JP5679936B2 - Word recognition device, word recognition method, and paper sheet processing device provided with word recognition device - Google Patents

Description

  Embodiments described herein relate generally to a word recognition device, a word recognition method, and a paper sheet processing apparatus including the word recognition device.

  2. Description of the Related Art Conventionally, a paper sheet processing apparatus such as a mail sorting machine that processes paper sheets such as mail has been put into practical use. Such a paper sheet processing apparatus takes in paper sheets that have been input into the input unit one by one, and acquires an image from the paper sheets. The paper sheet processing apparatus includes a word recognition device. The word recognition device recognizes a word on a paper sheet based on an image acquired from the paper sheet. The paper sheet processing apparatus identifies an address or other information written on the paper sheet based on the recognition result, and sorts the paper sheet into a predetermined sorting pocket.

  As a method of recognizing a word, an analytical method (Analytic Approach) and an overall method (Holistic Approach) are generally known. The analytical method and the overall method have a complementary relationship. For this reason, the word recognition apparatus can recognize a word with higher accuracy by using both the analytical method and the overall method.

  When recognizing a word by an analytical method, the word recognition device extracts a plurality of cutting point candidates based on the word image, and generates a plurality of character candidates including overlapping portions based on the extracted cutting point candidates. . Furthermore, the word recognition device calculates an evaluation value of the analytical method using the posterior probability ratio, and selects a correct combination from a plurality of character candidates based on the evaluation value.

  Further, when recognizing a word by the overall method, the word recognizing device recognizes the word by using, for example, a hidden Markov model (HMM).

Japanese Patent No. 4601835

  As a method of using both the analytical method and the overall method, recognition is first performed by the overall method, the word is divided into characters based on the recognition result, and each divided character is recognized by the analytical method. Thus, there is a method of verifying the word recognition result. However, when a misperception occurs in recognition by the overall method, the word recognition device fails to recognize characters, so that the word recognition device cannot obtain sufficient accuracy.

  Therefore, an object of the present invention is to provide a word recognition device, a word recognition method, and a paper sheet processing device including the word recognition device that can recognize words with higher accuracy.

  A word recognition device according to an embodiment includes a word dictionary that stores a plurality of words, an image receiving unit that receives an image including a word, a word image extracting unit that extracts a word image for each word from the image, and the word Character candidate extraction means for extracting character candidates from the image, character recognition means for performing character recognition on the character candidates, and words stored in the word dictionary based on the result of character recognition by the character recognition means Analytical matching means for calculating a first evaluation value every time, feature extraction means for extracting features from the word image, word model generation means for generating a word model for each word stored in the word dictionary, An overall matching means for calculating a second evaluation value indicating the probability of occurrence of the feature for each word model; and a feature probability calculation means for calculating a feature probability of occurrence of the feature; Integrated evaluation value calculating means for calculating a third evaluation value by multiplying the first evaluation value, the second evaluation value, and the inverse of the feature probability, and the first evaluation value calculated by the integrated evaluation value calculating means. Output means for outputting 3 evaluation values.

FIG. 1 is a diagram for explaining an example of a paper sheet processing apparatus according to an embodiment. FIG. 2 is a diagram for describing an example of a word recognition device according to an embodiment. FIG. 3 is a diagram for explaining processing of the word recognition device according to the embodiment. FIG. 4 is a diagram for explaining processing of the word recognition device according to the embodiment. FIG. 5 is a diagram for explaining processing of the word recognition device according to the embodiment. FIG. 6 is a diagram for explaining the processing of the word recognition device according to the embodiment. FIG. 7 is a diagram for explaining processing of the word recognition device according to the embodiment. FIG. 8 is a diagram for explaining processing of the word recognition device according to the embodiment. FIG. 9 is a diagram for explaining processing of the word recognition device according to the embodiment.

(First embodiment)
Hereinafter, a paper sheet processing apparatus and a light detection apparatus according to an embodiment will be described in detail with reference to the drawings.

FIG. 1 shows a configuration example of a paper sheet processing apparatus 100 according to an embodiment.
The paper sheet processing apparatus 100 reads an image from the paper sheet, recognizes destination information and a stamping position from the read image, and stamps the paper sheet to classify the paper sheet. The sheet processing apparatus 100 includes a supply unit 200, a separation roller 210, a conveyance path 220, an image reading unit 400, a stamping unit 460, a printing unit 470, a main control unit 500, a sorting processing unit 300, a word recognition unit 600, and an operation unit. 700, a display unit 800, and an input / output unit 900.

  The main control unit 500 controls the operation of each unit of the paper sheet processing apparatus 100 in an integrated manner. The main control unit 500 includes a CPU, a buffer memory, a program memory, a nonvolatile memory, and the like. The CPU performs various arithmetic processes. The buffer memory temporarily stores the results of calculations performed by the CPU. The program memory and the nonvolatile memory store various programs executed by the CPU, control data, and the like. The main control unit 500 can perform various processes by executing a program stored in the program memory by the CPU.

  The supply unit 200 stocks the paper sheets 1 to be taken into the paper sheet processing apparatus 100. The supply unit 200 collectively receives the stacked paper sheets 1.

  The separation roller 210 is installed at the lower end of the supply unit 200, for example. When the paper sheet 1 is loaded into the supply unit 200, the separation roller 210 contacts the lower end of the loaded paper sheet 1 in the stacking direction. The separation roller 210 rotates to take the sheets 1 set in the supply unit 200 one by one from the lower end in the stacking direction into the sheet processing apparatus 100.

  For example, the separation roller 210 takes in one sheet 1 every rotation. Thereby, the separation roller 210 can take in the paper sheets 1 at a constant pitch. The paper sheet 1 taken in by the separation roller 210 is introduced into the conveyance path 220.

