JPWO2007094078A1 - Character string search method and apparatus - Google Patents

Abstract

A pattern character string input unit for receiving a character string to be searched from a user in a character code and a handwritten stroke in one of or both kinds of expression formats, and a character code and a handwritten stroke in one or both of expressions A document management unit that manages document information having a character string consisting of a format and a character string in the document to be searched and the input character string to be searched are both converted into a character cutout graph. A character string search system comprising: a character string matching unit that extracts a portion where a pattern character string appears in a character string to be searched by matching; and a display unit that displays a document and a character string search result.

Description

  The present invention relates to a method and an apparatus for searching a character string composed of a character code and a handwritten stroke.

Character string search is a process of finding a character string (pattern) or a search key designated by a user or the like from a character string to be searched such as a sentence described in a document. The expression format of the character string used in the character string search is generally a character string made up of character code strings conforming to standards such as SJIS and Unicode. In recent years, the spread of touch panels that can scan the movement of pens on the screen, and digital pens that can acquire strokes written on paper (for example, see International Publication No. 01/71473) make it easy to write handwritten character strings. It has become possible to digitize. Therefore, a character string search method in the case where a search target and a pattern character string (search key) are expressed in a handwritten stroke format has become necessary.
As a method for searching for a character string, Japanese Patent Application Laid-Open No. 2005-251222 discloses a method 1 for collating sampling point sequences constituting a character string stroke with a dynamic programming method.
Japanese Patent Application Laid-Open No. 10-055409 discloses a method 2 in which a stroke is converted into a line of fine segments from a line direction feature or the like, and segment lines are collated.
Further, Japanese Patent Laid-Open No. 2002-259912 discloses a subset of strokes (character segmentation candidates) that may constitute a character from a stroke set that constitutes a character string, and an appropriate character code from the shape of each character segmentation candidate. A method 3 is disclosed in which a candidate character lattice that includes a set of character extraction candidates and character identification candidates is created and collated using character identification means for identifying the character.

However, since the conventional method 1 disclosed in Japanese Patent Laid-Open No. 2005-251222 and the conventional method 2 disclosed in Japanese Patent Laid-Open No. 10-055409 depend on the order of sampling points of the stroke, the same character shape can be written. If the order is different, there is a problem that collation becomes difficult. Further, when the search target and the handwritten character string of the keyword are entered by different users, there is a problem that the search accuracy is lowered. Furthermore, since these methods are stroke-based searches, there is a problem that the search target or pattern character string cannot be searched if the character string is expressed by a character code instead of a handwritten stroke.
Further, in the conventional method 3 disclosed in Japanese Patent Laid-Open No. 2002-259912, the character cutout candidates and the character identification candidates are not always correct. If the character cutout candidates and the character identification candidates include an error, the characters are correctly displayed. There was a problem that the column could not be searched.
The present invention has been made in view of such problems.
That is, an object of the present invention is to provide a character string search method that can be applied even when the writer and the writing order are different between the search target and the pattern.
Another object of the present invention is to provide a character string search method that allows an error in character extraction candidates and character identification candidates even if errors exist.
Furthermore, an object of the present invention is to provide a character string search method that allows an arbitrary combination of a character code and a handwritten stroke as an expression format of a character string to be searched and a pattern.
In order to solve the above-described problems, typical inventions disclosed in the present application are as follows.
A pattern character string input unit for receiving a character string to be searched from a user in one or both of two types of expression formats of a character code and a handwritten stroke, and either or both of a character code and a handwritten stroke A document management unit that manages information on a document having a character string in the expression format, and converts the character string in the document to be searched and the input character string to be searched into a character cutout graph, and a character cutout graph. A character string search system comprising: a character string matching unit that extracts a portion where a pattern character string appears in a search target character string by matching each other; and a display unit that displays a document or character string search result.
According to the present invention, a user can find a character string to be searched for in a document with high accuracy. You can also find other writing than yourself. Furthermore, there is an effect that all character strings in a document can be searched without distinguishing between character codes and handwritten strokes.

