JP2016212473A - Information processor and information processing program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an information processor for eliminating the need of information on a boundary between single characters as a teacher signal in machine learning in character recognition of a recognition object.SOLUTION: Reception means of the information processor receives semi-character pattern strings and correct answer text strings of a recognition object, creation means creates a first lattice including a sequence of character recognition results to the recognition object, and a second lattice including a sequence of the correct answer text strings to the recognition object by a combination of the semi-character pattern strings, and difference means respectively calculates expectation values of differential values of evaluation values of the sequences included in the first lattice and the second lattice created by the creation means, and calculates difference between the expectation value of the first lattice and the expectation value of the second lattice as a recognition parameter update value in machine learning.SELECTED DRAWING: Figure 6

Description

  The present invention relates to an information processing apparatus and an information processing program.

  Patent Document 1 has an object to provide a handwritten character recognition method and apparatus for recognizing an input character string continuously input by a user for convenience, and relates to single character recognition accuracy of a plurality of stroke combinations in an input character string. Various cutout patterns can be obtained by calculating the various features of the input character string, including features and spatial geometric features, and by using the probability model in which the probability model coefficients are estimated by the parameter estimation method through sample training. The step of calculating the cutout reliability of each stroke combination and the step of recognizing characters of various written patterns by the multi-template matching method and the best cutout path when performing the single character recognition of the plurality of stroke combinations Process and post-processing for optimizing recognition results, simple structure, low hardware requirements, high speed recognition A precision, is disclosed to have an advantage that can be implemented in embedded systems.

  Patent Document 2 has an object of realizing reading of a character string that is robust to character contact and continuous writing, and character extraction / feature extraction means detects a cut-out position candidate from a character string image, and character string reading means Extracts a character pattern candidate from the character string image based on the cutout position candidate, verifies the validity of every possible reading result using the character appearance probability calculating means, and the character appearance probability calculating means is more than the character string reading means, Character pattern candidate, character code, character state, and character code / character state of the character pattern candidate located immediately before the character pattern candidate are received, and the validity of the geometric connection with the immediately preceding character pattern candidate is determined as the character state transition probability And using a character template to evaluate the validity of a character pattern candidate belonging to a category, It is disclosed that the likelihood (score) belonging to a certain character category is calculated, and the character string reading means searches for and outputs a character string segmentation recognition result that maximizes the recognition score of the entire character string.

In Patent Document 3, it is an object to optimize an OCR character recognition dictionary and character segmentation parameters. The character string collation unit obtains a character recognition result output by the character recognition unit and a correct text of the input document image. Collated by dynamic programming, extracted cumulative distance value between two character strings, correspondence between characters and misrecognized locations, especially in the dynamic programming used here, from the grid points where the character code correctly corresponds In addition to the normal distance value, a penalty distance value is added to the path extending in the horizontal and vertical directions to accurately determine the correspondence between characters, and the maintenance unit outputs the character string matching unit. It is disclosed that a character recognition dictionary of a misrecognized character is corrected or a character cutout parameter is finely adjusted so that an accumulated distance value becomes as small as possible.
Non-Patent Documents 1 to 5 disclose techniques related to machine learning and character recognition techniques.

Special table 2012-520492 gazette JP 2000-207495 A JP 09-251518 A

John D. Lufferty, Andrew McCallum, and Fernando C.L. N. Pereira. “Conditional random fields: Probabilistic models for segmenting and labeling sequence data,” In Proceedings of the Feeling International IC. 282-289, San Francisco, CA, USA, 2001. Morgan Kaufmann Publishers Inc. Peng, Jian and Bo, Liefeng and Xu, Jinbo, “Conditional Neural Fields,” Advances in neural information processing systems, pp. 1419-1427, 2009. Zhou, Xiang-Dong and Liu, Cheng-Lin and Nakagawa, Masaki, "Online handwritten Japanese character string recognition using conditional random fields," IEEE Computer Society, In Proceedings of the 2009 10th International Conference on Document Analysis and Recognition, ICDAR'09 , Pp. 521-525, Washington, DC, USA, 2009. Zhou, Xiang-Dong and Wang, Da-Han and Tian, Feng and Liu, Cheng-Lin and Nakagawa, Masaki, "Handwritten Chinese / Japanese text recognition using semi-Markov conditional random fields," IEEE Trans, Pattern Analysis and Machine Intelligence , Vol. 35, no. 10, pp. 2413-2426, 2013. Zhou, Xiang-Dong and Zhang, Yan-Ming and Tian, Feng and Wang, Hong-An and Liu, Cheng-Lin, "Minimum-risk training for semi-Markov conditional random fields with application to handwritten Chinese / Japanese text recognition, "Pattern Recognition, Vol. 47, NO. 5, pp. 1904-1916, 2014, Elsevier.

  An object of the present invention is to provide an information processing apparatus and an information processing program that do not require information regarding a boundary between single characters as a teacher signal in machine learning in character recognition of a recognition target.

The gist of the present invention for achieving the object lies in the inventions of the following items.
The invention of claim 1 is a combination of a receiving means for receiving a quasi-character pattern string and a correct text string for a recognition object, a first lattice including a series of character recognition results for the recognition object, and a quasi-character pattern string. Creation means for creating a second lattice including a series of correct text strings for the recognition target, and expectation of a differential value of evaluation values of series included in the first lattice and the second lattice created by the creation means An information processing apparatus comprising: a difference unit that calculates each value and calculates a difference between the expected value of the first lattice and the expected value of the second lattice as a recognition parameter update amount in machine learning It is.

  The invention of claim 2 is characterized in that the creating means creates a third lattice obtained by integrating the first lattice and the second lattice, and uses the third lattice as the first lattice. The information processing apparatus according to claim 1.

  The invention of claim 3 further comprises storage means for storing a single character identification result of the single character pattern, wherein the creating means uses the single character identification result of the single character pattern stored in the storage means. The information processing apparatus according to claim 1 or 2, characterized in that

  The invention of claim 4 further comprises second storage means for storing a calculation result relating to a local energy function in the expected value calculation of the first lattice, wherein the difference means is the second storage means. The information processing apparatus according to claim 1, wherein the calculation result stored in is used for calculating the expected value of the second lattice.

  The invention according to claim 5 is included in the subset extracted by the creating means for creating the set of the first and second lattice pairs, the extracting means for extracting the subset from the set, and the extracting means. 5. The calculation means including the information processing apparatus according to claim 1 for calculating a recognition parameter update amount for each of the pairs, and the recognition parameter update corresponding to the set of recognition parameter update amounts. An update means for updating the recognition parameter by calculating the sum of the quantities, multiplying by a predetermined coefficient, and subtracting the value of the multiplication result from the current recognition parameter update quantity, the extraction means, The information processing apparatus is characterized in that the processing by the calculating means and the updating means is repeated.

  The invention according to claim 6 is the information processing apparatus according to any one of claims 1 to 5, wherein the recognition target is a character image or character stroke information.

  The invention of claim 7 is a computer that accepts a quasi-character pattern sequence and a correct text sequence for a recognition target, a first lattice including a sequence of character recognition results for the recognition target, and a quasi-character pattern sequence. A creation unit that creates a second lattice including a series of correct text strings for the recognition target by combination, and a derivative of the evaluation value of the series included in the first lattice and the second lattice created by the creation unit Information processing program for calculating expected values of values and functioning as difference means for calculating a difference between the expected value of the first lattice and the expected value of the second lattice as a recognition parameter update amount in machine learning It is.