  The conveyance path 220 is a conveyance unit that conveys the paper sheet 1 to each unit in the paper sheet processing apparatus 100. The conveyance path 220 includes a conveyance belt (not shown) and a drive pulley (not shown). The conveyance path 220 drives a drive pulley by a drive motor (not shown). The conveyor belt is operated by a driving pulley.

  The transport path 220 transports the paper sheet 1 taken in by the separation roller 210 in the direction of arrow a (transport direction a) at a constant speed by the transport belt. Note that the side closer to the separation roller 210 in the conveyance path 220 will be described as an upstream side, and the opposite side will be described as a downstream side.

  The image reading unit 400 acquires an image from the paper sheet 1 conveyed by the conveyance path 220. The image reading unit 400 includes, for example, illumination and an optical sensor. The illumination irradiates light onto the paper sheet 1 conveyed by the conveyance path 220. The optical sensor includes a light receiving element such as a Charge Coupled Device (CCD) and an optical system (lens). The optical sensor receives the reflected light reflected by the paper sheet 1 by the optical system, forms an image on the CCD, and acquires an electrical signal (image). The image reading unit 400 acquires an entire image of the paper sheet 1 by continuously acquiring images from the paper sheet 1 conveyed by the conveyance path 220. The image reading unit 400 supplies the acquired image to the main control unit 500. Note that the image reading unit 400 may include a video camera or the like.

  Based on the image received from the image reading unit 400, the main control unit 500 performs processing for determining the transport destination of the paper sheet 1. For this purpose, the main control unit 500 identifies an address (destination information) as a destination by causing the word recognition unit 600 to recognize a word in the image. The main control unit 500 generates an image such as a two-dimensional code or a barcode based on the destination information, and supplies the generated image to the printing unit 470.

  Further, the main control unit 500 specifies the position where a stamp or the like on the paper sheet 1 is pasted. Furthermore, the main control unit 500 controls the operation of the stamp unit 460 based on the specified stamp position.

  The stamp unit 460 stamps a stamp such as a date stamp on the paper sheet 1 based on the control of the main control unit 500. That is, the main control unit 500 controls the stamping unit 460 so as to stamp the stamp at a position overlapping the stamping position. For example, the stamp unit 460 stamps the stamp so as to be a split.

  The printing unit 470 prints an image such as a two-dimensional code or a barcode based on the control of the main control unit 500. That is, the printing unit 470 prints an image such as a two-dimensional code or a barcode supplied from the main control unit 500. For example, the printing unit 470 prints the above image with an ink containing a phosphor that emits excitation light when irradiated with ultraviolet rays.

  The sorting processing unit 300 sorts and stacks the sheets 1 based on the control of the main control unit 500. The sorting processing unit 300 includes, for example, a plurality of gates and stackers such as a first gate 310, a first stacker 320, a second gate 330, and a second stacker 340. Further, the sorting processing unit 300 further includes a plurality of gates and a plurality of stackers. A stacker is provided for each destination information, for example. A gate is provided for each stacker.

  The main control unit 500 can sort the sheets 1 by controlling each gate of the sorting processing unit 300. Thereby, the sorting processing unit 300 can stack the paper sheets 1 in different stackers for each piece of destination information of the paper sheets 1.

  The first gate 310 and the second gate 330 are provided downstream of the image reading unit 400, the stamping unit 460, and the printing unit 470 in the conveyance path 220. The first gate 310 and the second gate 330 each operate based on the control of the main control unit 500. The main control unit 500 controls the first gate 310 and the second gate 330 according to the destination information recognized by the above processing.

  The first gate 310 switches the transport destination of the paper sheet 1 between the first stacker 320 and the second gate 330. The second gate 330 switches the transport destination of the paper sheet 1 between the second stacker 340 and another stacker.

  The main control unit 500 supplies the image received from the image reading unit 400 to the word recognition unit 600 so that the word recognition unit 600 can recognize words in the image.

  The word recognition unit 600 recognizes a word in the received image. The word recognition unit 600 outputs the recognition result to the main control unit 500. The main control unit 500 identifies destination information and the like based on the recognition result by the word recognition unit 600.

  In addition, the main control unit 500 includes a memory that holds an image of the paper sheet 1 for which destination information could not be specified. In addition, the sorting processing unit 300 includes a stacker that accumulates the paper sheets 1 whose destination information could not be specified.

  The operation unit 700 receives various operation inputs from the operator through the operation unit. The operation unit 700 generates an operation signal based on an operation input by the operator, and transmits the generated operation signal to the main control unit 500.

  For example, the paper sheet processing apparatus 100 may include a VCS (Video Coding System) function. That is, the main control unit 500 of the paper sheet processing apparatus 100 causes the display unit 800 to display an image of the paper sheet 1 whose destination information could not be specified. The paper sheet processing apparatus 100 causes the operator to read the image of the paper sheet 1 displayed on the display unit 800 and causes the operation unit 700 to input destination information. Thereby, the paper sheet processing apparatus 100 can acquire correct destination information.

  The display unit 800 displays various screens based on the control of the main control unit 500. For example, the display unit 800 displays various operation guidance and processing results for the operator. Further, as described above, the display unit 800 may be configured to display an image of the paper sheet 1 for which the destination information is not specified. Note that the operation unit 700 and the display unit 800 may be integrally formed as a touch panel.

  The input / output unit 900 transmits / receives data to / from an external device connected to the paper sheet processing apparatus 100 or a storage medium. For example, the input / output unit 900 includes a disk drive, a USB connector, a LAN connector, or another interface capable of transmitting and receiving data. The paper sheet processing apparatus 100 can acquire data from an external device connected to the input / output unit 900 or a storage medium. In addition, the paper sheet processing apparatus 100 can transmit the processing result to an external device connected to the input / output unit 900 or a storage medium.