FIG. 1 is a configuration diagram of a character string search system according to an embodiment of this invention.
FIG. 2 is a diagram showing types of character string search that can be realized in the embodiment of the present invention.
FIG. 3 is a diagram showing an example of a document to be processed in the embodiment of the present invention.
4A and 4B are diagrams showing character extraction graphs of search objects and patterns according to the embodiment of the present invention.
FIG. 5 is a diagram for explaining a character cutout graph matching method according to the embodiment of this invention.
6A and 6B are diagrams showing a data structure of character code information according to the embodiment of the present invention.
7A and 7B are diagrams showing a data structure of stroke information according to the embodiment of the present invention.
FIG. 8 is a diagram showing a data structure of document information according to the embodiment of the present invention.
9A to 9D are diagrams showing a data structure of the character segmentation graph according to the embodiment of the present invention.
FIG. 10 is a diagram illustrating a data structure of a character string search result according to the embodiment of this invention.
FIG. 11 is a diagram illustrating a data structure of a character string search candidate according to the embodiment of this invention.
FIG. 12 is an explanatory diagram of preprocessing according to the embodiment of this invention.
FIG. 13 is an explanatory diagram of character cutout creation processing according to the embodiment of this invention.
FIG. 14 is an explanatory diagram of character string search processing according to the embodiment of this invention.
FIG. 15 is an explanatory diagram of the character cutout graph matching process according to the embodiment of this invention.