  According to the information processing apparatus of the first aspect, information regarding the boundary between single characters can be made unnecessary as a teacher signal in machine learning in character recognition to be recognized.

  According to the information processing apparatus of the second aspect, machine learning can be performed using the first lattice including the second lattice.

  According to the information processing apparatus of the third aspect, it is not necessary to perform character recognition for character recognition of overlapping single character patterns for the second and subsequent times.

  According to the information processing apparatus of claim 4, it is not necessary to perform the expected value calculation for the second and subsequent expected value calculations.

  According to the information processing apparatus of the fifth aspect, better recognition parameters can be obtained as compared with the case where the recognition parameters are not updated.

  According to the information processing apparatus of the sixth aspect, a character image or character stroke information can be set as a recognition target.

  According to the information processing program of the seventh aspect, in the machine learning in the recognition of the character to be recognized, the information regarding the boundary between the single characters can be made unnecessary as the teacher signal.

It is explanatory drawing which shows the example of the input character string pattern. It is explanatory drawing which shows the example of a quasi-character pattern row | line | column. It is explanatory drawing which shows the example of a single character pattern row | line candidate. It is explanatory drawing which shows the example of the output candidate lattice of a character string recognition. It is explanatory drawing which shows the example of a text string limitation lattice. It is a conceptual module block diagram about the structural example of 1st Embodiment (lattice preparation for learning). It is a conceptual module block diagram about the structural example of 1st Embodiment (recognition parameter update amount calculation). It is a conceptual module block diagram about the structural example of 2nd Embodiment (Japanese lattice creation). It is explanatory drawing which shows the example of the sum of the output candidate lattice of a text string limitation, and an output candidate lattice. It is a notional module block diagram about the structural example of 3rd Embodiment (a single character identification result is diverted). It is a notional module block diagram about the structural example of 4th Embodiment (border evaluation value is diverted). It is a notional module block diagram about the structural example of 5th Embodiment (SGD). It is a block diagram which shows the hardware structural example of the computer which implement | achieves this Embodiment.

The present embodiment relates to a character string recognition technique. Hereinafter, terms used in the description of this embodiment will be defined.
Character string recognition is a process of outputting a corresponding text string in response to an input of a character string pattern that is a character image or a character stroke. In particular, the present invention relates to a machine learning technique for character string recognition of a character string in which the number of characters and the boundary between single characters are unknown. In other words, because single characters are composed of partial and 旁, single characters are written in Japanese with multiple connected components, and single characters are written in contact with each other, so one connected component is written with multiple single characters. A character string such as

  In the description of the present embodiment, images and strokes are collectively referred to as “patterns”. That is, the input for character string recognition is a character string image or a character string stroke, which are collectively referred to as a character string pattern. Also, the type (reading) of a single character corresponding to the correct answer for character string recognition is called “text”. A boundary between single character patterns is called a “boundary”. The output of character string recognition is a single character pattern string divided by a boundary string and a text string that is a reading thereof, and these are collectively called an “output series”. Estimating the recognition parameters from the learning data is called “learning” (machine learning).

First, before describing the present embodiment, the premise or an information processing apparatus using the present embodiment will be described. This description is intended to facilitate understanding of the present embodiment.
A character string recognition process to which the present embodiment (that is, an information processing apparatus related to machine learning used for the recognition process) is applied will be described.
FIG. 1 shows an example of an input character string pattern. For the input shown in the example of FIG. 1, outputting a text string “Takuya Motomura” is a good character string recognition result.

  In the character string shown in the example of FIG. 1, since the boundary between characters is unknown, the region for single character identification is also unknown. Therefore, in the character string recognition here, first, a number of boundary candidates are detected, and a partial pattern that is a unit of a single character pattern or less is created. This partial pattern is called a quasi-character pattern. An example of the quasi-character pattern sequence is shown in FIG. In FIG. 2, the luminance is inverted for easy understanding.

このとき、単文字パタン候補を間引きしてもよい。例えば、Ｍ個以下の準文字パタンからなる単文字パタンを作成するという方法がある。Ｍ＝３としてこの方法を適用した単文字パタン列候補の例を図３に示す。つまり、単文字パタンを、３個以下の連続する準文字パタン（１個の準文字パタンを含む）によって構成する。例えば、左端にあるｘ_１を含む単文字パタンには、ｘ_１の単文字パタン、ｘ_１とｘ_２の単文字パタン、ｘ_１とｘ_２とｘ_３の単文字パタンがある。図３に示す例において、丸角長方形が単文字パタンを表す。その丸角長方形間をつなぐ直線（リンク）は、単文字パタンどうしが隣り合うことを示す。隣り合うノードを辿ることで単文字パタン列が得られる。また、ｂｏｓ，ｅｏｓは単文字パタン列の始点と終点を表す模式的なノードである。なお、良好な文字列認識結果が得られる単文字パタン列は、１文字目を（ｘ_１）、２文字目を（ｘ_２，ｘ_３）、３文字目を（ｘ_４，ｘ_５）、４文字目を（ｘ_６，ｘ_７）とするものであり、図３の例では、黒い丸角長方形（ノード３１０、ノード３２０、ノード３４０、ノード３６０）で図示している。 Subsequently, candidates for all single character pattern sequences are created for the quasi-character pattern sequence. A single character pattern is composed of one or more consecutive quasi-character patterns. Therefore, when the number of quasi-character patterns (the number of texts is unknown) is N, the total number of single-character pattern sequence candidates is the number shown in Equation 1.

At this time, single character pattern candidates may be thinned out. For example, there is a method of creating a single character pattern composed of M or less quasi-character patterns. An example of a single character pattern sequence candidate to which this method is applied with M = 3 is shown in FIG. That is, the single character pattern is composed of three or less consecutive quasi-character patterns (including one quasi-character pattern). For example, the single character pattern including x ₁ at the left end includes a single character pattern of x _1, a single character pattern of x ₁ and x _{2, and} a single character pattern of x ₁ , x _2, and x ₃ . In the example shown in FIG. 3, a rounded rectangle represents a single character pattern. A straight line (link) connecting the rounded rectangles indicates that single character patterns are adjacent to each other. A single character pattern string can be obtained by tracing adjacent nodes. Bos and eos are schematic nodes representing the start point and end point of the single character pattern string. In addition, the single character pattern sequence from which a good character string recognition result is obtained is that the first character is (x ₁ ), the second character is (x ₂ , x ₃ ), the third character is (x ₄ , x ₅ ), The fourth character is (x ₆ , x ₇ ), and in the example of FIG. 3, it is illustrated by a black rounded rectangle (node 310, node 320, node 340, node 360).

Subsequently, text candidates in a single character region (in the example of FIG. 3, a single character pattern indicated by a rounded rectangle) are created using a single character classifier. As a result, all single character pattern strings and text string candidates that can be output by the character string recognition are created.
At this time, text candidates may be thinned out. For example, there is a method of adopting the top K texts with high certainty of single character identification. FIG. 4 shows examples of all single character pattern strings and text string candidates that can be output by character string recognition using this method with K = 3. The example shown in FIG. 4 is an example in which text candidates are added to those shown in the example of FIG. 3, and the text is illustrated in a single character pattern. By tracing adjacent nodes, a single character pattern string and a text string are obtained. A good character string recognition result is that the first character is ((x ₁ , 'book'), the second character is ((x ₂ , x ₃ ), 'village'), the third character is ((x ₄ , x ₅ ), “taku”), and the fourth character is ((x ₆ , x ₇ ), “ya”). In the example of FIG. 4, node 410, node 420, node 440, node 460 This is shown in the figure.

