JP5224847B2 - Pattern recognition method, character recognition method, pattern recognition program, and character recognition program - Google Patents

Description

  The present invention relates to a pattern recognition method, a character recognition method, a pattern recognition program, and a character used in, for example, a character recognition device for recognizing characters described on an object to be read or a biometric matching device for recognizing a person based on biometric information. It relates to recognition programs.

Conventionally, various methods have been proposed for pattern recognition methods. For example, JP-A-2004-178280 (Patent Document 1) and JP-A-2001-283156 (Patent Document 2) disclose a pattern recognition method for address information having a hierarchical structure.
Japanese Patent Application Laid-Open Publication No. 2004-178280 discloses a method in which the sum of reliability of each word with respect to recognition result candidates in a certain hierarchy is used as an evaluation value, only the recognition result candidates having the highest E evaluation values are left, and the rest are discarded. Is disclosed.
Japanese Laid-Open Patent Publication No. 2001-283156 discloses a technique for discarding recognition result candidates in which characters in each word are not recognized at a certain rate or more with respect to the word length.

  Also, IEEE Trans. Pattern Analysis and Machine Intelligence, vol.11, no.1, pp.68-83, Jan.1989 (Non-Patent Document 1) uses the value obtained by dividing the posterior probability by the prior probability as the evaluation value. A method is disclosed in which a candidate having the maximum evaluation value is set as the next search target. The posterior probability is expressed as P (c | x). The posterior probability P (c | x) is defined to mean the probability that the candidate c is correct, with the output x obtained by the recognition processing as a condition. The prior probability is written as P (c). The prior probability P (c) is defined as meaning the probability that the candidate c is correct when the recognition process is not performed.

However, with the method disclosed in Japanese Patent Application Laid-Open No. 2004-178280, correct answer candidates may be discarded by mistake. This is because the recognition result candidates are rearranged using an evaluation value based on the sum of the reliability of each word.
Further, the method disclosed in Japanese Patent Laid-Open No. 2001-283156 may be unrealistic in actual operation. This is because if the recognition process for each pattern to be recognized is complex and detailed, a huge amount of processing time is required because all recognition results that are not discarded are also searched.
Also, the method disclosed in IEEE Trans.Pattern Analysis and Machine Intelligence, vol.11, no.1, pp.68-83, Jan.1989 takes time to obtain a correct recognition result. End up. In particular, when the prior probabilities of the patterns to be recognized are biased, it is considered that a huge amount of processing time is required. This is because a value obtained by dividing the posterior probability by the prior probability is used as the evaluation value.
JP 2004-178280 A JP 2001-283156 A `` IEEE Trans. Pattern Analysis and Machine Intelligence, vol.11, no.1, pp.68-83, Jan.1989 ''

  An object of one aspect of the present invention is to provide a pattern recognition method, a character recognition method, a pattern recognition program, and a character recognition program that can efficiently recognize a pattern.

Pattern recognition method according to an embodiment of the invention provides a method for use in an information processing apparatus that performs processing for recognizing a pattern in a plurality of stages, to expand the recognition candidates for the next step belonging to the recognition candidates, the recognition estimating a time required for recognition processing on the recognition candidates subordinate subsequent stage to the candidate, for each recognition candidate deployed, subject to full recognition processing result for the recognition processed recognition candidates was the posterior probability the An evaluation value is calculated based on the estimated time required for recognition processing, a recognition candidate is selected based on the calculated evaluation value for each recognition candidate , and a pattern recognition result is determined from the selected recognition candidate.

A character recognition method according to an aspect of the present invention is a method used in a character recognition device that performs processing for recognizing character information including information of a plurality of layers, and a word candidate of a next layer belonging to a word candidate is determined. Expand, estimate the time required for the recognition process for the next and subsequent recognition candidates subordinate to each word candidate , and for each expanded word candidate, display the result of all character recognition processing for the word recognition processed word candidate. An evaluation value is calculated based on the posterior probability as a condition and the estimated time required for the recognition process, word candidates are selected based on the calculated evaluation value for each word candidate , and the entire character information is selected from the selected word candidates The recognition result of is determined.

A pattern recognition program according to an aspect of the present invention is a program for causing a computer to perform a process of recognizing a pattern at a plurality of stages, and a function of developing a recognition candidate at a next stage belonging to a recognition candidate; A function that estimates the time required for recognition processing for the next and subsequent recognition candidates subordinate to each recognition candidate, and for each expanded recognition candidate, all recognition processing results for recognition candidates that have already been processed are used as conditions. A function for calculating an evaluation value based on the posterior probability and the estimated time required for the recognition process, a function for selecting a recognition candidate based on the calculated evaluation value for each recognition candidate, and pattern recognition from the selected recognition candidate Realize the function to determine the result.

A character recognition program according to an aspect of the present invention is a program for causing a computer to perform recognition processing of character information composed of information of a plurality of layers, and expands a word candidate at a next stage belonging to a word candidate. A function, a function for estimating the time required for recognition processing for recognition candidates for the next and subsequent layers subordinate to each word candidate, and all character recognition processing for word candidates subjected to character recognition processing for each expanded word candidate A function for calculating an evaluation value based on the posterior probability based on the result and the estimated time required for the recognition process, a function for selecting a word candidate based on the calculated evaluation value for each word candidate, and the selected word And a function for determining a recognition result of the entire character information from the candidates.

  According to one aspect of the present invention, it is possible to provide a pattern recognition method, a character recognition method, a pattern recognition program, and a character recognition program that can efficiently recognize a pattern.

Embodiments of the present invention will be described in detail with reference to the drawings.
First, the pattern recognition process as this embodiment will be schematically described.
In pattern recognition processing, it is often the case that high-speed processing is required due to the necessity of processing within a predetermined time. For example, as a typical example of the pattern recognition process, there is a process for recognizing entire character information composed of information of a plurality of layers such as address information. In address information recognition processing, it is often the case that an upper hierarchy with a small number of candidates is recognized first, and candidates of lower hierarchies are narrowed down using the result. In such a pattern recognition process consisting of a plurality of stages (pattern recognition of a multi-stage configuration), in order to speed up the whole process, a recognition process (that is, a processing speed) by a coarse classifier that operates at high speed in the initial stage process. (Recognition processing for selecting candidates with emphasis on). In this case, in the subsequent process, a method of sequentially narrowing down the processing range or processing target using the process result of the previous stage may be taken. Such a problem of selecting a candidate that is a correct answer from a plurality of candidates can be regarded as a search problem.

  That is, in pattern recognition with a multi-stage configuration, it is possible to narrow down the processing range or target of the subsequent stage using the recognition result of the previous stage. Such a method corresponds to performing a beam search. In the pattern recognition processing with a multi-stage configuration described in the present embodiment, a search technique for sequentially narrowing down candidates at each stage based on an evaluation value based on a posteriori probability or the like is applied. As a search method, for example, a method of sequentially narrowing down a predetermined number of candidates in each stage or a method of best priority search is assumed. The posterior probability is written as P (c | x), and means the probability that the candidate c is a correct answer on the condition of the output x obtained by the recognition process. The posterior probability P (c | x) is an approximate value calculated by an approximate calculation method described later. The prior probability is written as P (c) and means the probability that the candidate c is correct in the stage where the recognition process is not performed.