First, a configuration example of the character string search system of the present invention is shown. Next, a character string search processing flow for finding a pattern character string designated by the user from document information managed by the character string search system will be described.
As shown in FIG. 1, the character string search system of the present invention comprises the following parts. A pattern character string input unit 101 included in the character string search device 100 is an input interface that can be connected to a keyboard, a digital pen 105, or the like in order to input a character string (pattern character string) to be searched for in the character string search. The document management unit 102 is realized by a storage unit such as a hard disk and a control unit that controls reading and writing of the storage unit, and manages a document including character code and stroke information entered with a pen or the like.
The character string collating unit 103 executes each program module stored in the storage unit by the calculation unit to find a place where the pattern character string appears in the document, and inputs the document or pattern character string stored in the storage unit. A procedure for generating a character cut-out graph from a character string input from the section or performing collation is realized. The display unit 104 displays documents and character string search results.
The pattern character string input unit 101 receives a pattern character string in one of two types of expression formats: a character code format by keyboard input or the like, and a stroke format by digital pen or touch panel input.
The document management unit 102 manages each document information with the data structure shown in the table 800 of FIG. An item 801 is an ID for identifying document information. Items 802 and 803 respectively indicate the total number of character codes and the total number of strokes included in the document. Items 805 and 806 to 807 store the ID of each character code included in the document. The items 808 and 809 to 810 store the ID of each stroke included in the document. Further, the document has a character cutout graph (for example, FIGS. 4A and 4B) created from strokes in the document by the preprocessing shown in FIG. 12 described later. An item 804 indicates the total number of character cutout graphs included in the document. Items 811 and 812 to 813 store the ID of each character segmentation graph included in the document. Each data structure of the character code, stroke, and character cutout graph will be described later.
The character string matching unit 103 compares the input pattern character string with the character string in the document. Comparison of character strings will be described in detail later with reference to FIGS.
The display unit 104 displays a document accumulated in the document management unit, a character string search result, and the like on a monitor screen or the like for the user. The document may be printed on paper or the like.
An example of a document handled in the present invention is shown in FIG. The document 300 of this example is composed of a character string 301 made up of character codes and strokes 302 and 303 written with a pen 310, respectively. In this embodiment, a digital pen disclosed in International Publication No. 01/71473 is used as a stroke acquisition unit. The stroke described on the paper surface can be digitized by the digital pen.
Here, the stroke 303 is considered. There are two possible interpretations of the first half of the stroke: first, the first character is regarded as “# (sharp)”, or “well (kanji well)”. If the first character is regarded as “# (sharp)”, the character string after # can be regarded as a number. This interpretation can also be interpreted as “7” with one digit or “17” with two digits. In addition, if the first character is regarded as “I (well character well)”, the second character can be interpreted as “mouth” if the subsequent character is assumed to be a kanji character.
In this way, interpretations such as (1) # 7, (2) # 17, (3) Iguchi are established, and it is impossible for a human to discriminate just by looking at this part. Thus, there is imperfection in character recognition. Therefore, the method of recognizing a handwritten character string, converting it into a character code string, and collating the character string has a problem that the handwritten character string cannot be searched with high accuracy.
The character code information is held in the data structure shown in the table 600 of FIG. 6B. For example, the character code 651 “D” on the document 650 in FIG. 6A is represented by information described in the items 601 to 607.
An item 601 is an ID for identifying the character code. An item 602 indicates a character code value. Items 603 to 607 are character attribute information, and items 603 and 604 indicate the coordinates of the upper left point of the character rectangle on the document in millimeters. An item 605 indicates a font type, an item 606 indicates a font size, and an item 607 indicates a style such as italic or bold.
The stroke information is held in the data structure shown in the table 700 of FIG. 7B. For example, the stroke 751 on the document 750 shown in FIG. 7A is represented by information described in the items 701 to 704. An item 701 is an ID for identifying the stroke. An item 702 indicates a stroke entry start time.
An item 703 indicates the number of sampling points existing in the stroke. Information on each sampling point is held in a table 730. Item 704 indicates a pointer pointing to the head of the sampling point set corresponding to the stroke. Each sampling point has XY coordinate values 731 and 732 on the document, and the difference between the sampling point entry time and the stroke entry start time described in the item 702 is held in 733. Information about each of the other strokes is also stored in a data structure as shown in Tables 710 and 720. Further, even when handwritten characters are input by an input device other than a digital pen, the stroke information can be managed by a table 730 that disassembles each stroke and holds the stroke ID, the number of sampling points, the pointer, and the coordinates.
The character string search system 100 executes preprocessing shown in FIG. 12 when inputting a document to be searched. In step 1202, a document is input. Stroke information in the input document is acquired (step 1203), and a stroke set analysis process in step 1204 divides the stroke set in the document into figures and character strings. For example, the strokes on the document 300 in FIG. 3 are divided into a stroke set 302 that forms a circle and a balloon, and a stroke set 303 that forms an annotation character string.
Next, step 1205 is executed for each divided stroke set to obtain a character segmentation graph. A character segmentation graph is a representation of a plurality of possible character segmentation hypotheses as a single directed acyclic graph because it is difficult to uniquely determine the correct segmentation of characters.
For example, in the portions 411, 412 and 413 in FIG. 4A, the interpretation of the right portion of the stroke 401 can be both one character and two characters. Since the graph 410 has both 411 representing the interpretation result as one character and two edges 412 and 413 representing the interpretation result as two characters, the multiple hypotheses can be expressed. The obtained character cutout graph is stored in the document management unit 102 in the data structure shown in FIGS. 9B to 9D, and the character cutout graph ID is stored in the items 811 and 812 to 813 of the document data structure 800.