また、ｂ_１＝１、ｅ_Ｔ＝Ｎ、および、ｂ_ｔ−１＋１＝ｅ_ｔである。また、テキスト列をＹ＝（ｙ_１，…，ｙ_Ｔ）とする。以上において、Ｔは単文字パタン数とテキスト数を表し、Ｔ≦Ｎである。 Finally, an optimum output is selected from these character string recognition output candidates and output.
Now, let the input quasi-character pattern sequence be X = (x ₁ ,..., X _N ). N is the quasi-character pattern string length. Also, the output single character pattern string is S = (s ₁ ,..., S _T ). _{Incidentally, s t = (b t,} e t) is, _{b t} is the start number of the quasi-characters included in the single character pattern _{s t, e t} represents the end number. That is, a single character pattern s _t consists Equation 2 is a quasi-character pattern.

_{Further, b 1 = 1, e T} = N, and _{a b t-1 + 1 = e} t. The text string is Y = (y ₁ ,..., Y _T ). In the above, T represents the number of single character patterns and the number of texts, and T ≦ N.

Furthermore, assuming that v _t = (s _t , y _t ) and the output sequence of character string recognition is V = (v ₁ ,..., V _T ), the example shown in FIG. Can see. The selection of the optimum candidate is to search for V ^* that minimizes the energy function E (X, V, Θ) as shown in Equation 3. Note that Θ is a recognition parameter obtained by learning.

The energy function E (X, V, Θ) may be calculated as in Equation 4. This is the sum of local energy functions E (v _t−1 , v _t, X, Θ) calculated from elements of adjacent output series. It is known that the search of Equation 3 can be solved at high speed by a method called Viterbi-Algorithm by decomposing the energy function into parts in this way.

なお、数式５はパラメータΘの要素である。また、Ｆ，Ｖはパラメータのサイズである。また、ｆ_ｉ（ｓ_ｔ−１，ｓ_ｔ，Ｘ）やｆ_ｉ（ｓ_ｔ−１，ｓ_ｔ，ｙ_ｔ−１、ｙ_ｔ，Ｘ）は特徴量関数である。また、σ（・）はシグモイド関数等を用いればよい。

なお、特徴量関数は、非特許文献３、４、５や特許文献１等で示される方法を用いればよい。また、数式３では、ｖ_０＝（ｓ_０，ｙ_０）が必要となるが、これはＸや事前の知識に基づいて計算される仮想的なノード（すなわち、ｂｏｓ）を設定すればよい。例えば、ｓ_０＝（０，０）としてｙ_０にはスペースコードや、文字列の始端を表す任意のシンボル等を設定すればよい。 Furthermore, the local energy function E (v _t−1 , v _t , X, Θ) may be calculated as in Equation 6, Equation 7, or Equation 8. Equation 6 is a method shown in Non-Patent Document 1, Equation 7 is Non-Patent Document 2, and Equation 8 is a method shown in Non-Patent Documents 3, 4, and 5.

Equation 5 is an element of the parameter Θ. F and V are parameter sizes. F _i (s _t−1 , _{st t} , X) and f _i (s _t−1 , _{st t} , y _t−1 , y _t , X) are feature quantity functions. Moreover, a sigmoid function or the like may be used for σ (•).

For the feature amount function, a method shown in Non-Patent Documents 3, 4, 5 or Patent Document 1 may be used. Further, in Expression 3, v ₀ = (s ₀ , y ₀ ) is required, but this may be performed by setting a virtual node (that is, bos) calculated based on X and prior knowledge. For example, s ₀ = (0, 0) and y ₀ may be set with a space code or an arbitrary symbol representing the start of a character string.

The character string recognition techniques shown in Patent Documents 1, 2, and 3 and Non-Patent Documents 3, 4, and 5 are all designed to recognize character strings whose number of characters and boundaries are unknown, as in this embodiment. is there. In addition, a method for learning the recognition parameter is shown.
The techniques disclosed in Non-Patent Documents 3, 4, 5 and Patent Document 1 require a correct output sequence as a learning teacher signal. That is, since both a correct text string and a correct single-character pattern string (boundary string) are required, the learning data creation cost is high.
On the other hand, since the techniques shown in Patent Document 2 and Patent Document 3 require only a correct text string as a learning teacher signal, there is no cost for providing a single character pattern string and learning data creation cost. Is small. However, in the technique shown in Patent Document 2, in order to compensate for information of a single character pattern sequence that is not given as a teacher signal, boundary division is performed as learning preprocessing, and a single character pattern sequence is uniquely determined. Since the character pattern string serves as a learning teacher signal, a good recognition parameter cannot be obtained unless a somewhat good recognition parameter is used as an initial value for learning. The technique disclosed in Patent Document 3 uses an edit distance between a correct text string and a text string output by character string recognition as an objective function of learning, and the edit distance is a discrete amount based on text mismatch and generally Since the accuracy of the evaluation function is rough, a good recognition parameter cannot be obtained by learning. Furthermore, since the edit distance that is the objective function cannot be differentiated by the recognition parameter, it is necessary to perform learning by numerical differentiation with a high calculation cost, which is expensive.
As described above, in character string recognition, learning data creation costs increase in order to obtain good recognition parameters by conventional learning.

An outline of the present embodiment will be described.
In the present embodiment, the character string recognizer creates an output sequence candidate limited by the correct text string for the character string pattern, thereby reducing learning data creation cost. Specifically, an output candidate lattice limited to a text string as shown in the example of FIG. 5 is created for the quasi-character pattern string as shown in the example of FIG. Hereinafter, this is referred to as a text string limited lattice. In addition to the correct output series included in the output candidate lattice shown in the example of FIG. 4, the text string limited lattice has a correct text string but an output series with a different single character pattern string. That is, in addition to “node 510, node 520, node 540, and node 560” which are correct output sequences, an output sequence that is not correct (an output sequence including an incorrect boundary) is given as a teacher signal. .
However, in the learning of the present embodiment, by weighting each output sequence of the text string limited lattice with the evaluation value, the influence of the output sequence that is not correct is reduced as the learning progresses, and learning is performed. It will be good.

According to the present embodiment, similar to the techniques disclosed in Patent Document 2 and Patent Document 3, only a correct text string is required as a teacher signal, and learning data generation without a teacher signal provision cost for a single character pattern string is required. Learning of character string recognition with low cost is realized.
Furthermore, since this embodiment does not uniquely determine a single character pattern string as pre-processing, it does not depend on the initial value, and a good recognition parameter can be obtained by learning. Similar to Patent Documents 3, 4, 5 and Patent Document 1, it can be differentiated by the recognition parameter and has sufficient accuracy for the evaluation function, so that a good recognition parameter can be obtained by learning, and its calculation cost Is small.