  On the other hand, as a typical example of the pattern recognition process having a plurality of stages, there is a person recognition process using biological information such as a face image, or an address information recognition process including a plurality of layers of information. For example, the address information includes information on multiple layers such as information on the first layer (for example, city names), information on the second layer (for example, town names and street names), and information on the third layer (for example, street addresses and street numbers). Consists of information. The total number of address information configured by combining the information of each layer ranges from several million to several tens of millions. For this reason, it is not efficient to perform recognition processing for all address information. The pattern recognition method described in the present embodiment can be applied to recognition processing of information (for example, address information) composed of information of multiple layers, thereby obtaining recognition results efficiently and at high speed. Is possible.

  Here, as a technology related to word recognition processing, Reference 1 “Lumiyuki Hamamura, Takuma Akagi, Hiroyuki Mizutani, Bunpei Irie:“ Word Matching by Word Length Normalized Bayesian Estimation ”, Image Recognition and Understanding Symposium (MIRU2000) Lecture Proceedings II, pp.1-6 (Jul. 2000).], Reference 2 ("Hamamura Tomoyuki, Akagi Takuma, Irie Fumihei:" Evaluation function using posterior probabilities in word recognition ", IEICE Technical Report, PRMU2006- 92, (Oct. 2006). ”, Reference 3 (“ Lumiyuki Hamamura, Takuma Akagi, Bunpei Irie: “Analytical Word Recognition Using Posterior Probability—Normalization of Number of Characters” —, IEICE Technical Report, PRMU2006 -238, (Mar. 2007) ") discloses a method for calculating a posteriori probability as an evaluation value of a word candidate.

  In the above-mentioned document 1, by using an approximation that the recognition result of each character is independent, the posterior probability ratio of word candidates (ratio of prior probability and posterior probability) can be decomposed into the product of the posterior probability ratio of each character. Is described. Document 2 describes a method for calculating a normalized posterior probability ratio by performing expression expansion that avoids approximation that causes a large error in Document 1 above. Document 3 describes a method of calculating an extended posterior probability ratio using the fact that some characters are written in addition to the place of interest.

  However, the calculation of the normalized posterior probability ratio or the expanded posterior probability ratio described in Document 2 or Document 3 requires processing for all words in the word dictionary. This suggests that the larger the number of words in the word dictionary, the longer it takes to calculate the normalized posterior probability ratio or the expanded posterior probability ratio. On the other hand, in the present embodiment, expression expansion that can apply the idea of Document 3 will be described later even in a situation where only the processing for one word of interest is performed. That is, in the pattern recognition method described in the present embodiment, the posterior probability is calculated by a calculation formula that can be used for a technique that enables efficient candidate search.

Next, the relationship between multi-stage pattern recognition and a search problem will be described.
Here, a face recognition process and an address recognition process will be described as examples of a pattern recognition process having a plurality of stages.
First, an example of face recognition processing will be described.
Face recognition processing is a kind of biometrics. In the face recognition process, an individual is specified based on a person's face image as biometric information. The face recognition process is roughly divided to obtain a recognition result by the following three-stage process. That is, the face recognition process detects a person image from the image (input image) acquired as the first stage process, detects a face image from the person image detected as the second stage process, and performs the third stage. The face image detected as the process is compared with the face image of the registrant.

  The first stage process is a person detection process executed by applying a person detection classifier to the input image. The person detection classifier detects a person image by superimposing a person detection template whose position and size are changed on an input image. This first-stage processing is performed on a low-resolution image, for example, in order to obtain a processing result at high speed.

  The second stage process is a face detection process that is executed by applying a face detection classifier to the person image obtained in the first stage process. The face detection classifier detects a face image by superimposing a face detection template whose position and size are changed on a human image. This second-stage processing is performed on a high-resolution image in order to detect a face with high accuracy, for example.

  The third stage process is a face collation process for collating the face image obtained in the second stage process with the registrant's face image. The third stage process is performed on a high-resolution face image, for example, in order to identify the face with high accuracy.

FIG. 1 represents the process of each stage of the face recognition process by a search tree. Note that the search tree is a conceptual diagram showing a state of search when candidates in a plurality of layers are displayed as nodes. In the search tree shown in FIG. 1, it is assumed that the number of registrants is three.
In the search tree as shown in FIG. 1, each stage candidate is represented as a node. In the first stage of processing, the person detection classifier is run at various positions and sizes to obtain a plurality of person detection results. In the search tree shown in FIG. 1, each node in the first stage corresponds to a plurality of person detection results obtained as candidates in the first stage. That is, each node in the first stage subordinate to the root node in the search tree corresponds to a person detection result detected by the person detection classifier at one position and size. In the second stage process, a plurality of face detection results are obtained for each person detection result (person image) obtained as the first stage process. In the search tree shown in FIG. 1, each node in the second stage corresponds to a plurality of face detection results obtained as a processing result in the second stage. Accordingly, a plurality of second-stage nodes are subordinate to each first-stage node in the search tree. In the third-stage process, a matching result for the registered number of persons is obtained for each face detection result (face image) obtained as the second-stage process. In the search tree shown in FIG. 1, each node in the third stage corresponds to the collation result for each registrant obtained as the processing result in the third stage. Therefore, the second-tier nodes in the search tree are subordinate to the third-tier nodes corresponding to the registered number.

  The parent-child relationship of each node as shown in FIG. 1 means an inclusion relationship. For example, four nodes B to E are subordinate to the node A as one processing result of the first stage shown in FIG. Nodes B to E correspond to four second-stage processing results (face detection results) based on the processing results of node A. The node C is subordinate to three nodes F to H. Node F to node H correspond to three third-stage processing results (matching results with three registrants) based on node C as the second-stage processing results.

  In face recognition processing, it is ideal if all face recognition processing corresponding to all nodes in the third stage can be executed. However, in order to increase the processing speed, it is necessary to efficiently perform the third stage process (collation process) using the first stage process result and the second stage process result. This corresponds to a solution to the search problem of efficiently searching a search tree as shown in FIG.

Next, an example of address information recognition processing will be described.
FIG. 2 is a diagram illustrating an example of an address database. Here, it is assumed that the address information to be recognized is composed of information of a plurality of stages. In the example shown in FIG. 2, the address information to be recognized includes a city name in the first row, a town name in the second row, and an address in the third row.