FIG. 2 shows an example of the configuration of the word recognition unit 600 according to the first embodiment.
The word recognition unit 600 includes an image receiving unit 601, a word extraction unit 602, a character candidate extraction unit 603, a character recognition unit 604, a feature extraction unit 605, an analytical matching unit 610, an overall matching unit 620, a feature probability calculation unit 630, VCS 640, first word image storage unit 641, model learning unit 642, model storage unit 643, word model generation unit 644, word dictionary 645, prior probability calculation unit 651, prior probability storage unit 652, prior probability input unit 653, integration An evaluation value calculation unit 660 and a prior probability multiplication unit 670 are provided.

  The operation of the word recognition unit 600 is roughly divided into a recognition phase and a learning phase. First, the recognition phase will be described.

  The word recognizing unit 600 performs analytical matching, performs overall matching, calculates feature probabilities, and integrates these results and prior probabilities for each word. Thereby, the word recognition part 600 can calculate the evaluation value (posterior probability) for every word.

  In general, in pattern recognition, it is optimal to assign a pattern to a category having the maximum posterior probability. That is, when a pattern belongs to a category having the maximum posterior probability, the identification error is minimized. That is, the word recognizing unit 600 can identify one word and transmit it to the main control unit 500 by outputting the word for which the highest posterior probability is calculated as a recognition result. For example, the word recognition unit 600 may be configured to output an evaluation value for each word to the main control unit 500 as a recognition result. In this case, the main control unit 500 can specify the destination information in consideration of the posterior probabilities of a plurality of words and combinations with other words.

For example, when all character recognition results in word candidates in analytical matching are Y and features extracted from an image used for overall matching are X, the word recognition unit 600 uses the following formula 1 to generate the word w The posterior probability P (w | Y, X) is calculated.

  The left side of Equation 1 shows a posterior probability of a word in a case where a set of character recognition results by analytical matching and a set of feature extraction results used for overall matching are used as conditions. That is, the posterior probability P (w | Y, X) indicates an evaluation value for each word when analytical matching and overall matching are used together.

  The left side of Equation 1 can be expanded like the right side by Bayes' theorem. Furthermore, by assuming that the result of the analytical matching and the result of the overall matching are independent from each other, the right side of the first stage of Equation 1 can be approximated as shown on the right side of the second stage.

  Note that P (Y | w) / P (Y) in Equation 1 indicates the result of analytical matching (a posteriori probability ratio). In addition, P (X | w) in Equation 1 indicates the overall matching result (likelihood). Furthermore, P (X) in Formula 1 indicates the calculation result of the feature probability. Furthermore, P (w) indicates the prior probability for the word w.

  The word recognition unit 600 can calculate the a posteriori probability for each word by calculating each of the above terms and calculating Formula 1.

  First, analytical matching will be described. The image receiving unit 601 of the word recognition unit 600 receives the image of the paper sheet 1 (paper sheet image) from the main control unit 500. FIG. 3 shows an example of a paper sheet image. As shown in FIG. 3, the image receiving unit 601 receives a paper sheet image including words written on the paper sheet 1. FIG. 3 shows an example in which a destination and the like are described using English words. However, the present embodiment can be applied even if the destination described on the paper sheet 1 is in Japanese or another language. The image receiving unit 601 transmits the received paper sheet image to the word extracting unit 602.

  The word extraction unit 602 extracts word candidates (word images) from the paper sheet image received by the image receiving unit 601. For example, the word extraction unit 602 identifies and extracts an area that is likely to be segmented as a word by performing image processing on the paper sheet image. FIG. 4 shows an example of word candidates. As shown in FIG. 4, the word extraction unit 602 extracts word candidates from the paper sheet image.

  For example, the word extraction unit 602 extracts word candidates by recognizing spaces between words, for example. In addition, for example, the word extraction unit 602 may be configured to extract word candidates by extracting “city”, “town”, or other keywords that serve as delimiters. Further, the process of extracting the word candidates is not limited to the above method, and any process may be used. The word extraction unit 602 transmits the extracted word candidates to the character candidate extraction unit 603 and the feature extraction unit 605.

  Character candidate extraction unit 603 extracts character candidates from word candidates. The character candidate extraction unit 603 identifies and extracts a region that is likely to be segmented as a character by performing image processing on the word candidate (word image). FIG. 5 shows an example of processing for extracting character candidates from word candidates. As illustrated in FIG. 5, the character candidate extraction unit 603 extracts a plurality of cut point candidates from the character candidates, and extracts a plurality of character candidates including overlapping portions based on the extracted cut point candidates. That is, the character candidate extraction unit 603 identifies a region that is highly likely to be recognized as one character and extracts it as a character candidate. The character candidate extraction unit 603 transmits the extracted character candidates to the character recognition unit 604.

  The character recognition unit 604 performs character recognition for each character candidate and acquires a character recognition result. That is, the character recognition unit 604 obtains a character recognition result by comparing a character candidate image with a character recognition dictionary prepared in advance. The character recognition unit 604 transmits the character recognition result for each character candidate to the analytical matching unit 610.

  The word dictionary 645 stores words to be recognized as a list. FIG. 6 shows an example of the word dictionary 645. When recognizing a word, the word recognition unit 600 selects a correct word from the list in the word dictionary 645. The word dictionary 645 supplies the word list to the analytical matching unit 610.

  The analytical matching unit 610 calculates the posterior probability ratio for each word stored in the word dictionary 645 based on the character recognition result for each character candidate transmitted from the character recognition unit 604. Accordingly, the analytical matching unit 610 searches for a correct path (route) of the plurality of character candidates extracted by the character candidate extraction unit 603.

For example, the i-th character of the word w is c _i , the serial number of the character candidate corresponding to the i-th character is f (i), and the character recognition result of the character candidate corresponding to the i-th character is y _{f ( i) When} the number of characters of the word w is N, the posterior probability ratio P (Y | w) / P (Y) of the word w is approximated as shown in Equation 2 below.

For example, when the target word is “ham”, c ₁ = “h”, c ₂ = “a”, and c ₃ = “m”. In this case, N = 3. In this case, P (y _{f (i)} | c _i ) / P (y _{f (i)} ) represents the posterior probability ratio of the i-th character.