Hereinafter, examples of various preferred embodiments for realizing the present invention will be described with reference to the drawings.
<< First Embodiment >>
FIG. 6 shows a conceptual module configuration diagram of a configuration example of the first embodiment (learning lattice creation).
The module generally refers to components such as software (computer program) and hardware that can be logically separated. Therefore, the module in the present embodiment indicates not only a module in a computer program but also a module in a hardware configuration. Therefore, the present embodiment is a computer program for causing these modules to function (a program for causing a computer to execute each procedure, a program for causing a computer to function as each means, and a function for each computer. This also serves as an explanation of the system and method. However, for the sake of explanation, the words “store”, “store”, and equivalents thereof are used. However, when the embodiment is a computer program, these words are stored in a storage device or stored in memory. This means that control is performed so as to be stored in the apparatus. Modules may correspond to functions one-to-one, but in mounting, one module may be configured by one program, or a plurality of modules may be configured by one program, and conversely, one module May be composed of a plurality of programs. The plurality of modules may be executed by one computer, or one module may be executed by a plurality of computers in a distributed or parallel environment. Note that one module may include other modules. Hereinafter, “connection” is used not only for physical connection but also for logical connection (data exchange, instruction, reference relationship between data, etc.). “Predetermined” means that the process is determined before the target process, and not only before the process according to this embodiment starts but also after the process according to this embodiment starts. In addition, if it is before the target processing, it is used in accordance with the situation / state at that time or with the intention to be decided according to the situation / state up to that point. When there are a plurality of “predetermined values”, they may be different values, or two or more values (of course, including all values) may be the same. In addition, the description having the meaning of “do B when it is A” is used in the meaning of “determine whether or not it is A and do B when it is judged as A”. However, the case where it is not necessary to determine whether or not A is excluded.
In addition, the system or device is configured by connecting a plurality of computers, hardware, devices, and the like by communication means such as a network (including one-to-one correspondence communication connection), etc., and one computer, hardware, device. The case where it implement | achieves by etc. is also included. “Apparatus” and “system” are used as synonymous terms. Of course, the “system” does not include a social “mechanism” (social system) that is an artificial arrangement.
In addition, when performing a plurality of processes in each module or in each module, the target information is read from the storage device for each process, and the processing result is written to the storage device after performing the processing. is there. Therefore, description of reading from the storage device before processing and writing to the storage device after processing may be omitted. Here, the storage device may include a hard disk, a RAM (Random Access Memory), an external storage medium, a storage device via a communication line, a register in a CPU (Central Processing Unit), and the like.

An example of an embodiment of the minimum configuration of the present embodiment is shown in FIGS. According to the present embodiment, a recognition parameter update amount in learning is obtained.
The first embodiment (learning lattice creation) shown in the example of FIG. 6 includes a single character pattern sequence candidate creation module 610, an output candidate lattice creation 1 module 620, a text sequence limited lattice creation 1 module 630, and a single character identification 1 A module 640 and a learning lattice writing module 650 are provided.
The configuration shown in the example of FIG. 6 outputs the output candidate lattice 622 and the text string limited lattice 632 with the quasi-character pattern string 608 and the correct text string 606 as inputs (that is, information regarding boundaries is not necessary as an input). Hereinafter, the pair of these two lattices is referred to as a learning lattice 652.
The single character pattern sequence candidate creation module 610 is connected to the output candidate lattice creation 1 module 620 and the text sequence limited lattice creation 1 module 630, accepts the quasi-character pattern sequence 608 for the recognition target, and outputs the output candidate lattice creation 1 module. 620, the single character pattern string candidate 612 is passed to the text string limitation lattice creation 1 module 630.
The output candidate lattice creation 1 module 620 is connected to the single character pattern sequence candidate creation module 610, the single character identification 1 module 640, and the learning lattice writing module 650, and from the single character pattern sequence candidate creation module 610 to the single character pattern sequence The candidate 612 is received, the single character pattern 624 is passed to the single character identification 1 module 640, the single character identification result 642 is received from the single character identification 1 module 640, and the output candidate lattice 622 is passed to the learning lattice writing module 650. The output candidate lattice creation 1 module 620 creates an output candidate lattice 622 including a series of character recognition results for the recognition target.
The text string limited lattice creation 1 module 630 is connected to the single character pattern sequence candidate creation module 610, the single character identification 1 module 640, and the learning lattice export module 650, accepts the correct text sequence 606, and receives the single character pattern sequence candidate. The single character pattern sequence candidate 612 is received from the creation module 610, the single character pattern 634 is passed to the single character identification 1 module 640, the single character identification result 644 is received from the single character identification 1 module 640, and the text is output to the learning lattice writing module 650. Pass column-limited lattice 632. The text string restriction lattice creation 1 module 630 creates a text string restriction lattice 632 including a series of correct text strings for a recognition target by combining quasi-character pattern strings. The processing of the output candidate lattice creation 1 module 620 and the text string limited lattice creation 1 module 630 may be performed by one module.
The single character identification 1 module 640 is connected to the output candidate lattice creation 1 module 620 and the text string limited lattice creation 1 module 630, receives the single character pattern 624 from the output candidate lattice creation 1 module 620, and outputs the output candidate lattice 1. The single character identification result 642 is passed to the module 620, the single character pattern 634 is received from the text string limitation lattice creation 1 module 630, and the single character identification result 644 is passed to the text string limitation lattice creation 1 module 630.
The learning lattice writing module 650 is connected to the output candidate lattice creation 1 module 620 and the text string restriction lattice creation 1 module 630, and the output candidate lattice 622 is output from the output candidate lattice creation 1 module 620 to the text string restriction lattice creation 1. The text string limited lattice 632 is received from the module 630 and a learning lattice 652 is output.

First, the single character pattern string candidate creation module 610 creates a single character pattern string candidate 612 as shown in the example of FIG.
Subsequently, the output candidate lattice creation 1 module 620 creates an output candidate lattice 622 as shown in the example of FIG. Further, the text string limitation lattice creation module 630 creates a text string limitation lattice 632 as shown in the example of FIG. At this time, single character identification results 642 and 644 are obtained by the single character identification 1 module 640 for the single character patterns 624 and 634 included in the single character pattern string candidate 612.
Finally, the learning lattice writing module 650 writes the output candidate lattice 622 and the text limited lattice 632 as a pair.

The first embodiment (recognition parameter update amount calculation) shown in the example of FIG. 7 includes a learning lattice reading module 710, an expected value calculation 1 module 720, an expected value calculation 1 module 730, and a difference calculation module 740. Yes.
The configuration shown in the example of FIG. 7 receives a learning lattice 652 (output of the learning lattice writing module 650), which is a learning sequence candidate, and a recognition parameter 706, and outputs a recognition parameter update amount 742.
The learning lattice reading module 710 is connected to the expected value calculation 1 module 720 and the expected value calculation 1 module 730, receives the learning lattice 652, receives the output candidate lattice 712 in the expected value calculation 1 module 720, and calculates the expected value. The text string limited lattice 714 is passed to one module 730.
The expected value calculation 1 module 720 is connected to the learning lattice reading module 710 and the difference calculation module 740, receives the recognition parameter 706, receives the output candidate lattice 712 from the learning lattice reading module 710, and sends it to the difference calculation module 740. The output candidate lattice expectation value 722 is passed.
The expected value calculation 1 module 730 is connected to the learning lattice reading module 710 and the difference calculation module 740, receives the recognition parameter 706, receives the text string limited lattice 714 from the learning lattice reading module 710, and receives the difference string calculation module 740. Is passed the text string limited lattice expectation value 732. The expected value calculation 1 module 730 includes the evaluation value of the series included in the output candidate lattice 622 created by the output candidate lattice creation 1 module 620 and the text column restriction lattice 632 created by the text column restriction lattice creation 1 module 630. Each expected value of the differential value is calculated, and the difference between the expected value of the first lattice and the expected value of the second lattice is calculated as the recognition parameter update amount in machine learning.
The difference calculation module 740 is connected to the expected value calculation 1 module 720 and the expected value calculation 1 module 730, the output candidate lattice expected value 722 is output from the expected value calculation 1 module 720, and the text string is limited from the expected value calculation 1 module 730. A lattice expectation value 732 is received.