  FIG. 3 is a diagram illustrating a notation example of address information. FIG. 3 shows an example of an image given as an input image for recognition processing. For example, an image as shown in FIG. 3 is information obtained by reading a medium on which address information is described with a scanner or the like. For an image including address information as shown in FIG. 3, line candidate detection, word candidate extraction, character candidate extraction, individual character recognition for each character candidate, and the like are executed. When individual character recognition results are obtained, a plurality of word candidates are selected based on the individual character recognition results. In the following description, processing for matching each word candidate and each word in the address database will be described.

  FIG. 4 is a diagram illustrating an example of word candidates obtained from the address information image illustrated in FIG. 3. In the example shown in FIG. 4, candidates I1 to I10 are detected as word candidates. FIG. 5 shows the matching process using a search tree. That is, a pair of the word candidate Ii and a word in the address database corresponds to one node of the search tree, and matching one set corresponds to searching for one node.

  As address information recognition processing, it is ideal to match all words in the address database with all word candidates. However, the words stored in the address database are enormous. For this reason, in the pattern recognition method having a multi-stage configuration, a matching process is performed from an upper layer to efficiently search for a solution (determine the recognition result of the entire address information).

  In the multi-stage pattern recognition as described above, a procedure is generally performed in which processing is performed in order starting from the processing of the upper level, and after narrowing down candidates for each level, processing of the next level is performed. For example, in the example of FIG. 1, all processes are performed on the first-stage nodes, and the number of nodes is narrowed down to n. Next, all the nodes in the second stage subordinate to the nodes in the first stage that have been narrowed down are processed, and the number of nodes in the second stage is narrowed down to n. Further, all the nodes in the third stage subordinate to the narrowed-down nodes in the second stage are processed, and the optimum recognition result is determined from the nodes in the third stage. When such a series of processing flows is viewed from the viewpoint of a search problem, it corresponds to performing a beam search.

  On the other hand, a method called best priority search is known as a solution to an excellent search problem. The best priority search method is a method in which all unprocessed child nodes subordinate to a processed node are stored, and the node with the highest evaluation value is selected as the next search target. . For example, in the example shown in FIG. 1, assuming that node A and node C have been processed and other nodes are unprocessed, the unprocessed child nodes of node A and node C are node B, There are six nodes D, E, F, G, and H. In this case, the next search target is selected from the six nodes B, D, E, F, G, and H. Even in the pattern recognition of a multi-stage configuration, if the best priority search as described above can be used, it is considered that the search efficiency can be improved.

  Generally, in beam search, it is only necessary to compare between nodes in the same stage. For this reason, for example, the degree of similarity output by the classifier at that stage may be used as the evaluation value. However, the best priority search requires comparison between nodes at different stages. For this reason, in the best priority search, it is meaningless to use the output of the discriminator as it is or to compare it. That is, the evaluation value used in the existing beam search cannot be used for the best priority search.

  Therefore, in the present embodiment, a pattern recognition method using the posterior probability as an evaluation value is provided. In general, it is difficult to directly calculate the posterior probability defined in Non-Patent Document 3 and the like described above. However, according to the methods described in Documents 1, 2, and 3, the posterior probability can be calculated approximately from the output of individual character recognition, for example, in word recognition. In the barcode recognition process of the present embodiment, an approximate posterior probability as an evaluation value is calculated by applying the calculation techniques described in the above-mentioned documents 1, 2, and 3.

Next, a method for calculating the posterior probability as an evaluation value used for the best priority search will be described.
Here, the processing result corresponding to the node ni is assumed to be xi. For example, when applying the best priority search to face recognition recognition, the output of the classifier at each stage corresponding to each node ni corresponds to the processing result xi. When applying the best priority search to address information recognition processing, the result of character recognition in each word candidate at each stage corresponding to each node ni corresponds to the processing result xi. In the following description, X, ni, Ui, and Xetc are defined as follows. Let X be all processing results corresponding to all processed nodes. Let Ui be the set of nodes that can be reached by following the parent node of ni. Assume that ni is not an element of the set Ui. Of the nodes not included in Ui, all processing results of processed nodes are set as Xetc.

FIG. 6 is a diagram illustrating the state of each node represented by the search tree. In FIG. 6, a black circle is a processed node, a white circle and a double circle are unprocessed nodes, and a double circle is a node having a processed node as a parent. That is, in the example shown in FIG. 6, a node represented by a double circle is a candidate for the next search target. For each of these nodes, a posterior probability P (ni | X) is calculated. Here, the posterior probability P (ni | X) means not only the node ni but also the probability that all nodes included in Ui are satisfied. For example, in the example of address information recognition processing, the address is written because it is not the probability that a word in a single hierarchy (town, etc.) is written, but the word in a higher hierarchy is also written at the same time. It corresponds to the probability. In the example of face recognition processing, if the child node is satisfied, the parent node is automatically satisfied.

  Equation (2) uses an approximation that the processing result of each node included in Ui occurs independently of other processing results. In equation (3), an approximation of P (xj | ni) ≈P (xj | nj) is used. By using Expression (3), the posterior probability P (ni | X) of the node ni can be calculated approximately. In the example of face recognition processing, xj is the output of a single classifier. Therefore, by collecting data, P (xj | nj) and P (xj) can be easily obtained, and the posterior probability can be calculated. On the other hand, in the example of address information recognition processing, xj includes a plurality of character recognition results in word candidates. For this reason, the calculation of Formula (3) is not simple. The posterior probability calculation method in the address information recognition process will be described later in detail.

Equation (3) further defines

And can be transformed. If the ratio between the prior probability and the posterior probability is called the posterior probability ratio, it can be expressed as “the posterior probability ratio of the node ni is the product of the posterior probability ratios of its parent node”.

Next, a method for calculating the posterior probability ratio in the address information recognition process will be described.
In the address information recognition process, xj includes a plurality of character recognition results in word candidates. For this reason, it is often not easy to calculate the posterior probability ratio P (xj | nj) / P (xj) in equation (3). In the above documents 1, 2, and 3, several formulas for calculating the posterior probability ratio are proposed. They are organized by paying attention to the denominator P (xj) of the posterior probability ratio. First, in Document 1, all character recognition results of each character candidate included in xj are approximated as independent. This may cause a large error and may make calculation difficult due to a remaining character candidate structure (described later) or a path selection term. The normalized posterior probability ratio described in Document 2 is a large error because the denominator calculation is expanded for all words in the word dictionary by transforming the denominator to P (xj) = Σ _k P (xj, wk). It has succeeded in avoiding the approximate deformation that causes At the same time, we succeeded in successfully canceling difficult-to-calculate terms with the denominator and numerator. Furthermore, in the extended posterior probability ratio described in Document 3, the approximation error is reduced by considering that some characters are written in places other than the word candidate of interest. Considering that some characters are written, terms that reduce the convenience of calculation occur. However, as with the normalized posterior probability ratio, the denominator calculation is expanded with all the words in the word dictionary, so that the difficult-to-calculate terms are canceled with the denominator and the numerator.