The data structure of the character cutout graph is shown in the table 900 of FIG. 9D. Here, the case of the character segmentation graph shown in the graph 950 will be described. An item 901 is an ID for identifying a character cutout graph. An item 902 indicates the ID of a document in which the character cutout graph is described. Items 903 and 904 indicate the total number of nodes and edges of the character segmentation graph, respectively. Items 905 to 908 indicate the information of the first edge, the item 905 indicates the number of the start node of the edge, and the item 906 indicates the number of the end node of the edge. Each edge has a character cutout candidate consisting of a stroke set and a character identification candidate for the character cutout candidate, and items 907 and 908 indicate their IDs. Such edge information is repeatedly expressed by the number described in the item 904.
The data structure of character extraction candidates is as shown in Table 920. An item 921 indicates an ID for identifying the character segmentation candidate. An item 922 indicates the number of strokes included in the character cutout candidates. Items 923 to 924 indicate the ID of each stroke included in the character candidate.
The data structure of character identification candidates is as shown in Table 930. An item 931 indicates an ID for identifying the character identification candidate. An item 931 indicates the number of character identification results output by the character identification process. Items 933 and 934 indicate the first character identification result. An item 933 indicates the character code for which the character is identified. An item 934 indicates the degree of similarity at that time. In the case of this example, the first place is “# (sharp)”, and the second place is “well” of the Chinese character with a close difference of 0.02.
The process for creating the character segmentation graph shown in step 1205 of FIG. 12 will be described in detail with reference to FIG. In step 1302, after inputting the corresponding stroke set, in step 1303, the stroke set is cut into subsets that can be considered to constitute characters. This operation is equivalent to creating a character cutout candidate at the edge of the character cutout graph.
At this time, it may be difficult to uniquely determine the cutout. For example, it may be difficult to determine whether the right portion of the stroke set 401 on the document 400 in FIG. 4A is “7” 1 character or “1” and “7”. In such a case, by creating an edge 411 and a pair of edges 412 and 413, the interpretation of both “7” 1 character and “1” “7” 2 characters can be expressed simultaneously in one cutout graph. it can. Therefore, by using the character cutout graph, it is possible to perform a character string search that allows incomplete character cutout.
In the next step 1304, a character code is identified from the stroke shape of each character extraction candidate. The obtained character identification candidates are held in the data structure shown in the table 930 of FIG. 9C. For example, in the case of the edge 951 in FIG. 9A described above, it is difficult to identify “# (sharp)” or “well” only from the shape of the character extraction candidate, but both of them can be obtained from the data structure shown in Table 930. Can be expressed simultaneously. Therefore, by using the character segmentation graph, it is possible to perform a character string search that allows incomplete character identification.
Finally, in step 1305, character cutout candidates and character identification candidate information are stored in association with each edge, and a character cutout graph is output.
Note that the character cutout graph can be created not only for strokes in a document but also for character codes. By converting each character code into a character cut edge for a character code string, a single-way character cut graph can be created. At this time, it is assumed that the character identification candidate of each edge has a similarity of 1.0 as a result of the corresponding character code 1.
In this manner, a character string is searched for the input pattern character string using the character segmentation graph obtained in the preprocessing (FIG. 14).
First, in step 1402, a pattern character string to be searched is input. The expression form of the pattern character string corresponds to both a character code and a handwritten stroke. Next, in step 1403, a character cutout graph is created from the pattern character string. A character cutout graph created from this pattern character string is hereinafter referred to as a pattern character cutout graph.
In accordance with steps 1203 to 1205 in FIG. 12 described above and a character cutout graph generation procedure similar to that described with reference to FIG. 13, a pattern character cutout graph is generated from a character code or a handwritten stroke.
Next, if there is a character cutout graph for the document managed by the document management unit 102, it is acquired (steps 1404 and 1405). This is called a search target character segmentation graph. The previous pattern character cutout graph is compared with the search target character cutout graph to detect the appearance of the pattern character cutout graph (step 1406). Details of this processing will be described later.
The result of collating the character cutout graph, that is, the character string search result is held in the data structure shown in the table 1000 of FIG.
An item 1001 indicates an ID for identifying the character string search result. An item 1002 is an index value of the result, and a larger value means a more reliable search result. Items 1003 and 1004 are IDs of the search target character cutout graph and the pattern character cutout graph, respectively.
An item 1005 indicates the number of matching portions of each edge of the search target character cutout graph and the pattern character cutout graph. Items 1006 and 1007 indicate information on the first matching portion, and indicate the corresponding edge number in the search target character cutout graph and the corresponding edge number in the pattern character cutout graph, respectively.
In step 1407, the character string search result obtained in step 1406 is registered as a character string search candidate. The data structure of the character string search candidate is shown in Table 1100 of FIG. According to the score of the character string search result, the ID of the character string search result is registered in this table. Accordingly, the ID 1103 of the character string search result # 1 in the table 1100 always indicates the ID to the character string search result having the highest score. Actually, a maximum value such as up to 10 candidates may be set for the total number of character string search results 1102. In this case, character string search result IDs with higher candidate ranks are removed from the list.
The processing of steps 1405 to 1407 is executed for all the search target character cutout graphs. When all the processing is completed, a character string search candidate is output (1408).
A character string search result is obtained by the processing described above. Since the pattern character string and the expression format of the search target character string are converted to a character cutout graph and collation is performed, it is possible to search without distinguishing either the character code or the handwritten stroke originally (FIG. 2).