  The analytical matching unit 610 calculates the posterior probability ratio P (Y | w) / P (Y) of the word w by multiplying the posterior probability ratio of the i-th character over i = 1 to N. be able to. In other words, the analytical matching unit 610 calculates the posterior probability ratio for each word by calculating the above Equation 2 based on the character recognition result for each word in the word list.

The analytical matching unit 610 includes a character probability calculation unit 611, a first calculation unit 612, and a second calculation unit 613. The character probability calculation unit 611 calculates the numerator of each factor on the right side of Equation 2. That is, the character probability calculation unit 611 calculates P (y _{f (i)} | c _i ) for each character of a certain word w.

The first calculation unit 612 calculates each factor on the right side of Equation 2. That is, the first calculation unit 612 calculates P (y _{f (i)} ) that is the denominator of the right side, and divides P (y _{f (i)} | c _i ) that is the numerator by the calculated value. Note that P (y _{f (i)} ) is the probability that the character recognition result y _{f (i)} will appear.

  The second calculation unit 613 calculates the right side of Equation 2. That is, the second calculation unit 613 multiplies all the factors on the right side of Formula 2 that is the calculation result of the first calculation unit 612. Thereby, the analytical matching unit 610 can calculate the posterior probability ratio P (Y | w) / P (Y) of the word w. The analytical matching unit 610 outputs the calculated posterior probability ratio P (Y | w) / P (Y) to the integrated evaluation value calculation unit 660.

  Next, overall matching will be described. The feature extraction unit 605 in FIG. 2 receives the image of the word candidate from the word extraction unit 602 as described above. The feature extraction unit 605 extracts a feature X that is a set of vectors based on the received word candidate images.

  For example, the feature extraction unit 605 extracts the feature X using the brightness gradient information after blurring the word candidate image as a 128-dimensional vector. The feature extraction unit 605 extracts a plurality of features while shifting a region of interest (a region of interest) in the word candidate image from left to right in the image.

  The feature extraction unit 605 may be configured to use the density value of the pixel of the word candidate image as a feature. Furthermore, the feature extraction unit 605 may be configured to use a pattern density value obtained by further simplifying a word candidate image as a feature.

  Through the above processing, the feature extraction unit 605 extracts a plurality of features from one word candidate image. The feature extraction unit 605 outputs the extracted feature X to the overall matching unit 620 and the feature probability calculation unit 630.

  The model storage unit 643 stores a character model for each character or a word model for each word. The model storage unit 643 may be configured to store a word model corresponding to each word in the word dictionary 645. Further, the model storage unit 643 may be configured to store an arbitrary word model corresponding to an arbitrary word.

  The word model generation unit 644 generates a word model corresponding to each word in the word dictionary 645 using the character model and the word model stored in the model storage unit 643. The word model generation unit 644 outputs the generated word model to the overall matching unit 620.

  For example, the word model generation unit 644 reads a character model stored in the model storage unit 643 and generates a word model by connecting the character models according to the words in the word dictionary 645. When a word model corresponding to a word in the word dictionary 645 is stored in the model storage unit 643, the word model generation unit 644 directly uses the word model stored in the model storage unit 643 as an overall matching unit 620. Output to.

  The overall matching unit 620 calculates a likelihood P (X | w) based on the feature X extracted by the feature extraction unit 605 and the word model output from the word model generation unit 644. The likelihood P (X | w) is a probability that the feature X extracted by the feature extraction unit 605 is output from the word model output from the word model generation unit 644. The likelihood P (X | w) is the same as the numerator of the second factor on the right side of Equation 1.

  The overall matching unit 620 calculates a likelihood P (X | w) by using a Viterbi algorithm.

  The Viterbi algorithm is an algorithm that calculates a state sequence having the highest possibility (likelihood) of outputting a given array when model parameters are known. That is, the Viterbi algorithm is a dynamic programming algorithm that searches for a sequence of hidden events that result in feature X.

  The overall matching unit 620 calculates a likelihood P (X | w) as a probability that the feature X appears using the Viterbi algorithm with the word model output from the word model generation unit 644 as a known parameter. That is, the likelihood P (X | w) indicates the probability that the feature X appears from the word model corresponding to the word w. The overall matching unit 620 outputs the calculated likelihood P (X | w) to the integrated evaluation value calculation unit 660.

  Next, calculation of the feature probability will be described. The feature probability calculation unit 630 in FIG. 2 outputs a feature X from an arbitrary word based on the feature X extracted by the feature extraction unit 605 and the arbitrary word model stored in the model storage unit 643. Probability P (X) is calculated.

  When an arbitrary word is c *, the feature probability P (X) can be expressed as P (X) = P (X | c *). That is, the feature probability P (X) indicates the probability that the feature X is output from an arbitrary word c *. Therefore, the feature probability calculation unit 630 can calculate the feature probability P (X) using the Viterbi algorithm described above. The feature probability P (X) is a constant value for each feature X regardless of the word.

  A model (arbitrary word model) corresponding to an arbitrary word c * is generated by a method using, for example, an ergodic hidden Markov model (ergic HMM). The arbitrary word model is generated in advance and stored in the model storage unit 643. That is, the feature probability calculation unit 630 acquires an arbitrary word model stored in the model storage unit 643, and calculates the feature probability P (X) by the above Viterbi algorithm based on the arbitrary word model and the feature X. To do. The feature probability calculation unit 630 outputs the calculated feature probability P (X) to the integrated evaluation value calculation unit 660.

  FIG. 7 shows an example of an ergodic model. The arbitrary word model is generated by learning parameters using an ergodic model that allows transition between arbitrary states as shown in FIG.