  However, in order to be able to use the posterior probability ratio at the time of searching, it is desirable that calculation is possible by processing only one word of interest without processing all words. Therefore, in the present embodiment, it is assumed that the calculation of the denominator is not expanded to all the words in the word dictionary and “some character is written” by the following expression expansion. This shows that there is an advantage of the extended posterior probability ratio, that is, there is an advantage that the approximation accuracy is higher than when all the character recognition results are independent, and a term that is difficult to calculate can also be canceled.

  The word candidate corresponding to the node ni is Li, and the word is wi. Searching for node ni corresponds to matching word wi with word candidate Li. Ai is a set of all character candidates in the word candidate Li, and each character candidate is aεAi.

  FIG. 7 is a diagram illustrating an example of character candidates. The structure shown in FIG. 7 is called a lattice structure. That is, here, it is assumed that the character candidates have a lattice structure as shown in FIG. In the following description, ai, ri, Li, Si, xi, wi, and the like are defined as follows. Let ri be the character recognition result of the character candidate ai. Assume that all character recognition results in the word candidate Li are ri and the character candidate structure is Si. The character candidate structure refers to information other than character recognition results such as adjacent information between character candidates or the number of character candidates. The processing result xi described above is defined as xi = (ri, Si). The jth character of the word wi is assumed to be cijεC (C is a set of alphabets). Let C * be an arbitrary character string.

Assume that the set of all routes from the left end to the right end in the word candidate Li is Fi = {fp}, p = 1, 2,..., The route is fp = (af1p, af2p,...), And afjpεAi. af (j + 1) p is located adjacent to the right side of afjp. FIG. 7 shows an example of the route fp with a bold line. A set of character candidates on the path fp is E′p = {afjp}, j = 1, 2,..., E′p but a set of character candidates included in Ai is Ep. Ep∩E′p = φ and Ep∪E′p = Ai. The posterior probability ratio P (xj | nj) / P (xj) is modified as follows.

The numerator P (xj | nj) on the left side of the equation (4) is P (xj | Lj, wj, Lp1, wp1, Lp2, wp2,..., Where the nodes obtained by tracing its parent node are np1, np2,. ). When approximation is performed that xj is not affected by information other than information related to the word candidate Lj, P (xj | Lj, wj) is obtained.
In the denominator of Expression (4), P (xj) ≈P (xj | Lj, C *) is an approximation that some character is written in any word candidate as described above. Because.
Equation (6) uses an approximation that the probability of other routes is negligible compared to the probability of the route having the maximum probability. Subsequently, the following approximation is performed.

  Here, Kj is a constant not depending on p. This approximation is an approximation that the probability that any path is correct is equally likely.

By using the equations (7) and (8), the equation (6) is calculated as follows. (If the length does not coincide with the word wj in any fp, the expression (6) becomes 0. The following calculation is performed in other cases.)

However, match () is defined as follows.

Equation (10) uses an approximation that the character recognition results are independent of each other. In the transformation from Equation (10) to Equation (11), a value that does not depend on fp

The denominator is divided by Note that in Equation (9), terms relating to the character candidate structure Sj and the path fp that are difficult to calculate can be canceled.

  Equation (11) is a formula for calculating the posterior probability ratio proposed in the present embodiment. In equation (11), the approximation accuracy is improved by the same idea as the extended posterior probability ratio, and an extra term can be canceled. Furthermore, it can be calculated by processing only the one word wj of interest, and can also be used at the time of search.

  As described above, in the pattern recognition method according to the present embodiment, a part of pattern recognition processing including a plurality of stages is regarded as a search problem. In the pattern recognition method, each candidate obtained at each stage is set as a node, and the obtained node is selectively processed based on the posterior probability. Thereby, in the said pattern recognition method, the pattern recognition process which consists of multiple steps can be performed efficiently and at high speed.

  In the pattern recognition method, when selecting a node at each stage to be processed, a technique such as best priority search can be applied. In the best priority search, it is necessary to compare nodes at different stages. As an evaluation value for that, the posterior probability is used in the pattern recognition method. The posterior probability is calculated by the product of the posterior probability ratio (ratio of prior probability and posterior probability) at each node. In particular, when specializing in address information recognition, the posterior probability ratio of each node is calculated in a form that can be calculated even during a search.

Hereinafter, application examples of the pattern recognition method as described above will be described.
FIG. 8 is a diagram illustrating a configuration example of the information processing apparatus 11 having a pattern recognition function according to the pattern recognition method as described above.
In the example illustrated in FIG. 8, an image input device 12 is connected to the information processing device 11. The image input device 12 is configured by a scanner or a camera. The image input device 12 acquires an image to be processed by the information processing device 11. For example, the image input device 12 includes a scanner or a camera. The image input device 12 supplies the acquired image information to the information processing device 11. The image input device 12 may read the image information stored in the recording medium and supply the image information read from the recording medium to the information processing device 11.

  The information processing apparatus 11 functions as a pattern recognition apparatus or a character recognition apparatus. The information processing apparatus 11 recognizes desired information (for example, address information or a face image) included in the image supplied from the image input apparatus 12 by pattern recognition processing.

  The information processing apparatus 11 is realized as a computer, for example. In the configuration example shown in FIG. 8, the information processing apparatus 11 includes an image interface (I / F) 21, a processor 22, a working memory 23, a program memory 24, a data memory 25, an output interface (I / F) 26, and the like. have. That is, the information processing apparatus 11 includes, for example, a data input / output unit as the image interface 21 and the output interface 26, a control unit as the processor 22, a variety of storage such as the working memory 23, the program memory 24, and the data memory 25. This is realized by a computer having a unit.

The image interface 21 is an interface for capturing an image supplied from the image input device 12. That is, the image interface 21 is an interface for acquiring an image to be subjected to pattern recognition processing.
FIG. 9 is a diagram showing an example of an image to be subjected to pattern recognition processing that is captured by the image interface 21. The example illustrated in FIG. 9 is an example of a read image of a paper sheet in which address information including information of a plurality of layers is described. In the example illustrated in FIG. 9, an example of an image that is a target of address information recognition processing including information of a plurality of layers as pattern recognition processing is illustrated.

  The processor 22 executes various processing functions in the information processing apparatus 11. The processor 22 is constituted by an arithmetic unit such as a CPU, for example. The processor 22 implements various processing functions by executing programs stored in the program memory 24 or the data memory 25. For example, the processor 22 includes a pattern recognition unit 22a that performs pattern recognition processing as one of functions realized by executing a program. A configuration example of the pattern recognition unit 22a will be described in detail later.

  The working memory 23 is a storage unit for temporarily storing data. The working memory 23 is composed of, for example, a RAM (random access memory). The program memory 24 is a storage unit that stores a control program, control data, and the like. The program memory 24 is composed of, for example, a ROM (read only memory). The data memory 25 is a large-capacity storage unit for storing data. The data memory 25 is constituted by, for example, a hard disk drive (HDD).