Finally, character cutout graph matching will be described with reference to FIGS. In the present embodiment, a description will be given of an application of dynamic programming as a graph matching method.
First, a character extraction graph of a search target and a pattern is input (1502). It is assumed that the search target character cutout graph nodes are T0, T1,..., Tm, and the pattern character cutout graph nodes are P0, P1,. Next, a collation table of (n + 1) rows (m + 1) columns is created, and an initial value 0 is substituted for each value (1503). Each value of the collation table corresponds to the score of each node PiTj (0 ≦ i ≦ n, 0 ≦ j ≦ m) in the matrix 500 of FIG.
Next, the score of each node in the matrix 500 is calculated for each Pi row (step 1504). In each row, calculate one by one from the left column. For example, assume that the matrix node P1T2 (501) is calculated. At this time, it is calculated by adding the transition score from the score of the already calculated matrix node connected to P1T2.
The target of the transition source matrix node PaTb is an object in which 0 ≦ a ≦ 1, 0 ≦ b ≦ 2 and an edge (PaTb, P1T2) having PaTb as a start point and P1T2 as an end point exists in the matrix 500. When the edge (Pa, Pb) exists in the pattern character cutout graph 560 and (Tb, T2) also exists in the search target character cutout graph 550, the matrix edge (PaTb, P1T2) exists in the matrix 500.
In the case of this example, the transition source nodes are P0T0 and P0T1. In addition, a gap transition (longitudinal or lateral transition) is also allowed in consideration of an extra stroke (insertion). That is, in this example, P0T2 when insertion occurs in the pattern character cutout graph and P1T1 when insertion occurs in the search target character cutout graph are also set as the transition source nodes. At this time, the score S (P1T2) of P1T2 is calculated using the PaSb score S (PaTb) and the transition score T (PaTb, P1T2) from PaTb to P1T2.
S (P1T2) = max (S (PaTb) + T (PaTb, P1T2)) (Equation 1)
It can be expressed.
Several methods of calculating the transition score T (PaTb, P1T2) are conceivable. In this embodiment, one method using information of character identification candidates is shown. Assume that the similarity of character identification when the character identification result of the edge (Pa, P1) is the character code C is s (Pa, P1, C). However, if there is no corresponding character identification result in C, s (Pa, P1, C) = γ << 1.0. At this time, the transition score is as follows.
T (PaTb, P1T2) = max (s (Pa, P1, C) + s (Tb, T2, C)) (Equation 2)
In Equation 2, the sum of the similarities is used, but a product may be used.
T (PaTb, P1T2) = max (s (Pa, P1, C) × s (Tb, T2, C)) (Equation 3)
In the case of a product, if one of the two similarities is low, the product is integrated, so the value is lower than the sum. Therefore, in the case of products, it is advantageous if there is a common character code with high similarity.
The transition cost in the case of gap transition is defined as follows:
T (PiTj-1, PiTj) = α << 1.0 (Equation 4)
T (Pi-1Tj, PiTj) = β << 1.0 (i ≠ 1, n), 1.0 (i = 1 or n) (Equation 5)
Here, when i = 1 or n in the gap transition in the horizontal direction (initial gap), the reason why the gap score is a perfect score of 1.0 is that the pattern character string appears at the beginning / end of the character string to be searched. However, the initial gap penalty was set to 1.0.
In this way, the score of each node in the matrix is calculated in order from left to right and from the top row to the bottom row, and the collation table is filled. After the collation table calculation, the collation result of the character cutout graph, that is, the character string search result is obtained from the symbol table (step 1505). Contrary to the calculation order, the matrix node PnTm at the lower right corner is traced in the reverse order in the upper left direction.
Assume that PiTj is currently focused. Of the PaTb of the transition source node of PiTj, the node with the highest score is set as the next node of interest, and the edge numbers of edges (Tb, Tj) and (Pa, Pi) are extracted and temporarily stored. When it finally reaches P0T0, a set of search object edge numbers and pattern edge numbers is registered in the items 1006 and 1007 and 1008 and 1009,... In FIG. A column search result 1000 is completed.
Here, the characteristics of the matrix 500 will be supplemented. Edges (PaTb, PiTj) in the matrix are always a ≦ i and b ≦ j. Therefore, since the score can be calculated in order from the upper line to the lower line, dynamic programming can be applied.
If the character string to be searched is only a character code, the character segmentation graph 550 has a single-chain structure without branching, and therefore j−1 ≦ b ≦ j. That is, each edge in the matrix is an edge that changes only at most once in the horizontal direction.
Similarly, when the pattern character string is only a character code, the character cutout graph 560 has a single-chain structure without branching, and therefore i-1 ≦ a ≦ i. In other words, each edge in the matrix is an edge that transitions to at most one in the vertical direction.
Only when both the search target character string and the pattern character string are handwritten, there are transition edges larger than 1 in both vertical and horizontal directions, as indicated by edges 511 and 512. In other words, by creating such an edge in the matrix, the graph can be collated.
Finally, the character string search result is output (step 1506), and the character cutout graph collation is terminated.
The above is the description of the embodiment according to the present invention.
When performing a search, a character string search may be performed by limiting the range of documents to be searched or by limiting the character string to be searched to only a character code or a handwritten stroke.
In addition, since the pattern character string is converted into a character cutout graph and searched, the search can be performed even when the pattern character string includes a combination of a character code and a stroke. This is because it is possible to create a pattern character cutout graph by converting each graph into a character cutout graph and combining the end node and the start end node of each graph. Similarly, a search target character string can be searched by collecting character strings formed by combinations of character codes and strokes into one search target character cutout graph using coordinate information or the like.
In the above-described embodiment, the similarity of the character identification result is used in the transition score formula (Formula 2 or Formula 3). However, the rank of the character identification result is used, or the line segment direction is determined from the character cutout candidate pattern. A feature amount such as a feature may be extracted and an inner product operation may be performed between the feature amounts.
However, when a character cutout graph is created from a character code, there are no character cutout candidates. For example, a character code and a feature amount of a character cutout candidate extracted from the standard character form of the code are registered in pairs. By preparing a character cutout candidate feature amount dictionary and acquiring and substituting the character cutout candidate feature amount corresponding to the character code from the dictionary, it is possible to calculate a transition score using the character cutout candidate feature amount.

