented characters are written, and selecting a character dictionary of the estimated language A character recognition device that performs character recognition. In recognition of accented characters,
Extracting means for extracting accent marks from accented characters;
Identification means for identifying accent marks extracted by the extraction means;
Determining means for determining the language from the accent marks identified by the identifying means;
A character recognition device characterized by selecting a character dictionary of a language determined by the determination means and performing character recognition. Either a document composed of words containing accented characters found in a language system such as German, or a document composed of words containing characters that represent accented characters as alternative characters In a character recognition device that recognizes a character from
A first word dictionary comprising a word string containing words containing accented characters;
A second word dictionary including a word string including characters in which accented characters are represented by alternative characters;
Determining means for determining the presence or absence of accent marks in the document;
Recognizing means for recognizing characters of the document;
Extracting means for extracting a word string from the character recognized by the recognition means;
When the presence of accent marks in the document is determined by the determination means, a character string that matches by comparison between the word string extracted by the extraction means and the first word dictionary is output as a word. Output means for, when it is determined that there is no accent mark in the document, outputting a character string that matches as a word by comparing the word string extracted by the extraction means with the second word dictionary;
A character recognition device comprising: The apparatus of claim 11, wherein the accent mark is an umlaut.

Images