  The data memory 25 is provided with a dictionary database 25a used for pattern recognition processing. For example, when the information processing apparatus 11 recognizes address information as a pattern recognition process, the dictionary database 25a is configured as an address database in which address information is stored. When the information processing apparatus 11 performs personal authentication using biometric information such as a face image as a pattern recognition process, the dictionary database 25a is configured as a biometric information database in which registrant's biometric information is stored. Here, it is assumed that the information processing apparatus 11 recognizes address information. Therefore, it is assumed that the dictionary database 25a is an address database.

  FIG. 10 is a diagram illustrating a configuration example of an address database as the dictionary database 25a. In the configuration example shown in FIG. 10, the dictionary database (address database) 25a stores address information including words of a plurality of layers (CITY layer, STREET layer, DIRECTION layer). That is, the dictionary database 25a stores the information of each lower layer so that the information of the next lower layer is subordinate to the information of each layer.

  In the example illustrated in FIG. 10, words such as “STOCKHOLM” (word D1), “GOTEBORG” (word D2), “ABCDE” (word D3),... Exist in the CITY hierarchy, and “STREET hierarchy includes“ There are words such as “AGATAN” (word D4), “TOSHIBA” (word D5), “BGATAN” (word D6),..., And words such as “EAST”, “WEST”, “NORTH”, etc. in the DIRECTION hierarchy. Is present. Further, in the example shown in FIG. 10, one word “STOCKHOLM” (word D1) in the CITY hierarchy is subordinate to two words “AGATAN” (word D4) and “TOSHIBA” (word D5) in the STREET hierarchy. doing.

  The output interface 26 is an interface for outputting information obtained by the processor 22 to the outside. For example, the recognition result obtained by the pattern recognition process in the processor 22 is output to the outside by the output interface 26.

Next, the configuration of the pattern recognition unit 22a will be described.
FIG. 11 is a diagram illustrating a configuration example of the pattern recognition unit 22a. In the following description, it is mainly assumed that the information processing apparatus 11 recognizes character information including information of multiple layers such as address information.
The pattern recognition unit 22a includes a recognition control unit 30, a candidate extraction unit 31, a node expansion unit 32, an evaluation value calculation unit 33, a node selection unit 34, a determination unit 35, and the like. As described above, the pattern recognition unit 22a is a function realized by the processor 22 executing a program. That is, the candidate extraction unit 31, the node expansion unit 32, the evaluation value calculation unit 33, the node selection unit 34, and the determination unit 35 are also functions realized by the processor 22 executing the program.

The recognition control unit 30 controls a function of controlling the entire pattern recognition process in the pattern recognition unit 22a. The candidate extraction unit 31 extracts information that is a candidate for each layer in the recognition result from the input image supplied from the image input device 12 via the image interface 21. For example, when the information processing apparatus 11 recognizes character information including information of a plurality of layers such as address information, the candidate extraction unit 31 performs a process of extracting word candidates of each layer from the input image.
12 is a diagram showing an example of word candidates extracted from the input image shown in FIG. In the example illustrated in FIG. 12, alphabets are assumed as characters constituting the address information to be recognized. For this reason, in the example shown in FIG. 12, seven word candidates are extracted. It is assumed that the seven word candidates shown in FIG. 12 exist at positions P1 to P7 indicated by position information in the image.

  The node expansion unit 32 generates each node for constructing a search tree for the candidates extracted by the candidate extraction unit 31. The node expansion unit 32 is a process for obtaining a next hierarchical node belonging to each node. That is, the node expansion unit 32 generates a search tree composed of nodes in a plurality of hierarchies by selecting all candidates that can be candidates for the next hierarchy for each candidate in a certain hierarchy.

  For example, a node indicating that the word candidate at the position P6 shown in FIG. 12 is the word D1 in the address database 25a shown in FIG. 10 is represented as (D1, P6). Here, in the address database 25a shown in FIG. 10, the word D1 (“STOCKHOLM” in the “CITY” hierarchy) includes the word D4 (“AGATAN” in the “STREET” hierarchy) and the word D5 (“TOSHIBA” in the “STREET” hierarchy). ]) Belongs to. If the position P6 is information on the CITY hierarchy, the STREET hierarchy is either the position P7 or the position P3 in the input image shown in FIG. It can be discriminated by the rules of notation. According to these situations, the node expansion unit 32 has four nodes (D4, P7), (D4, P3), (D5, P7), and (D5, P3) as nodes belonging to the node (D1, P6). Expand.

  The evaluation value calculation unit 33 calculates the evaluation value of each node generated by the node expansion unit 32. For example, the evaluation value calculation unit 33 calculates the evaluation value of each node by executing recognition processing for each candidate as each node. In the present embodiment, the evaluation value calculation unit 33 calculates the posterior probability calculated by the above-described method as the evaluation value.

  The node selection unit 34 selects a node to be finally evaluated among the nodes. Based on the evaluation value calculated by the evaluation value calculation unit 33, the node selection unit 34 determines whether or not each node is to be a node to be finally evaluated. For example, the node selection unit 34 selects a predetermined number (ie, top N) nodes in descending order of evaluation values calculated by the evaluation value calculation unit 33 for each layer. Further, when the best priority search is applied as the search method, the node selection unit 34 selects the one having the highest evaluation value from a plurality of terminal nodes to be described later regardless of each hierarchy.

  The determination unit 35 determines a final recognition result as a series of pattern recognition processes. When the node of the final hierarchy is selected by the node selection unit 34, the determination unit 35 determines the final recognition result based on the nodes of the final hierarchy. For example, when a plurality of nodes in the final hierarchy are obtained by the node selection unit 34, the determination unit 35 outputs a recognition result based on the node having the maximum evaluation value as a final recognition result. In addition, the determination unit 35 finally determines each recognition result (one or more recognition results) based on each node having an evaluation value equal to or higher than a predetermined value among the nodes in the final hierarchy obtained by the node selection unit 34. May be output as a recognition result. Furthermore, when the evaluation value of the node in the final hierarchy obtained by the node selection unit 34 is less than a predetermined value, the determination unit 35 outputs that the pattern cannot be recognized as a final recognition result. You may do it.

Next, first, second, and third processing examples of pattern recognition processing in the information processing apparatus 11 configured as described above will be described.
First, a first processing example of pattern recognition processing in the information processing apparatus 11 will be described.
FIG. 13 is a flowchart for explaining the flow of processing as a first processing example of pattern recognition processing.
First, an image to be subjected to pattern recognition processing supplied from the image input device 12 is taken into the information processing device 11 by the image interface 21 (step S10). When an image to be subjected to pattern recognition processing is captured by the image interface 21, the processor 22 starts pattern recognition processing by the pattern recognition unit 22a. That is, the recognition control unit 30 of the pattern recognition unit 22a first executes a process of extracting word candidates from the input image by the candidate extraction unit 31 (step S11). For example, when an image as shown in FIG. 9 is given, the candidate extraction unit 31 extracts word candidates as shown in FIG. At this time, the candidate extraction unit 31 provides identification information for identifying each extracted word candidate and specifies information indicating the position of each word candidate.