First, the learning lattice reading module 710 reads the learning lattice 652 and outputs the output candidate lattice 712 and the text string limited lattice 714.
Subsequently, the expected value calculation 1 modules 720 and 730 calculate Func (G, Θ) of Equation 9 for each lattice. Equation 9 takes a lattice G and a recognition parameter Θ as inputs. Note that Θ is a vector having a length corresponding to the number of dimensions of the parameter. Further, edges is a set of pairs of all adjacent nodes included in the lattice G. In the example shown in FIGS. 4 and 5, this corresponds to a pair of two nodes connected by a solid line. For the sake of simplicity, the expressions related to the local energy function are omitted as in Expressions 10 and 11. Further, α (v), β (v), and Z are calculated as Equation 12, Equation 13, Equation 14, Equation 15, and Equation 16, respectively. In Equation 12, prevs (v) is a set of nodes immediately before the node v included in the lattice G. In Equation 13, posts (v) is a set of nodes immediately after the node v included in the lattice G.
Subsequently, the difference calculation module 740 calculates the parameter update amount _∇Θ L (X, Y, Θ) of Equation 17. The output candidate lattice 712 is G (X). Further, the text string limitation lattice 714 is set to G (X, Y).

In Expression 17, Func (G (X, Y), Θ) calculated from the text string limitation lattice G (X, Y) corresponds to the calculation of Expression 18. P (S | X, Y, Θ) in Expression 18 is an evaluation value of the single character pattern string S when the quasi-character pattern string X, the correct text string Y, and the recognition parameter Θ are given. Calculate as follows. Further, as in Non-Patent Document 4, the calculation may be performed using Equation 20. That is, Equation 18 calculates the sum of the parameter update amounts calculated from the respective series included in the text string limited lattice, weighted by the evaluation value. Since Equation 19 satisfies the definition of probability, Equation 18 can be called an expected value. As a result, the evaluation value other than the correct output sequence becomes smaller as the recognition parameter Θ is improved by learning, so that the influence of an erroneous teacher signal is reduced and good learning is realized. Note that Func (G (X, Y), Θ) calculated from the output candidate lattice G (X) is the same as the techniques shown in Non-Patent Documents 3 and 4.
Further, the recognition parameter update amount may be calculated using the text string limitation lattice as in the technique disclosed in Non-Patent Document 5.

In the implementation, the learning lattice writing module 650 shown in the example of FIG. 6 and the learning lattice reading module 710 shown in the example of FIG. 7 may be removed, and FIG. 6 and FIG. 7 may be connected.
However, in learning, a large number of quasi-character pattern sequences 608 are often used as learning data, and recognition parameter updating is often repeated. The recognition parameter update amount 742 is calculated from the learning lattice 652 as shown in the embodiment shown in the example of FIG. Therefore, first, a large number of learning lattices 652 are created using the embodiment shown in the example of FIG. 6 and held in the storage device, and then stored in the storage device according to the embodiment shown in the example of FIG. By reading from the apparatus and calculating the parameter update amount, learning lattice creation is only required for the first time, and the calculation cost is reduced.

<< Second Embodiment >>
FIG. 8 is a conceptual module configuration diagram of a configuration example of the second exemplary embodiment (Japanese lattice creation).
The second embodiment includes a single character pattern sequence candidate creation module 610, an output candidate lattice creation 1 module 620, a text sequence limited lattice creation 1 module 630, a single character identification 1 module 640, a sum lattice creation module 850, and a learning lattice. A writing module 650 is included. A sum lattice creation module 850 is added to the configuration shown in the example of FIG. 6 (another form of the embodiment shown in the example of FIG. 6). In addition, the same code | symbol is attached | subjected to the site | part of the same kind as the above-mentioned embodiment, and the overlapping description is abbreviate | omitted (hereinafter the same).
In the second embodiment, a lattice (sum lattice 852) that is the sum of the text string limited lattice 632 and the output candidate lattice 622 is created. According to the second embodiment, a better recognition parameter update amount 742 than the first embodiment shown in the examples of FIGS. 6 and 7 can be obtained. As described above, in the case where the top K texts with high single-character identification certainty are adopted as the output candidate lattice 622 of the first embodiment (up to the top three in the example of FIG. 4). In some cases, the correct text is not included therein. The second embodiment deals with such a case.

The output candidate lattice creation 1 module 620 is connected to the single character pattern sequence candidate creation module 610, the single character identification 1 module 640, and the sum lattice creation module 850, and from the single character pattern sequence candidate creation module 610, the single character pattern sequence candidate 612 is received, the single character pattern 624 is passed to the single character identification 1 module 640, the single character identification result 642 is received from the single character identification 1 module 640, and the output candidate lattice 622 is passed to the sum lattice creation module 850.
The text string limited lattice creation 1 module 630 is connected to the single character pattern string candidate creation module 610, the single character identification 1 module 640, the sum lattice creation module 850, and the learning lattice export module 650, and accepts the correct text string 606. The single character pattern sequence candidate generation module 610 receives the single character pattern sequence candidate 612, passes the single character pattern sequence 634 to the single character identification 1 module 640, receives the single character identification result 644 from the single character identification 1 module 640, and receives the sum lattice. The text string limited lattice 632 is passed to the creating module 850 and the learning lattice writing module 650.
The sum lattice creation module 850 is connected to the output candidate lattice creation 1 module 620, the text string limited lattice creation 1 module 630, and the learning lattice export module 650, and the output candidate lattice creation 1 module 620 receives the output candidate lattice 622. The text string limitation lattice 632 is received from the text string limitation lattice creation 1 module 630, and the sum lattice 852 is passed to the learning lattice writing module 650. The sum lattice creation module 850 creates a sum lattice 852 obtained by integrating the output candidate lattice 622 created by the output candidate lattice creation 1 module 620 and the text column restriction lattice 632 created by the text column restriction lattice creation 1 module 630. The lattice 852 is treated as the output candidate lattice 622 in the first embodiment. The integration here is a so-called sum (logical sum processing). Specifically, when the node included in the text string limitation lattice 632 is not in the output candidate lattice 622, the node is added to the output candidate lattice 622. If the node included in the text string limitation lattice 632 is in the output candidate lattice 622, nothing is done (the node is not added to the output candidate lattice 622). This process is performed for all nodes in the text string limitation lattice 632.
The learning lattice writing module 650 is connected to the sum lattice creation module 850 and the text string limited lattice creation 1 module 630, and the sum lattice creation 850 from the sum lattice creation module 850 and the text string from the text string restriction lattice creation 1 module 630. The limited lattice 632 is received and a learning lattice 652 is output.

  In the second embodiment, a learning lattice 652 that is a learning sequence candidate and a recognition parameter 706 are input, and a recognition parameter update amount 742 is output. Unlike the example of FIG. 6, the sum lattice creation module 850 creates a lattice (sum lattice 852) that is the sum of the output candidate lattice 622 and the text string limited lattice 632. The sum lattice 852 includes a text column limited lattice 632. That is, all output sequences included in the text string limited lattice 632 are included in the sum lattice 852. An example of the Japanese lattice 852 is shown in FIG. The example shown in FIG. 9 is the sum of the lattices shown in the examples of FIGS. Note that the nodes shown in the example of FIG. 5 are indicated by dotted-line rounded rectangles in the example of FIG.