  Moreover, the arbitrary word model can also be comprised using the character model of each character, for example. As an example of the configuration method, an arbitrary character string can be represented by connecting the character models of all characters in parallel and allowing the transition from the model end to the head so that the arbitrary number of repetitions is possible. An example of this model is shown in FIG. Transition probabilities between character models connected in parallel may be equal, or an arbitrary value may be set. The configuration method is not limited to this, and an arbitrary configuration method is conceivable. For example, if there are frequently appearing partial character strings such as “or”, “er”, and “st”, a model representing the partial character string may be added to the parallel connection.

  The example shown in FIG. 7 has an advantage that the likelihood calculation process can be performed at high speed because the number of states is smaller than that in FIG. On the other hand, since the example shown in FIG. 7 needs to store the ergodic model parameters separately from the character model, the example shown in FIG. 8 that can divert the parameters of the character model can reduce the memory capacity. There is an advantage.

  The integrated evaluation value calculation unit 660 integrates the calculation results of the analytical matching unit 610, the overall matching unit 620, and the feature probability calculation unit 630. The integrated evaluation value calculation unit 660 includes the posterior probability ratio P (Y | w) / P (Y) calculated by the analytical matching unit 610 and the likelihood P (X | w) calculated by the overall matching unit 620. And the integrated probability value {P (Y | w) / P (Y)} · {P (X | w) / P (X) based on the feature probability P (X) calculated by the feature probability calculation unit 630 )} Is calculated.

  That is, the integrated evaluation value calculation unit 660 multiplies the posterior probability ratio P (Y | w) / P (Y), the likelihood P (X | w), and the inverse of the feature probability P (X). The integrated evaluation value calculation unit 660 outputs the calculated integrated evaluation value {P (Y | w) / P (Y)} · {P (X | w) / P (X)} to the prior probability multiplication unit 670.

  Prior probability multiplication section 670 applies the integrated evaluation value {P (Y | w) / P (Y)} · {P (X | w) / P (X)} calculated by integrated evaluation value calculation section 660 for each word. Is multiplied by the prior probability P (w). Thereby, the word recognizing unit 600 can obtain the posterior probability P (w | Y, X) as the calculation result of Equation 1 above.

  Prior probability storage unit 652 stores prior probability P (w) for each word as a table. The prior probability P (w) is a probability indicating the frequency with which a word in the paper sheet 1 is described. By adjusting this value and creating a table, the posterior probability P (w | Y, X) of a word inappropriate as an address can be suppressed.

  For example, a barcode on the paper sheet 1 may be recognized as a word such as “11111111”. Even in such a case, by setting a low value as a prior probability P (w) in a word such as “11111111” in advance, the word recognition unit 600 can determine the posterior probability P (w |) of the word “11111111”. Y, X) can be prevented from being calculated as a high value. That is, by setting in advance a low value as the prior probability P (w) for a word that is easily misrecognized, it is possible to prevent the word recognition unit 600 from causing erroneous recognition.

  For example, when the appearance frequency of all the words is uniform, the prior probability P (w) may be a constant value.

  Prior probability multiplication section 670 reads prior probability prior probability P (w) stored for each word in prior probability storage section 652, and integrates evaluation values {P (Y | w) / P (Y)} · {P Multiply (X | w) / P (X)}. Prior probability multiplication section 670 outputs the result of multiplication, that is, posterior probability P (w | Y, X), to main control section 500.

  Through the above processing, the main control unit 500 can acquire a recognition result (evaluation value) for each word. The main control unit 500 can specify the destination information in consideration of the posterior probabilities of a plurality of words and combinations with other words. For example, the main control unit 500 can infer a combination of words suitable as destination information.

  As described above, the word recognizing unit 600 calculates the posterior probability ratio P (Y | w) / P (Y) by analytical matching by the above-described units, and calculates the likelihood P (X | w) by overall matching. The feature probability P (X) is calculated by calculating the feature probability. The word recognition unit 600 includes a posteriori probability ratio P (Y | w) / P (Y), likelihood P (X | w), feature probability P (X), and a prior probability P set in advance for each word. By integrating (w), the posterior probability P (w | Y, X) for each word can be calculated.

  Note that the word recognition unit 600 may be configured to output the word for which the highest posterior probability P (w | Y, X) is calculated to the main control unit 500 as a recognition result. In this case, the word recognition unit 600 can identify one word as a recognition result and transmit it to the main control unit 500.

  Further, as described above, the word recognition unit 600 may be configured to output the posterior probability P (w | Y, X) for each word to the main control unit 500 as a recognition result. In this case, the main control unit 500 can specify the destination information in consideration of the posterior probabilities P (w | Y, X) of a plurality of words and combinations with other words.

Next, the learning phase will be described.
The VCS 640 shown in FIG. 2 is a module for causing the operator of the paper sheet processing apparatus 100 to input correct destination information of the paper sheet 1 whose destination information has not been recognized by the word recognition unit 600, for example. The VCS 640 includes, for example, an operation unit 700 and a display unit 800 illustrated in FIG. Further, for example, the word recognition unit 600 may be configured to include a module capable of operation and display as the VCS 640 separately from the operation unit 700 and the display unit 800.

  The VCS 640 displays an image of the paper sheet 1 whose destination information could not be specified. The VCS 640 causes the operator to read the displayed image of the paper sheet 1 and input destination information. For example, the VCS 640 causes the operator to input a correct word for each word candidate. As a result, the VCS 640 can associate the word image with the correct destination information (correct answer).

  The VCS 640 outputs the word image and correct destination information (correct answer) to the first word image storage unit 641 and the prior probability calculation unit 651.

  First, word model learning will be described. The first word image storage unit 641 stores the word image input by the VCS 640 and the correct answer in association with each other.

  The model learning unit 642 learns one or more of each character model, each word model, and an arbitrary character model using the word image stored in the first word image storage unit 641 and its correct answer.

  The model learning unit 642 performs model learning using, for example, a Baum-Welch algorithm. The Baumwelch algorithm is an algorithm that searches for unknown parameters in a hidden Markov model. The Baumwelch algorithm can estimate model parameters from the sequence output by the model.