  When word candidates are extracted from the input image, the recognition control unit 30 starts a search process for each extracted word candidate. First, the recognition control unit 30 performs processing for setting a root node of a search tree. This corresponds to starting generation of a search tree on the buffer. That is, the recognition control unit 30 stores the root node in a buffer provided in the working memory 23 or the like (step S12). When the root node is set, the recognition control unit 30 sets “L = 1” as an initial value to the variable L indicating the hierarchy being processed (step S13).

  When the variable L is set, the recognition control unit 30 performs processing for generating each node of the L hierarchy and calculating an evaluation value of each node (steps S14 to S17). That is, the recognition control unit 30 takes out one node already stored in the buffer (step S14). For example, when L = 1, the recognition control unit 30 extracts the root node from the buffer. In addition, when L = 2, the recognition control unit 30 sequentially extracts each node of the first hierarchy stored in the buffer.

  When one node is extracted from the buffer (step S14), the recognition control unit 30 obtains each node of the next lower layer (L layer) belonging to the node extracted by the node expansion unit 32. Is performed (step S15). For example, when L = 1, the node expansion unit 32 stores, in the buffer, each node corresponding to the first layer candidate among the candidates extracted by the candidate extraction unit 31 as each node subordinate to the root node. . Further, in the example of extracting word candidates shown in FIG. 12, when the node extracted in step 14 is (D1, P6), the node expansion unit 32, as described above, (D4, P7). , (D4, P3), (D5, P7), and (D5, P3).

  When each node of the L layer belonging to the node extracted by the node expansion unit 32 is obtained, the evaluation value calculation unit 33 calculates an evaluation value for each obtained node (step S16). Here, it is assumed that the posterior probability approximately obtained by the recognition processing for each node is calculated as the evaluation value by the above-described calculation formula.

  When the evaluation value of each node developed by the evaluation value calculation unit 33 is calculated, the recognition control unit 30 determines whether or not an unprocessed node exists in the buffer (step S17). That is, the recognition control unit 30 determines whether or not there is an unprocessed node among nodes that can become a parent node with respect to a node in the L hierarchy. When it is determined that there is an unprocessed node in the buffer (step S17, NO), the recognition control unit 30 returns to step S14 and takes out the next node from the buffer, thereby performing the processes in steps S14 to S16. Repeatedly.

  If it is determined that there is no unprocessed node in the buffer (step S17, YES), the recognition control unit 30 is ranked in the order of the evaluation value of each node calculated in step 16 by the node selection unit 34. N nodes are selected and stored in the buffer (step S18). That is, in the process of step S18, each node selected by the node selection unit 34 (node with the highest evaluation value) is obtained as an L-layer node (that is, an L-layer candidate). Here, the number of nodes selected by the node selection unit 34 in the order of evaluation values is set to a value that does not cause a candidate to be correct to be discarded. However, as the number of nodes to be selected increases, the processing speed decreases. Therefore, the number of nodes to be selected should be appropriately set according to the operation mode such as the nature of the pattern to be recognized and the required processing time.

  When the top N nodes are stored in the buffer as nodes in the L layer, the recognition control unit 30 determines whether the L layer is the final layer in the pattern recognition. If it is determined that the L layer is not the final layer (NO in step S19), the recognition control unit 30 increments the variable L by updating the variable L to “L = L + 1” (step S20). ). When the change L is incremented, the recognition control unit 30 proceeds to step S14 and performs processing for each updated node in the L hierarchy.

  If it is determined by the above determination that the L layer is the last layer (step S19, YES), the recognition control unit 30 performs a process of determining a final recognition result by the determination unit 35 (step S21). . In this case, the determination unit 35 determines the final recognition result of the pattern recognition process based on the evaluation value of each node. For example, when evaluating the evaluation value of each node based on a predetermined threshold, the determination unit 35 outputs each candidate specified by each node having an evaluation value equal to or higher than the predetermined threshold as a recognition result. When the recognition result of the pattern recognition process is uniquely determined, the determination unit 35 recognizes the candidate specified by the node that is the maximum evaluation value (or the node that is the predetermined evaluation threshold and the maximum evaluation value). Output as.

  As described above, in the first process example, in the pattern recognition process including a plurality of stages, the search is performed by narrowing down the evaluation value to the top N nodes for each layer. Thus, according to the first processing example, it is unlikely that the correct answer candidate is erroneously discarded, and it becomes possible to select a candidate that is correct at high speed.

Next, a second processing example of the pattern recognition processing in the information processing apparatus 11 will be described.
The second processing example is an example of pattern recognition processing in which the best priority search is applied as a method for searching for a candidate that is a correct answer from a plurality of hierarchical candidates. As described above, the best priority search is a search method that preferentially processes a node having the highest evaluation value as a comparison target even for nodes in different hierarchies. Here, it is assumed that the evaluation value of each node is calculated by the calculation method as described above.

  FIG. 14 is a flowchart for explaining the flow of processing as a second processing example of pattern recognition processing in the information processing apparatus 11.

  First, an image to be subjected to pattern recognition processing supplied from the image input device 12 is taken into the information processing device 11 by the image interface 21 (step S30). When an image to be subjected to pattern recognition processing is captured by the image interface 21, the processor 22 starts pattern recognition processing by the pattern recognition unit 22a. That is, the recognition control unit 30 of the pattern recognition unit 22a first executes a process of extracting word candidates from the input image by the candidate extraction unit 31 (step S31).

  When word candidates are extracted from the input image, the recognition control unit 30 starts a search process for each extracted word candidate. First, the recognition control unit 30 performs processing for setting a root node of a search tree. This corresponds to starting generation of a search tree on the buffer. That is, the recognition control unit 30 stores the root node in a buffer such as the working memory 23 (step S32). When the root node is set, the recognition control unit 30 performs a process of sequentially searching for a node having the maximum evaluation value.

  That is, the recognition control unit 30 takes out one node having the maximum evaluation value from the nodes already stored in the buffer (step S33). In the state where only the root node is stored in the buffer, the recognition control unit 30 takes out the root node from the buffer. When a plurality of nodes are stored in the buffer, the recognition control unit 30 takes out the node having the maximum evaluation value from each node stored in the buffer regardless of the hierarchy of each node.

  When one node is extracted from the buffer, the recognition control unit 30 determines whether or not the extracted node is a terminal node (step S34). Here, the term “end node” refers to a node having no subordinate node. That is, the terminal node is a node that is a terminal in a search tree composed of root nodes. For example, FIG. 15 is a conceptual diagram illustrating a configuration example of a search tree. In the example shown in FIG. 15, each white circle and black circle are nodes indicating each state in the search. Further, in the example shown in FIG. 15, black circles indicate terminal nodes that have no lower-level nodes (no subordinate nodes).