The learning lattice 652 in the second embodiment corresponds to the embodiment shown in the example of FIG. 6 in which the output candidate lattice 622 is replaced with the sum lattice 852.
The text string limited lattice 632 includes an output sequence that is not correct. On the other hand, the sum lattice 852 in the second embodiment in which the text string limitation lattice 632 is included also includes an output sequence that is not the correct answer. As a result, the expected value is calculated for each lattice and the difference is calculated as shown in Equation 17, so that the influence of the output sequence that is not correct is offset and a better recognition parameter update amount 742 is obtained.

<< Third Embodiment >>
FIG. 10 is a conceptual module configuration diagram of a configuration example of the third embodiment (using a single character identification result).
The third embodiment includes a single character pattern sequence candidate creation module 610, an output candidate lattice creation 2 module 1020, a text sequence limited lattice creation 2 module 1030, a single character identification 2 module 1040, a single character identification result storage module 1060, learning A lattice writing module 650. A single character identification result storage module 1060 is added to the configuration shown in the example of FIG. 6, and a single character identification 2 module 1040 is substituted for the single character identification 1 module 640 and an output candidate lattice is created instead of the output candidate lattice 1 module 620 Instead of the two module 1020 and the text string limitation lattice creation 1 module 630, the text string limitation lattice creation 2 module 1030 is used.
According to the third embodiment, the calculation cost of single character identification of overlapping single character patterns is reduced. This embodiment is another embodiment of the configuration shown in the examples of FIGS. In addition, you may combine with the structure shown in the example of FIG. That is, single character pattern sequence candidate creation module 610, output candidate lattice creation 2 module 1020, text sequence limited lattice creation 2 module 1030, single character identification 2 module 1040, single character identification result storage module 1060, sum lattice creation module 850, learning A lattice writing module 650 may be provided.

The single character pattern sequence candidate creation module 610 is connected to the output candidate lattice creation 2 module 1020, the text sequence limited lattice creation 2 module 1030, and the single character identification 2 module 1040, receives the quasi-character pattern sequence 608, and outputs the output candidate lattice. The single character pattern string candidate 612 is passed to the creation 2 module 1020, the text string limited lattice creation 2 module 1030, and the single character identification 2 module 1040.
The output candidate lattice creation 2 module 1020 is connected to the single character pattern sequence candidate creation module 610, the single character identification result storage module 1060, and the learning lattice writing module 650. The column candidate 612 is received, and the output candidate lattice 622 is passed to the learning lattice writing module 650. The output candidate lattice creation 2 module 1020 performs the same processing as the output candidate lattice creation 1 module 620, but uses the single character identification result of the single character pattern stored in the single character identification result storage module 1060.
The text string limited lattice creation 2 module 1030 is connected to the single character pattern string candidate creation module 610, the single character identification result storage module 1060, and the learning lattice write module 650. The pattern string candidate 612 is received, and the text string limited lattice 632 is passed to the learning lattice writing module 650. The text string limitation lattice creation 2 module 1030 performs the same processing as the text string limitation lattice creation 1 module 630, but uses the single character identification result of the single character pattern stored in the single character identification result storage module 1060.
The single character identification 2 module 1040 is connected to the single character pattern string candidate creation module 610 and the single character identification result storage module 1060, and receives the single character pattern string candidate 612 from the single character pattern string candidate creation module 610. The single character identification 2 module 1040 performs the same processing as the single character identification 1 module 640, but stores the result in the single character identification result storage module 1060.
The single character identification result storage module 1060 is connected to the output candidate lattice creation 2 module 1020, the text string limited lattice creation 2 module 1030, and the single character identification 2 module 1040. The single character identification result storage module 1060 stores a single character identification result of a single character pattern.
The learning lattice writing module 650 is connected to the output candidate lattice creation 2 module 1020 and the text string restriction lattice creation 2 module 1030, and the output candidate lattice 622 is output from the output candidate lattice creation 2 module 1020 to the text string restriction lattice creation 2. A text string limited lattice 632 is received from the module 1030 and a learning lattice 652 is output.

  In the third embodiment, first, the single character identification 2 module 1040 performs single character identification for all the single character patterns in the single character pattern string, and holds the result. Subsequently, the output candidate lattice creation 2 module 1020 and the text string restriction lattice creation 2 module 1030 create a lattice (output candidate lattice 622, text string restriction lattice 632) with reference to the single character identification result. As shown in the examples of FIGS. 4, 5, and 8, the output candidate lattice 622 and the text string limited lattice 632 have a common single character identification result, so that single character identification is performed once as in the third embodiment. By holding and referring to the result, the calculation cost for single character identification is reduced.

<< Fourth Embodiment >>
FIG. 11 is a conceptual module configuration diagram of a configuration example of the fourth embodiment (diverting the boundary evaluation value).
The fourth embodiment includes a learning lattice reading module 710, an expected value calculation 2 module 1120, an expected value calculation 3 module 1130, a local energy function calculation result storage module 1150, and a difference calculation module 740. The local energy function calculation result storage module 1150 is added to the configuration shown in the example of FIG. 7, and an expected value calculation 2 module 1120 instead of the expected value calculation 1 module 720 and an expected value calculation 3 instead of the expected value calculation 1 module 730. The module 1130 is used.
According to the fourth embodiment, the calculation result of the energy function is reduced by diverting the calculation result of the overlapping local energy function. This embodiment is another form of the embodiment of FIG.

The learning lattice reading module 710 is connected to the expected value calculation 2 module 1120 and the expected value calculation 3 module 1130, receives the learning lattice 652, receives the output candidate lattice 712 in the expected value calculation 2 module 1120, and calculates the expected value. The text string limited lattice 714 is passed to the three modules 1130.
The expected value calculation 2 module 1120 is connected to the learning lattice reading module 710, the local energy function calculation result storage module 1150, and the difference calculation module 740, accepts the recognition parameter 706, and receives the output candidate lattice from the learning lattice reading module 710. 712 is received and the output candidate lattice expected value 722 is passed to the difference calculation module 740. The expected value calculation 2 module 1120 performs processing equivalent to the expected value calculation 1 module 720, but stores the calculation result in the local energy function calculation result storage module 1150.
The expected value calculation 3 module 1130 is connected to the learning lattice reading module 710, the local energy function calculation result storage module 1150, and the difference calculation module 740, receives the recognition parameter 706, and is limited to the text string from the learning lattice reading module 710. The lattice 714 is received and the text string limited lattice expected value 732 is passed to the difference calculation module 740. The expected value calculation 3 module 1130 performs the same processing as the expected value calculation 1 module 730, but the calculation result related to the local energy function common to the expected value calculation 2 module 1120 is stored in the local energy function calculation result storage module 1150. The stored calculation result is used.
The local energy function calculation result storage module 1150 is connected to the expected value calculation 2 module 1120 and the expected value calculation 3 module 1130. The local energy function calculation result storage module 1150 stores a calculation result related to a local energy function common to the expected value calculation 2 module 1120 and the expected value calculation 3 module 1130 in the expected value calculation by the expected value calculation 2 module 1120. .
The difference calculation module 740 is connected to the expected value calculation 2 module 1120 and the expected value calculation 3 module 1130, the output candidate lattice expected value 722 is expected from the expected value calculation 2 module 1120, and the text string is limited from the expected value calculation 3 module 1130. A lattice expectation value 732 is received.