  For example, the model learning unit 642 generates a model by the Baum Welch algorithm using the word images stored in the first word image storage unit 641 and the correct answer thereof. The model learning unit 642 outputs the generated model to the model storage unit 643. The model storage unit 643 stores the received model.

  The model learning unit 642 may be configured to update a model already stored in the model storage unit 643.

  Next, learning of prior probabilities will be described. The prior probability calculation unit 651 counts the frequency for each word based on the correct destination information of the word image input by the VCS 640. That is, the prior probability calculation unit 651 calculates the prior probability P (w) for each word by counting and counting the number of words included in the destination information for each word. The prior probability calculation unit 651 stores the calculated prior probability P (w) for each word in the prior probability storage unit 652.

  Prior probability input section 653 can change prior probability P (w) stored in prior probability storage section 652. The prior probability input unit 653 rewrites the prior probability P (w) stored in the prior probability storage unit 652 to a value corresponding to the operation based on, for example, an operation input by the operation unit 700 illustrated in FIG.

  The prior probability input unit 653 is a value corresponding to the prior probability P (w) stored in the prior probability storage unit 652 based on an operation input by a module that can be operated separately from the operation unit 700. The configuration may be rewritten as follows.

  Thereby, a low value can be set as the prior probability P (w) for the above-described words that are easily misrecognized. Thereby, it is possible to prevent the word recognition unit 600 from causing erroneous recognition.

  According to such a configuration, the word recognizing unit 600 can use an analytical technique (analytic matching) and an overall technique (global matching) in combination. Further, the word recognition unit 600 can calculate the posterior probability P (w | Y, X) with higher accuracy by calculating the feature probability P (X) based on the arbitrary word model as described above. it can. As a result, it is possible to provide a word recognition device, a word recognition method, and a paper sheet processing device including the word recognition device that can recognize words with higher accuracy.

  In the above embodiment, either the analytical matching or the overall matching may be performed first. In addition, when the word recognition unit 600 includes a configuration capable of processing the analytical matching and the global matching in parallel, the configuration may be such that the analytical matching and the global matching are processed in parallel.

  In the above embodiment, when the word recognition unit 600 specifies one word as the recognition result, the main control unit 500 uses the word for which the highest posterior probability P (w | Y, X) is calculated as the recognition result. However, the present invention is not limited to this configuration. Since the feature probability P (X) is constant when the word images are the same, the word recognition unit 600 calculates the posterior probability P (w | Y, X) with P (X) in Equation 1 as an arbitrary value. It may be configured to. In other words, the word recognition unit 600 determines the posterior probability P (w | Y) based on the posterior probability ratio P (Y | w) / P (Y), the likelihood P (X | w), and the prior probability P (w). X) can be calculated.

  Further, as described above, when outputting the posterior probability P (w | Y, X) for each word to the main control unit 500, which is the upper level, the word recognition unit 600 calculates the feature probability P (X), A posteriori probability P (w | Y, X) is calculated based on Equation 1. Thereby, the word recognition unit 600 can output the posterior probability P (w | Y, X) as the evaluation of each word to the main control unit 500. The main control unit 500 can specify the destination information with higher accuracy in consideration of the posterior probability P (w | Y, X) for each word and the combination of each word.

  In the embodiment described above, the word recognition unit 600 determines the posterior probability based on the posterior probability ratio P (Y | w) / P (Y), the likelihood P (X | w), and the prior probability P (w). Although described as a configuration for calculating P (w | Y, X), it is not limited to this configuration. For example, when it is not necessary to consider the prior probability or when the prior probability is a constant value, the word recognition unit 600 ignores the prior probability P (w) of Equation 1 or replaces it with a predetermined value, and the posterior The probability P (w | Y, X) may be calculated.

(Second Embodiment)
FIG. 9 shows an example of the configuration of the word recognition unit 600 according to the second embodiment.
The word recognition unit 600 includes an image receiving unit 601, a word extraction unit 602, a character candidate extraction unit 603, a character recognition unit 604, a feature extraction unit 605, an analytical matching unit 610, an overall matching unit 620, a feature probability calculation unit 630, VCS 640, first word image storage unit 641, model learning unit 642, model storage unit 643, word model generation unit 644, word dictionary 645, prior probability calculation unit 651, prior probability storage unit 652, prior probability input unit 653, integration An evaluation value calculation unit 660, a prior probability multiplication unit 670, a second word image storage unit 681, a parameter learning unit 682, and a parameter storage unit 683 are provided.

  The feature probability calculation unit 630 includes a head feature probability calculation unit 631, a conditional feature probability calculation unit 632, a joint probability feature calculation unit 633, a previous feature probability calculation unit 634, and a total product calculation unit 635. The same components as those in the first embodiment are denoted by the same reference numerals, and detailed description thereof is omitted.

  Note that the operation of the word recognition unit 600 according to the second embodiment is roughly divided into a recognition phase and a learning phase. First, the recognition phase will be described.

The feature probability calculation unit 630 calculates the feature probability P (X) based on the feature X extracted by the feature extraction unit 605 and the parameter stored by the parameter storage unit 683. As described above, the feature extraction unit 605 extracts the feature X, which is a set of vectors, based on the word candidate images. This feature X has T feature vectors x ₁ , x ₂ , x ₃ ... X _T. In this case, the feature vector _xt indicates the t-th feature vector. A feature vector x _t-1 indicates a feature vector immediately before the feature vector x _t .

Assuming the above, the feature probability calculation unit 630 calculates the feature probability P (X) based on the following Equation 3.

The head feature probability calculation unit 631 of the feature probability calculation unit 630 calculates the first factor P (x ₁ ) on the right side of Equation 3. The first factor P (x ₁ ) indicates the probability (first feature probability) that x ₁ is extracted as the _first feature vector. The leading feature probability calculation unit 631 calculates the first factor P (x ₁ ) based on the parameters stored in the parameter storage unit 683.