  If it is determined by the above determination that the node is not a terminal node (NO in step S34), the recognition control unit 30 performs node expansion processing on the node extracted by the node expansion unit 32 (step S35). As described above, the node expansion process is a process for obtaining each node in the next lower hierarchy (L hierarchy) belonging to the extracted node. When the root node is extracted, the node expansion unit 32 buffers each node corresponding to the first layer candidate among the candidates extracted by the candidate extraction unit 31 as each node subordinate to the root node. Store.

  When each node belonging to the node extracted by the node expansion unit 32 is obtained, the evaluation value calculation unit 33 calculates an evaluation value for each obtained node (step S36). Here, it is assumed that the posterior probability approximately obtained by the recognition process for each node is calculated as the evaluation value by the above-described calculation method.

  When the evaluation value of each node developed by the evaluation value calculation unit 33 is calculated, the recognition control unit 30 associates each node with each evaluation value and stores them in the buffer (step S37). When the obtained nodes and evaluation values are stored in the buffer, the recognition control unit 30 returns to step S33 and repeatedly executes the above-described processing. The processes in steps S33 to S37 are repeatedly executed until the node extracted in step S33 is determined to be a terminal node. As a result, a search tree is obtained until the node having the maximum evaluation value reaches the terminal node.

  That is, when it is determined that the node extracted in step S34 is a terminal node (YES in step S34), the recognition control unit 30 determines whether the evaluation value of the node is equal to or greater than a predetermined threshold value. The determination is made by the unit 35 (step S38). If it is determined by this determination that the evaluation value of the node is less than the predetermined threshold (NO in step S38), the recognition control unit 30 determines whether there is an unprocessed node other than the node in the buffer. Is determined (step S39).

  If it is determined that there is an unprocessed node in the buffer (step S39, NO), the recognition control unit 30 returns to step S33, and extracts a node having the maximum evaluation value from the unprocessed nodes other than the node. As a result, the processes of steps S33 to S37 are executed. If it is determined that there is no unprocessed node in the buffer (step S39, YES), the recognition control unit 30 determines that no candidate whose evaluation value is equal to or greater than a predetermined threshold has not been obtained. finish.

  Further, when it is determined by the determination in step S38 that the evaluation value of the node is equal to or greater than a predetermined threshold (YES in step S38), the recognition control unit 30 is specified by the node by the determination unit 35. The pattern indicated by each layer candidate is output as the final recognition result (step S40). As a second processing example, all nodes having a predetermined threshold value or more may be output as the final recognition result. This is because even if the final recognition result is obtained in step S40, the process proceeds to step S39, and the processes after step S33 are executed on the unprocessed nodes existing in the buffer. It is feasible. As a result, a plurality of recognition results whose node evaluation values are equal to or greater than a predetermined threshold can be output as final recognition results.

As described above, in the second processing example, in the pattern recognition processing including a plurality of stages, the candidates as recognition results are narrowed down by the best priority search, and based on the candidates whose evaluation values are equal to or higher than a predetermined threshold among the narrowed candidates. As a result, the final recognition result is obtained. Thus, according to the second processing example, recognition result candidates can be narrowed down efficiently, and a final recognition result can be obtained from the narrowed down candidates.
Further, it is necessary for the best priority search applied to the second processing example. For this reason, in the second processing example, the posterior probability approximately calculated by the above-described calculation formula is used as the evaluation value. As a result, according to the second processing example, it is possible to compare the nodes at different stages, and the best priority search can be realized.

Next, a third processing example of the pattern recognition processing in the information processing apparatus 11 will be described.
Similar to the second processing example, the third processing example is an example of pattern recognition processing in which the best priority search is applied as a method for searching for a candidate that is a correct answer from a plurality of hierarchical candidates. The third processing example is a modification of the second processing example. In the third processing example, the value obtained by dividing the posterior probability of each node by the estimated processing time is used as the evaluation value of each node.

  FIG. 16 is a flowchart for explaining the flow of processing as a third processing example of the pattern recognition processing in the information processing apparatus 11. Note that steps S50 to S60 shown in FIG. 16 are equivalent to steps S30 to S40 shown in FIG. 14 described as the second processing example.

  First, an image to be subjected to pattern recognition processing supplied from the image input device 12 is taken into the information processing device 11 by the image interface 21 (step S50). When an image to be subjected to pattern recognition processing is captured by the image interface 21, the processor 22 starts pattern recognition processing by the pattern recognition unit 22a. That is, the recognition control unit 30 of the pattern recognition unit 22a first executes a process of extracting word candidates from the input image by the candidate extraction unit 31 (step S51).

  When word candidates are extracted from the input image, the recognition control unit 30 starts a search process for each extracted word candidate. First, the recognition control unit 30 performs processing for setting a root node of a search tree. This corresponds to starting generation of a search tree on the buffer. That is, the recognition control unit 30 stores the root node in a buffer such as the working memory 23 (step S52). When the root node is set, the recognition control unit 30 performs a process of sequentially searching for a node having the maximum evaluation value.

  That is, the recognition control unit 30 takes out one node having the maximum evaluation value from the nodes already stored in the buffer (step S53). When one node is extracted from the buffer, the recognition control unit 30 determines whether or not the extracted node is a terminal node (step S54). If it is determined by the above determination that the node is not a terminal node (NO in step S54), the recognition control unit 30 performs node expansion processing on the node extracted by the node expansion unit 32 (step S55).

When each node belonging to the node extracted by the node expansion unit 32 is obtained, the evaluation value calculation unit 33 calculates an evaluation value for each obtained node (step S56). In the third processing example, a value obtained by dividing the posterior probability by the estimated processing time is calculated as the evaluation value.
That is, the evaluation value calculation unit 33 first calculates a posteriori probability for each node obtained by the node expansion unit 32 (step S61). The posterior probability of each node is approximately calculated by the recognition process for each node by the calculation method described above.

When the posterior probability of each node is calculated, the evaluation value calculation unit 33 performs a process of estimating the time required for the process related to each node (step S62). Here, it is assumed that the processing time required for the recognition processing of the lower hierarchy to which each node belongs is estimated. For example, a method of assuming such processing time based on the total number of characters of words in a lower layer can be considered.
When the posterior probability and estimated processing time of each node are obtained, the evaluation value calculation unit 33 divides the posterior probability calculated in step S61 for each node by the estimated processing time calculated in step S62. The value is calculated as an evaluation value for each node (step S63).

  When the evaluation value of each node obtained by the evaluation value calculation unit 33 is calculated by such a method, the recognition control unit 30 associates each node with each evaluation value and stores them in the buffer. (Step S57). When the obtained nodes and evaluation values are stored in the buffer, the recognition control unit 30 returns to step S53 and repeatedly executes the above-described processing.

  That is, when it is determined that the node extracted in step S54 is a terminal node (YES in step S54), the recognition control unit 30 determines whether or not the evaluation value of the node is equal to or greater than a predetermined threshold value. The determination is made by the unit 35 (step S58). When it is determined by this determination that the evaluation value of the node is less than the predetermined threshold (NO in step S58), the recognition control unit 30 determines whether there is an unprocessed node other than the node in the buffer. Is determined (step S59).

