JP7206605B2 - Information processing equipment - Google Patents

Description

The present invention relates to an information processing device.

The method disclosed in Patent Document 1 performs character recognition on an image on an input form, obtains a degree of similarity as a result of the character recognition, and compares the obtained degree of similarity with the previously registered character recognition. , and based on the result of this comparison, the character recognition result is output without requiring manual verification processing, or the character recognition result is output as a character recognition candidate based on the result of the above comparison. Alternatively, based on the result of the comparison, manual new input and confirmation of the character recognition result are presented to prompt the manual input process.

In the method disclosed in Patent Document 2, when changing from the old version of the character recognition software to the new version of the character recognition software, during the transition from the old version of the software to the new version of the software, both the new version and the old version of the software are used in the actual system. Recognize characters. As a result, information on the recognition accuracy of both the new version and the old version of the software is statistically collected, and the recognition accuracy of both is compared. Then, when the accuracy of the new version is higher than the accuracy of the old version, the introduction of the new version software is decided. On the other hand, if the recognition accuracy of the old version software is better, it is possible to operate in parallel using the good points of both the old and new versions without completely changing to the new version software.

In the method disclosed in Patent Document 3, character information is read from an input document by OCR, and recognition processing is performed by a recognition processing unit. The operator inputs the character information on the input document from the keyboard, and the CPU compares the key-inputted character data with the character-recognized recognition data. The configuration is such that verify input is performed by displaying an abnormality. For example, character data that is judged to have an error in the recognition data because the character data entered with the key matches the input document, and character data that is judged to have an error in not only the recognition data but also the character data entered with the key. is displayed as an error by inverting (whitening out), and input data with a high possibility of input error can be automatically detected.

The apparatus disclosed in Patent Document 4 has an image reading means for reading a form (form) in which data is entered as an electronic image form, and a read electronic image form having different properties, i.e., misrecognition is not common or OCR recognition means that performs OCR recognition with two (or more) OCR engines that are rarely used, and characters whose recognition results match are automatically stored in the database, and even if they match, characters that do not match OCR and database storage means for storing characters with low engine recognition reliability in a database after confirmation and correction.

The classification means of the information processing apparatus disclosed in Patent Document 5 classifies a character recognition object into one of three types, and the extraction means performs the character recognition when the character recognition target is classified into the first type by the classification means. extracting a target character recognition result, and extracting the character recognition result of the character recognition target when the classification means classifies the character recognition target into the second type, and extracting the character recognition target manually. The second control means controls to manually input the character recognition target by a plurality of persons when the character recognition target is classified into the third type by the classification means.

Patent Documents 6 to 11 disclose various calculation methods for the recognition accuracy of character recognition.

Japanese Patent Application Laid-Open No. 2003-346080 JP 2004-171326 A JP-A-05-274467 JP 2010-073201 A JP 2016-212812 A JP-A-5-40853 JP-A-5-20500 JP-A-5-290169 JP-A-8-101880 JP-A-9-134410 JP-A-9-259226

In the case of judging an input by a judging means, in order to obtain the accuracy rate of the judgment of the judging means, for example, the judging result of the judging means for each input is judged by a method with higher judging accuracy (for example, a human check). There is a method of judging whether or not the input is correct, and calculating the proportion of correct judgment results for all of those inputs. However, the determination by the method with high determination accuracy is more expensive than the determination by the determination means. Otherwise, a method with high determination accuracy should be used instead of the determination means from the beginning. Therefore, it is costly to make determinations using this method for all inputs.

The present invention obtains the correct answer rate of the judging means at a lower cost than the method of obtaining the correct answer rate of the judging means by judging the correct/wrong answer of the judging result of the judging means for all inputs by a different method. for the purpose.

The invention according to claim 1 comprises recognition means for executing character recognition for an input, outputting the recognition result and recognition accuracy of the character recognition, and confirming whether the recognition result is correct or wrong, and confirming that the recognition result is correct. In the case of , the recognition result is adopted, in the case of an error, a correct recognition result for the input is obtained, and confirmation means for adopting the obtained recognition result; output control means for outputting the recognition result of the recognition means without intervening, and outputting the recognition result adopted by the confirmation means if the recognition accuracy is less than the threshold; accuracy rate calculation means for calculating a ratio of inputs within a first range that are confirmed as correct by the confirmation means as an accuracy rate of the recognition means in the first range; and the accuracy rate in the first range. estimating means for estimating the accuracy rate of the recognizing means in a second range out of the range equal to or greater than the threshold based on the above.

The invention according to claim 2 is the information processing apparatus according to claim 1, wherein the first range is a range from a value greater than 0 determined according to a predetermined criterion to the threshold value.