The parameter storage unit 683 accumulates parameters calculated by learning based on a plurality of word images. This parameter indicates the probability distribution formed by the feature vectors x ₁ , x ₂ , x ₃ ... Of the feature X extracted based on the word image. That is, the parameter storage unit 683 stores the probability distribution parameters formed by the feature vectors. The parameter storage unit 683 stores, for example, the mixing ratio, average vector, or covariance matrix of each Gaussian distribution if the model is modeled with a mixed Gaussian distribution.

The condition feature probability calculation unit 632 calculates the individual factor P (x _t | x _t−1 ) that is the content of the second factor bag on the right side of Equation 3. P (x _t | x _t−1 ) indicates a conditional probability that each feature vector excluding the first feature vector appears on the condition that the feature vector arranged immediately before the feature vector. That is, the feature probability calculation unit 630 calculates P (x _t | x _t−1 ) according to the combination of T−1.

Note that P (x _t | x _t−1 ) can be expressed as shown in Equation 4 below.

The conditional feature probability calculation unit 632 includes a joint probability feature calculation unit 633 and a previous feature probability calculation unit 634. The joint probability feature calculation unit 633 calculates P (x _t , x _t−1 ), which is the numerator on the right side of Equation 4. P (x _t , x _t−1 ) indicates the probability (simultaneous probability) that the previous vector x _t−1 and the feature vector x _t appear simultaneously. The joint probability feature calculation unit 633 calculates P (x _t , x _t−1 ) based on the parameters stored in the parameter storage unit 683.

Further, the previous feature probability calculation unit 634 calculates P (x _t−1 ) that is the denominator of the right side of Equation 4. P (x _t−1 ) indicates the probability (previous feature probability) that the previous vector x _t−1 appears. The previous feature probability calculation unit 634 calculates P (x _t−1 ) based on the parameters stored by the parameter storage unit 683.

The total product calculation unit 635 multiplies the output of the head feature probability calculation unit 631 by all the outputs of the condition feature probability calculation unit 632. That is, the total product calculation unit 635 multiplies P (x ₁ ) by P (x _t | x _t−1 ) from t = 2 to T. Thereby, the feature probability calculation unit 630 can calculate the right side of Equation 3. As a result, the feature probability calculation unit 630 can calculate the feature probability P (X). The feature probability calculation unit 630 outputs the calculated feature probability P (X) to the integrated evaluation value calculation unit 660. The subsequent processing is the same as in the first embodiment.

Next, the learning phase will be described.
The VCS 640 shown in FIG. 9 has the same configuration as the VCS 640 shown in FIG. However, the VCS 640 according to the second embodiment may be configured to be able to collect at least the word images of the paper sheet 1 whose words are not recognized by the word recognition unit 600. The VCS 640 outputs the word image to the second word image storage unit 681.

  The second word image storage unit 681 stores the word image input by the VCS 640. Moreover, the structure which stores directly the image of the paper sheet 1 read by the image reading part 400 shown in FIG. 1 may be sufficient. Further, the second word image storage unit 681 may be configured in the same manner as the first word image storage unit 641.

The parameter learning unit 682 uses parameters used by the head feature probability calculation unit 631, the joint probability feature calculation unit 633, and the previous feature probability calculation unit 634 based on the word images stored in the second word image storage unit 681. To learn. That is, the parameter learning unit 682 calculates a plurality of feature vectors x ₁ , x ₂ , x ₃ ... Based on the word image, and learns probability distribution parameters from the plurality of feature vectors. The parameter learning unit 682 stores the learned parameter in the parameter storage unit 683.

  According to such a configuration, the word recognition unit 600 learns in advance the parameters of the probability distribution of the feature vector based on the feature of the word image. The word recognition unit 600 calculates the feature probability P (X) based on the learned parameter. By calculating the posterior probability P (w | Y, X) using the feature probability P (X) calculated in this way, the word recognition unit 600 can increase the posterior probability P (w | Y, X) with higher accuracy. ) Can be calculated. As a result, it is possible to provide a word recognition device, a word recognition method, and a paper sheet processing device including the word recognition device that can recognize words with higher accuracy.

  In the above-described embodiment, the word recognition unit 600 has been described as a configuration in which the calculations of Formulas 1 to 4 are performed as they are, but the configuration is not limited thereto. The word recognition unit 600 may be configured to calculate by taking the logarithm of each term of Equations 1 to 4. By using the logarithm in this way, the part that was a multiplication is replaced with addition. Also, the part that was the division is replaced with subtraction.

  It should be noted that the functions described in the above embodiments are not limited to being configured using hardware, but can be realized by causing a computer to read a program describing each function using software. Each function may be configured by appropriately selecting either software or hardware.

  Note that the present invention is not limited to the above-described embodiment as it is, and can be embodied by modifying the constituent elements without departing from the scope of the invention in the implementation stage. Further, various inventions can be formed by appropriately combining a plurality of constituent elements disclosed in the embodiment. For example, some components may be deleted from all the components shown in the embodiment. Furthermore, you may combine suitably the component covering different embodiment.

  DESCRIPTION OF SYMBOLS 1 ... Paper sheets, 100 ... Paper sheet processing apparatus, 200 ... Supply part, 210 ... Separation roller, 220 ... Conveyance path, 300 ... Sorting processing part, 400 ... Image reading part, 500 ... Main control part, 600 ... Word Recognizing unit, 601 ... Image receiving unit, 602 ... Word extracting unit, 603 ... Character candidate extracting unit, 604 ... Character recognizing unit, 605 ... Feature extracting unit, 610 ... Analytical matching unit, 611 ... Character probability calculating unit, 612 ... 1st operation part, 613 ... 2nd operation part, 620 ... Overall matching part, 630 ... Feature probability calculation part, 631 ... Leading feature probability calculation part, 632 ... Conditional feature probability calculation part, 633 ... Simultaneous probability feature calculation Part 634 ... previous feature probability calculation part 635 ... total product calculation part 640 ... VCS, 641 ... first word image storage part, 642 ... model learning part, 643 ... model storage part, 644 ... word model generation part, 45 ... Word dictionary, 651 ... Prior probability calculation unit, 652 ... Prior probability storage unit, 653 ... Prior probability input unit, 660 ... Integrated evaluation value calculation unit, 670 ... Prior probability multiplication unit, 681 ... Second word image storage unit 682 ... Parameter learning unit, 683 ... Parameter storage unit, 700 ... Operation unit, 800 ... Display unit, 900 ... Input / output unit.