  If it is determined that there is an unprocessed node in the buffer (step S59, NO), the recognition control unit 30 returns to step S53, and extracts a node having the maximum evaluation value from the unprocessed nodes other than the node. As a result, the processes of steps S53 to S57 are executed. If it is determined that there is no unprocessed node in the buffer (step S59, YES), the recognition control unit 30 determines that no candidate whose evaluation value is equal to or greater than a predetermined threshold has not been obtained. finish.

  If it is determined in step S58 that the evaluation value of the node is greater than or equal to a predetermined threshold (YES in step S58), the recognition control unit 30 is identified by the node by the determination unit 35. The pattern indicated by each layer candidate is output as the final recognition result (step S60). As a third processing example, all nodes having a predetermined threshold value or more may be output as a final recognition result. This is because even when the final recognition result is obtained in step S60, the process proceeds to step S59, and the processes after step S53 are executed on the unprocessed nodes existing in the buffer. It is feasible. As a result, a plurality of recognition results whose node evaluation values are equal to or greater than a predetermined threshold can be output as final recognition results.

  As described above, in the third processing example, in the pattern recognition processing including a plurality of stages, a value obtained by dividing the posterior probability for each candidate by the estimation processing time is used as an evaluation value, and candidates as recognition results are narrowed down by the best priority search. The final recognition result is obtained based on candidates that have an evaluation value equal to or higher than a predetermined threshold among the narrowed candidates. As a result, according to the third processing example, candidates for recognition results can be narrowed down efficiently based on an evaluation value that takes processing time into account, and a final recognition result can be obtained from the narrowed candidates.

The process of each step in the face recognition process according to the pattern recognition method of the present embodiment is expressed by a search tree. It is a figure which shows the example of an address database. It is a figure which shows the example of a description of address information. It is a figure which shows the example of the word candidate obtained from the image of the address information shown in FIG. It is a conceptual diagram which shows a matching process with a search tree. It is a figure which shows the state of each node represented by a search tree. It is a figure which shows the example of a character candidate. It is a figure which shows the structural example of the information processing apparatus which has a pattern recognition function. It is a figure which shows the example of the image used as the object of a pattern recognition process. It is a figure which shows the structural example of the address database as a dictionary database. It is a figure which shows the structural example of a pattern recognition part. It is a figure which shows the example of the word candidate extracted from the input image shown in FIG. It is a flowchart for demonstrating the flow of the process as a 1st process example of a pattern recognition process. It is a flowchart for demonstrating the flow of a process as the 2nd process example of the pattern recognition process in information processing apparatus. It is a conceptual diagram which shows the structural example of a search tree. 12 is a flowchart for explaining a flow of processing as a third processing example of pattern recognition processing in the information processing apparatus.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 11 ... Information processing apparatus, 21 ... Image interface, 22 ... Processor, 22a ... Pattern recognition part, 23 ... Working memory, 24 ... Program memory, 25 ... Data memory, 25a ... Dictionary database (address database), 26 ... Output interface, DESCRIPTION OF SYMBOLS 30 ... Recognition control part, 31 ... Candidate extraction part, 32 ... Node expansion | deployment part, 33 ... Evaluation value calculation part, 34 ... Node selection part, 35 ... Determination part.

Claims (8)

A pattern recognition method used in an information processing apparatus for performing pattern recognition processing in a plurality of stages,
Expand the next stage recognition candidates belonging to the recognition candidates,
Estimate the time required for the recognition process for the next and subsequent recognition candidates that depend on each recognition candidate,
For each of the expanded recognition candidates, an evaluation value is calculated based on the posterior probability on the condition of all recognition processing results for recognition candidates that have been recognized and the estimated time required for the recognition processing ,
Select recognition candidates based on the calculated evaluation values for each recognition candidate ,
Determining a pattern recognition result from the selected recognition candidates;
A pattern recognition method characterized by the above. The selection of the recognition candidates selects a predetermined number of recognition candidates in descending order of the evaluation value for each stage.
The pattern recognition method according to claim 1, wherein: The selection of the recognition candidate is to select the recognition candidate having the largest evaluation value from among the recognition candidates that have calculated the evaluation value,
The expansion of the recognition candidate expands the next stage recognition candidate belonging to the recognition candidate having the largest selected evaluation value.
The pattern recognition method according to claim 1, wherein: The posterior probability of the recognition candidate includes the probability that the recognition processing result for the recognition candidate is output on the condition of the recognition candidate, the probability that the recognition processing result for the recognition result is output, and the previous one of the recognition candidate. Calculated based on posterior probabilities for stage recognition candidates,
The pattern recognition method according to claim 1, wherein the pattern recognition method is a pattern recognition method. A character recognition method used in a character recognition device that performs processing for recognizing character information composed of information of a plurality of layers,
Belonging to the word candidate to expand the word candidate of the next stage layer,
Estimate the time required for the recognition process for the next and subsequent recognition candidates subordinate to each word candidate,
For each expanded word candidate, an evaluation value is calculated based on the posterior probability on the condition of all character recognition processing results for the word candidates already subjected to character recognition processing and the estimated time required for the recognition processing ,
Select word candidates based on the calculated evaluation value for each word candidate ,
Determining a recognition result of the entire character information from the selected word candidates;
A character recognition method characterized by the above. The posterior probability of the word candidate is a probability that a character recognition process result for the word candidate is output on the condition of the word candidate, a probability that a character recognition process result for the recognition result is output, and one of the word candidates. Calculated based on the posterior probability for the previous word candidate,
The character recognition method according to claim 5 , wherein: A program for causing a computer to perform processing for recognizing a pattern in multiple stages,
The ability to expand the next stage of recognition candidates belonging to recognition candidates,
A function for estimating the time required for recognition processing for recognition candidates in the next and subsequent stages depending on each recognition candidate;
A function for calculating an evaluation value based on the posterior probability and the estimated time required for the recognition processing for each of the expanded recognition candidates based on the result of all recognition processing for recognition candidates that have already been recognized;
A function for selecting recognition candidates based on the calculated evaluation value for each recognition candidate;
A function for determining the recognition result of the pattern from the selected recognition candidates;
Pattern recognition program for realizing A program for causing a computer to perform processing for recognizing character information including information of a plurality of layers,
The ability to expand the next stage of word candidates belonging to word candidates,
A function for estimating the time required for recognition processing for recognition candidates in the next and subsequent layers subordinate to each word candidate;
A function for calculating an evaluation value based on the posterior probability and the estimated time required for the recognition processing for each expanded word candidate based on the result of all character recognition processing for the word candidate that has undergone character recognition processing;
A function for selecting word candidates based on the calculated evaluation value for each word candidate;
A function that determines the recognition result of the entire character information from the selected word candidates,
Character recognition program for realizing

Images