  Applicable to document management system. This is particularly effective in a document management system that manages annotations written by hand on a document consisting of character codes.

Claims (11)

A character string search method in a character string search system having a pattern character string input unit, a document management unit, and a character string collation unit,
In the pattern character string input unit, a pattern character string input step for receiving a pattern character string, which is a character string to be searched, from the user in one or both of two types of character code and handwritten stroke;
Search target that reads the character string to be searched from the document management unit that manages the information of the document having the character string input in one or both of the character code and the handwritten stroke as the search target character string information A character string reading step;
A character string matching step for converting the search target character string and the pattern character string into character cutout graphs, and extracting portions where the pattern character strings appear in the document by matching the character cutout graphs,
And a display step of displaying the appearance location of the extracted pattern character string as a character string search result. The character string search method according to claim 1,
The character string matching method, wherein the character string matching step realizes the character cutout graph matching by dynamic programming. The character string search method according to claim 2,
In the character string collating step, a character identification result is used for a score calculation used in the dynamic programming method. The character string search method according to claim 2 or 3,
A character string search method characterized in that, in the character string matching step, a geometric feature value generated from a character segmentation candidate is used for score calculation used in the dynamic programming. A character string search method according to any one of claims 1 to 4,
Preparing in advance character code graph correspondence information in which a character code and a character cutout graph corresponding to the character code are associated;
In the character string collating step, the character string search method characterized by converting the pattern character string or the search target character string input by a character code into a character cutout graph using the character code graph correspondence information. . A pattern character string input unit for receiving a pattern character string, which is a character string to be searched, in either or both of a character code and a handwritten stroke;
A document management unit for managing information on a document having a character string input in one or both of a character code and a handwritten stroke as information on a search target character string;
A character string matching unit that converts the search target character string and the pattern character string into character cutout graphs, and extracts portions where the pattern character string appears in the document by matching the character cutout graphs,
A character string search system comprising: a display unit configured to display an appearance portion of the extracted pattern character string as a character string search result. The character string search system according to claim 6,
The character string collating unit includes a character string collating unit that realizes character cutout graph collation by dynamic programming. The character string search system according to claim 7,
The character string collating unit uses a character identification result for score calculation used in the dynamic programming method. The character string search system according to claim 7 or 8,
The character string search system, wherein the character string matching unit uses a geometric feature amount generated from a character segmentation candidate for score calculation used in the dynamic programming. The character string search system according to any one of claims 6 to 9,
The character string search system stores in advance character code graph correspondence information in which a character code and a character cutout graph corresponding to the character code are associated with each other,
The character string collating unit converts the pattern character string or the search target character string input by a character code into a character cutout graph using the character code graph correspondence information. . A string search system,
A digital pen that converts input of handwritten characters into stroke information of electronic data and outputs it, an input device that accepts input of stroke information from the digital pen, and a storage device that stores document information including a search target character string And an arithmetic unit,
The input device receives input of stroke information of a handwritten character added to the document or stroke information of a pattern character string which is a character string to be searched from the digital pen, and transmits the input to the arithmetic device,
The arithmetic unit is
The stroke information of the handwritten characters added to the document or the search target character string included in the document stored in the storage device is converted into a search target character cutout graph, and the stroke information or character code of the input pattern character string Convert the pattern character string input in to a pattern character cutout graph,
Further, the search is performed in order to search for a search target character string input in either or both of handwritten characters and character codes by a pattern character string input in either or both of handwritten characters and character codes. A character string search system characterized by using collation between a target character cutout graph and the pattern character cutout graph.

Images