Claims (14)

A word dictionary for storing multiple words,
An image receiving means for receiving an image including a word;
Word image extraction means for extracting a word image for each word from the image;
Character candidate extraction means for extracting character candidates from the word image;
Character recognition means for performing character recognition on the character candidates;
Analytical matching means for calculating a first evaluation value for each word stored in the word dictionary based on a result of character recognition by the character recognition means;
Feature extraction means for extracting features from the word image;
Word model generation means for generating a word model for each word stored in the word dictionary;
An overall matching means for calculating a second evaluation value indicating a probability that the feature appears for each word model;
A feature probability calculating means for calculating a feature probability that the feature appears;
Integrated evaluation value calculation means for calculating a third evaluation value by multiplying the first evaluation value, the second evaluation value, and the inverse of the feature probability;
Output means for outputting the third evaluation value calculated by the integrated evaluation value calculating means;
A word recognition device comprising:   The word recognition apparatus according to claim 1, wherein the feature probability calculation unit calculates a probability that the feature appears from an arbitrary word model representing an arbitrary word as the feature probability.   The word recognition device according to claim 2, wherein the arbitrary word model allows transition between arbitrary states in the model. First word image storage means for storing a word image and a correct answer of the word image;
Model learning means for learning the word model and the arbitrary word model using the word image and the correct answer stored by the first word image storage means;
The word recognition device according to claim 3, further comprising: Model storage means for storing a character model for each character;
The word model generation means generates the arbitrary word model using the character model and the word model for each word stored in the word dictionary.
The word recognition device according to claim 2. First word image storage means for storing a word image and a correct answer of the word image;
Model learning means for learning the character model using the word images and correct answers stored by the first word image storage means;
The word recognition device according to claim 5, further comprising: The feature extraction means extracts a plurality of feature vectors forming a permutation as features,
The feature probability calculation means calculates a leading feature probability that the leading feature vector of the feature appears, and each feature vector excluding the leading feature vector appears on the condition that the feature vector is arranged immediately before the feature vector. Calculating a conditional probability for each feature vector, and calculating the feature probability based on the leading feature probability and the conditional probability;
The word recognition device according to claim 1. The feature probability calculating means calculates each previous feature probability that a feature vector arranged immediately before each feature vector appears, and each feature vector and a feature vector arranged immediately before the feature vector appear simultaneously. Calculating a joint probability for each feature vector, and calculating the feature probability based on the head feature probability, each joint probability, and each previous feature probability;
The word recognition device according to claim 7. A second word image storage means for storing a word image;
Parameter learning means for learning a parameter used for the head feature probability calculation and a parameter used for the conditional probability calculation using the word image stored by the second word image storage means;
Further comprising
The word recognition device according to claim 8. Prior probability storage means for storing prior probabilities for each word;
Posterior probability calculation means for calculating a fourth evaluation value based on the prior probability stored in the prior probability storage means and the third evaluation value;
Further comprising
The output means outputs the fourth evaluation value calculated by the posterior probability calculation means;
The word recognition device according to any one of claims 1 to 9. A prior probability input means for inputting a value of the prior probability;
The prior probability storage means changes the value of the stored prior probability to the value input by the prior probability input means.
The word recognition device according to claim 10. A prior probability calculating means for receiving a recognition result of the word specified by the paper sheet processing apparatus provided with the word recognition device, and calculating the value of the prior probability for each word based on the received recognition result; Equipped,
The prior probability storage means changes the value of the stored prior probability to the value calculated by the prior probability calculation means.
The word recognition device according to claim 10 or 11. A word recognition method used in a word recognition device having a word dictionary for storing a plurality of words,
Take an image containing a word,
Extracting a word image for each word from the image;
Extracting character candidates from the word image;
Character recognition is performed on the character candidates,
Calculating a first evaluation value for each word stored in the word dictionary based on the result of the character recognition;
Extracting features from the word image;
Generating a word model for each word stored in the word dictionary;
Calculating a second evaluation value indicating a probability that the feature appears for each word model;
Calculating a feature probability that the feature appears;
Multiplying the first evaluation value, the second evaluation value, and the inverse of the feature probability to calculate a third evaluation value;
Outputting the third evaluation value;
Word recognition method. A capturing means for capturing paper sheets;
Conveying means for conveying the paper sheet;
Image reading means for reading an image including a word on the paper sheet;
A word dictionary for storing multiple words,
Word image extraction means for extracting a word image for each word from the image;
Character candidate extraction means for extracting character candidates from the word image;
Character recognition means for performing character recognition on the character candidates;
Analytical matching means for calculating a first evaluation value for each word stored in the word dictionary based on character recognition by the character recognition means;
Feature extraction means for extracting features from the word image;
Word model generation means for generating a word model for each word stored in the word dictionary;
An overall matching means for calculating a second evaluation value indicating a probability that the feature appears for each word model;
A feature probability calculating means for calculating a feature probability that the feature appears;
Integrated evaluation value calculation means for calculating a third evaluation value by multiplying the first evaluation value, the second evaluation value, and the inverse of the feature probability;
A recognition unit for recognizing the destination information of the paper sheet based on the third evaluation value calculated by the integrated evaluation value calculating unit;
A sorting processing unit for sorting the paper sheets based on the destination information recognized by the recognition unit;
A paper sheet processing apparatus comprising:

Images