In the fourth embodiment, the expected value calculation 2 module 1120 calculates Formula 9, and holds the value of _∇Θ E (v ′, v), Formula 10 and Formula 11. Subsequently, the expected value calculation 3 module 1130 calculates Formula 9 with reference to the value held in the local energy function calculation result storage module 1150. If the corresponding _ΘΘ E (v ′, v) and the values of Equations 10 and 11 are not held, a new calculation is performed. As shown in the examples of FIGS. 4, 5, and 8, the output candidate lattice 712 and the text string limited lattice 714 have a common boundary. Therefore, the local energy function calculated once as in the fourth embodiment is used. By holding and referring to the value, the calculation cost of Equation 9 is reduced.

<< Fifth Embodiment >>
FIG. 12 is a conceptual module configuration diagram of a configuration example of the fifth embodiment (Stochastic Gradient Descent (probabilistic gradient descent method, hereinafter referred to as SGD)).
The fifth embodiment includes a learning lattice creation module 1210, a learning lattice full set storage module 1220, a subset extraction module 1230, a recognition parameter update amount calculation module 1240, a recognition parameter storage module 1250, and a recognition parameter update module 1260. doing.
According to the fifth embodiment, better recognition parameters can be obtained by using a plurality of learning lattices for learning.
In the fifth embodiment, a set of learning lattices and an initial value of a recognition parameter are input, and the recognition parameter is sequentially updated and output. A set of learning lattices is created in advance from a plurality of character string patterns and their correct text strings according to the embodiment shown in the examples of FIG. 6, FIG. 8, FIG.

The learning lattice creation module 1210 is connected to the learning lattice full set storage module 1220 and receives the correct text string set 1206 and the quasi-character pattern string set 1208. The learning lattice creation module 1210 creates a set of pairs of the output candidate lattice 622 and the text sequence limited lattice 632 using the correct text sequence set 1206 and the quasi-character pattern sequence set 1208. Specifically, it is created according to the embodiment shown in the examples of FIG. 6, FIG. 8, FIG.
The learning lattice full set storage module 1220 is connected to the learning lattice creation module 1210 and the subset extraction module 1230. The learning lattice full set storage module 1220 stores the set created by the learning lattice creation module 1210.
The subset extraction module 1230 is connected to the learning lattice full set storage module 1220 and the recognition parameter update amount calculation module 1240, and passes the learning lattice subset 1232 to the recognition parameter update amount calculation module 1240. The subset extraction module 1230 extracts the learning lattice subset 1232 from the set stored in the learning lattice full set storage module 1220.
The recognition parameter update amount calculation module 1240 is connected to the subset extraction module 1230, the recognition parameter storage module 1250, and the recognition parameter update module 1260. The recognition parameter update amount calculation module 1240 receives the learning lattice subset 1232 from the subset extraction module 1230 and sends it to the recognition parameter update module 1260. The recognition parameter update amount set 1242 is passed. The recognition parameter update amount calculation module 1240 calculates a recognition parameter update amount set 1242 for each pair included in the learning lattice subset 1232 extracted by the subset extraction module 1230. Specifically, it is created according to the embodiment shown in the examples of FIGS.
The recognition parameter storage module 1250 is connected to the recognition parameter update amount calculation module 1240 and the recognition parameter update module 1260. The recognition parameter storage module 1250 stores a recognition parameter (corresponding to the recognition parameter 706 in the above-described embodiment), and is updated by the recognition parameter update module 1260.
The recognition parameter update module 1260 is connected to the recognition parameter update amount calculation module 1240 and the recognition parameter storage module 1250, and receives the recognition parameter update amount set 1242 from the recognition parameter update amount calculation module 1240. The recognition parameter update module 1260 calculates the sum of the corresponding recognition parameter update amounts for the recognition parameter update amount set 1242, multiplies by a predetermined coefficient, and calculates the value of the multiplication result from the current recognition parameter update amount. By subtracting, the recognition parameter in the recognition parameter storage module 1250 is updated.
Then, the processing by the subset extraction module 1230, the recognition parameter update amount calculation module 1240, and the recognition parameter update module 1260 is repeated.

In the fifth embodiment, first, the learning lattice creation module 1210 uses the quasi-character pattern sequence set 1208 and the correct text sequence set 1206 according to the embodiments shown in the examples of FIG. 6, FIG. 8, FIG. Create a set of learning lattices.
Subsequently, the subset extraction module 1230 extracts the learning lattice subset 1232 that is the subset from the learning lattice set.

Subsequently, the recognition parameter update amount calculation module 1240 calculates the recognition parameter update amount for the learning lattice included in the learning lattice subset 1232 according to the embodiment shown in the example of FIGS. The recognition parameter update amount set 1242 is output.
Subsequently, the recognition parameter update module 1260 updates the recognition parameter by multiplying the previous sum by a predetermined coefficient and subtracting this value from the current recognition parameter.
The above processing is repeated.
By using a plurality of learning lattices, a better recognition parameter update amount can be obtained, so that a better recognition parameter can be obtained.

  A hardware configuration example of the information processing apparatus according to the present embodiment will be described with reference to FIG. The configuration illustrated in FIG. 13 is configured by, for example, a personal computer (PC) or the like, and illustrates a hardware configuration example including a data reading unit 1317 such as a scanner and a data output unit 1318 such as a printer.

  A CPU (Central Processing Unit) 1301 includes various modules described in the above-described embodiments, that is, a single character pattern sequence candidate creation module 610, an output candidate lattice creation 1 module 620, a text string limited lattice creation 1 module 630, a single Character identification 1 module 640, learning lattice writing module 650, learning lattice reading module 710, expected value calculation 1 module 720, expected value calculation 1 module 730, difference calculation module 740, sum lattice creation module 850, output candidate lattice creation 2 Module 1020, text string limited lattice creation 2 module 1030, single character identification 2 module 1040, expected value calculation 2 module 1120, expected value calculation 3 module 1130, The control unit executes processing according to a computer program describing an execution sequence of each module such as a learning lattice creation module 1210, a subset extraction module 1230, a recognition parameter update amount calculation module 1240, and a recognition parameter update module 1260.

  A ROM (Read Only Memory) 1302 stores programs used by the CPU 1301, calculation parameters, and the like. A RAM (Random Access Memory) 1303 stores programs used in the execution of the CPU 1301, parameters that change as appropriate during the execution, and the like. These are connected to each other by a host bus 1304 including a CPU bus or the like.

  The host bus 1304 is connected via a bridge 1305 to an external bus 1306 such as a PCI (Peripheral Component Interconnect / Interface) bus.

  A keyboard 1308 and a pointing device 1309 such as a mouse are input devices operated by an operator. The display 1310 includes a liquid crystal display device or a CRT (Cathode Ray Tube), and displays various types of information as text or image information.

  An HDD (Hard Disk Drive) 1311 includes a hard disk (may be a flash memory or the like), drives the hard disk, and records or reproduces a program executed by the CPU 1301 and information. The hard disk realizes functions as a single character identification result storage module 1060, a local energy function calculation result storage module 1150, a learning lattice full set storage module 1220, a recognition parameter storage module 1250, and the like. Further, various other data, various computer programs, and the like are stored.

  The drive 1312 reads data or a program recorded on a removable recording medium 1313 such as a magnetic disk, an optical disk, a magneto-optical disk, or a semiconductor memory, and reads the data or program into an interface 1307 and an external bus 1306. , The bridge 1305, and the RAM 1303 connected via the host bus 1304. The removable recording medium 1313 can also be used as a data recording area similar to a hard disk.

  The connection port 1314 is a port for connecting the external connection device 1315 and has a connection unit such as USB and IEEE1394. The connection port 1314 is connected to the CPU 1301 and the like via the interface 1307, the external bus 1306, the bridge 1305, the host bus 1304, and the like. A communication unit 1316 is connected to a communication line and executes data communication processing with the outside. The data reading unit 1317 is, for example, a scanner, and executes document reading processing. The data output unit 1318 is, for example, a printer, and executes document data output processing.

  Note that the hardware configuration of the information processing apparatus illustrated in FIG. 13 illustrates one configuration example, and the present embodiment is not limited to the configuration illustrated in FIG. 13, and the modules described in the present embodiment are executed. Any configuration is possible. For example, some modules may be configured with dedicated hardware (for example, Application Specific Integrated Circuit (ASIC), etc.), and some modules are in an external system and connected via a communication line Alternatively, a plurality of systems shown in FIG. 13 may be connected to each other via communication lines so as to cooperate with each other. In particular, in addition to personal computers, portable information communication devices (including mobile phones, smartphones, mobile devices, wearable computers, etc.), information appliances, robots, copiers, fax machines, scanners, printers, multifunction devices (scanners, printers, An image processing apparatus having two or more functions such as a copying machine and a fax machine) may be incorporated.

  Note that the above-described various embodiments may be combined (for example, adding or replacing a module in one embodiment in another embodiment), and processing contents of each module The technique described in the background art may be employed.

The program described above may be provided by being stored in a recording medium, or the program may be provided by communication means. In that case, for example, the above-described program may be regarded as an invention of a “computer-readable recording medium recording the program”.
The “computer-readable recording medium on which a program is recorded” refers to a computer-readable recording medium on which a program is recorded, which is used for program installation, execution, program distribution, and the like.
The recording medium is, for example, a digital versatile disc (DVD), which is a standard established by the DVD Forum, such as “DVD-R, DVD-RW, DVD-RAM,” and DVD + RW. Standard “DVD + R, DVD + RW, etc.”, compact disc (CD), read-only memory (CD-ROM), CD recordable (CD-R), CD rewritable (CD-RW), Blu-ray disc ( Blu-ray (registered trademark) Disc), magneto-optical disk (MO), flexible disk (FD), magnetic tape, hard disk, read-only memory (ROM), electrically erasable and rewritable read-only memory (EEPROM (registered trademark)) )), Flash memory, Random access memory (RAM) ), SD (Secure Digital) memory card, and the like.
The program or a part of the program may be recorded on the recording medium for storage or distribution. Also, by communication, for example, a local area network (LAN), a metropolitan area network (MAN), a wide area network (WAN), a wired network used for the Internet, an intranet, an extranet, or a wireless communication It may be transmitted using a transmission medium such as a network or a combination of these, or may be carried on a carrier wave.
Furthermore, the program may be a part of another program, or may be recorded on a recording medium together with a separate program. Moreover, it may be divided and recorded on a plurality of recording media. Further, it may be recorded in any manner as long as it can be restored, such as compression or encryption.

606 ... Correct text string 608 ... Quasi-character pattern string 610 ... Single character pattern string candidate creation module 612 ... Single character pattern string candidate 620 ... Output candidate lattice creation 1 module 622 ... Output candidate lattice 624 ... Single character pattern 630 ... Text string limitation Lattice creation 1 module 632 ... Text string limited lattice 634 ... Single character pattern 640 ... Single character identification 1 module 642 ... Single character identification result 644 ... Single character identification result 650 ... Learning lattice export module 652 ... Learning lattice 706 ... Recognition parameters 710 ... Learning lattice reading module 712 ... Output candidate lattice 714 ... Text string limited lattice 720 ... Expected value calculation 1 module 722 ... Output candidate lattice expected value 730 ... Expected value calculation 1 module 732 ... Text column limit Expected constant lattice 740 ... Difference calculation module 742 ... Recognition parameter update amount 850 ... Sum lattice creation module 852 ... Sum lattice 1020 ... Output candidate lattice creation 2 module 1030 ... Text string limited lattice creation 2 module 1040 ... Single character identification 2 module 1060 Single character identification result storage module 1120 ... Expected value calculation 2 module 1130 ... Expected value calculation 3 module 1150 ... Local energy function calculation result storage module 1206 ... Correct text string set 1208 ... Quasi-character pattern string set 1210 ... Learning lattice creation module 1220 ... Learning lattice all set storage module 1230 ... Subset extraction module 1232 ... Learning lattice subset 1240 ... Recognition parameter update amount calculation module 12 42 ... Recognition parameter update amount set 1250 ... Recognition parameter storage module 1260 ... Recognition parameter update module

Claims (7)

Accepting means for accepting a quasi-character pattern sequence and a correct text sequence for the recognition target;
Creating means for creating a second lattice including a series of correct text strings for the recognition target by combining a first lattice including a series of character recognition results for the recognition target and a quasi-character pattern sequence;
Expected values of the differential values of the evaluation values of the series included in the first lattice and the second lattice created by the creating means are calculated, and the expected value of the first lattice is used as a recognition parameter update amount in machine learning. An information processing apparatus comprising: difference means for calculating a difference between a value and an expected value of the second lattice. The said creation means creates the 3rd lattice which integrated the said 1st lattice and the said 2nd lattice, This 3rd lattice is made into the 1st lattice. The Claim 1 characterized by the above-mentioned. Information processing device. Storage means for storing a single character identification result of the single character pattern;
The information processing apparatus according to claim 1, wherein the creating unit uses a single character identification result of a single character pattern stored in the storage unit. A second storage means for storing a calculation result related to a local energy function in the expected value calculation of the first lattice;
The information processing apparatus according to claim 1, wherein the difference unit uses a calculation result stored in the second storage unit for calculating an expected value of the second lattice. Creating means for creating a set of pairs of the first lattice and the second lattice;
Extraction means for extracting a subset from the set;
Calculation means including the information processing apparatus according to any one of claims 1 to 4, for calculating a recognition parameter update amount for each pair included in the subset extracted by the extraction means;
For the set of recognition parameter update amounts, the sum of the corresponding recognition parameter update amounts is calculated, multiplied by a predetermined coefficient, and the value of the multiplication result is subtracted from the current recognition parameter update amount. An updating means for updating the parameters,
An information processing apparatus characterized by repeating the processing by the extracting unit, the calculating unit, and the updating unit. The information processing apparatus according to any one of claims 1 to 5, wherein the recognition target is a character image or character stroke information. Computer
Accepting means for accepting a quasi-character pattern sequence and a correct text sequence for the recognition target;
Creating means for creating a second lattice including a series of correct text strings for the recognition target by combining a first lattice including a series of character recognition results for the recognition target and a quasi-character pattern sequence;
Expected values of the differential values of the evaluation values of the series included in the first lattice and the second lattice created by the creating means are calculated, and the expected value of the first lattice is used as a recognition parameter update amount in machine learning. An information processing program for functioning as a difference means for calculating a difference between a value and an expected value of the second lattice.

Images