In the invention according to claim 3, the estimation means is configured such that the accuracy rate calculated by the accuracy rate calculation means corresponds to a first representative value of the recognition accuracy in the first range, and the second The accuracy rate corresponding to the second representative value of the recognition accuracy in the range is set between the accuracy rate corresponding to the first representative value and a predetermined maximum accuracy rate at the maximum value that the recognition accuracy can take. 3. The information processing apparatus according to claim 1, wherein estimation is performed by linear interpolation.

In the invention according to claim 4, the accuracy rate calculation means obtains the accuracy rate for each of a plurality of ranges in which the recognition accuracy is less than the threshold value, and the estimation means calculates the accuracy rate for each of the plurality of ranges. 3. The information processing apparatus according to claim 1, wherein said accuracy rate in said second range is estimated based on a tendency of change according to said recognition accuracy of .

In the invention according to claim 5, the accuracy rate calculation means obtains the accuracy rate for each of a plurality of ranges in which the recognition accuracy is less than the threshold value, and the estimation means calculates the accuracy rate for each of the plurality of ranges. and the recognition accuracy, estimating a function for obtaining the accuracy rate corresponding to the recognition accuracy, and estimating the accuracy rate in the second range using the estimated function. is an information processing device.

In the invention according to claim 6, the estimation means obtains a probability density function of the recognition accuracy from the distribution of occurrence frequencies of the recognition accuracy, and estimates the accuracy rate in the second range using the probability density function. An information processing apparatus according to claim 1 .

According to a seventh aspect of the present invention, a computer comprises recognition means for executing character recognition on an input, outputting the recognition result and recognition accuracy of the character recognition , confirming whether the recognition result is correct or wrong, and If the input is correct, the recognition result is adopted; if the input is incorrect, the correct recognition result for the input is obtained and the obtained recognition result is adopted ; Output control means for controlling to output the recognition result of the recognition means without intervening, and to output the recognition result adopted by the confirmation means if the recognition accuracy is less than the threshold; accuracy rate calculation means for calculating a ratio of inputs within a first range that are confirmed as correct by the confirmation means as an accuracy rate of the recognition means in the first range; Based on this, the program functions as estimation means for estimating the accuracy rate of the recognition means in a second range out of the range equal to or greater than the threshold value.

According to the invention according to claim 1, 3 or 7, it is lower than the method of obtaining the correct answer rate of the judgment means by judging the correct/wrong answer of the judgment result of the judgment means for all inputs by another method. The accuracy rate of the judging means can be obtained from the cost.

According to the invention according to claim 2, the second range having higher validity than the case of calculating the accuracy rate for the entire range from 0 to the threshold value and estimating the accuracy rate of the second range based on this accuracy rate can be estimated.

According to the fourth, fifth or sixth aspect of the present invention, it is possible to estimate an accuracy rate with higher validity than in the case of estimating the accuracy rate in the second range by linear interpolation.

It is a figure which illustrates the functional structure of the information processing apparatus of embodiment. FIG. 10 is a diagram for explaining an example of a method of estimating the accuracy rate in an area where the recognition accuracy is equal to or higher than a threshold; FIG. 4 is a diagram for explaining how to calculate a probability density function of recognition accuracy; FIG. 10 is a diagram for explaining another example of a method of estimating the accuracy rate in an area where the recognition accuracy is equal to or higher than the threshold; FIG. 11 is a diagram for explaining still another example of the method of estimating the accuracy rate in an area where the recognition accuracy is equal to or higher than the threshold; It is a figure which illustrates the internal structure of a confirmation process part.

FIG. 1 shows an embodiment of an information processing apparatus according to the present invention.

This information processing apparatus determines a character string included in input image data (“input image data”) using an OCR 10 and a confirmation processing unit 18 .

The OCR 10 includes a recognition processing section 12 and a recognition accuracy calculation section 14 . The recognition processing unit 12 recognizes character strings included in the input image data by performing known OCR (optical character recognition) processing on the input image data. The recognition processing unit 12 outputs a text code indicating the recognized character string. The recognition accuracy calculator 14 calculates the recognition accuracy of the text code recognized from the input image data. The recognition accuracy is a degree of likelihood that the text code of the recognition result correctly represents the character string (which may be handwritten) included in the input image data. The higher the recognition accuracy, the higher the possibility that the text code of the recognition result is correct (that is, correctly represents the character string in the input image data). The possibility that the recognition result is correct is hereinafter referred to as the recognition rate or the accuracy rate. The OCR 10 may output a plurality of different recognition results of the input image data in association with the recognition accuracy in descending order of recognition accuracy. The unit in which the OCR 10 performs character recognition (that is, the unit in which the recognition result is output) is not particularly limited. Any unit or the like may be used.

The method of character recognition used by the OCR 10 and the method of calculating the recognition accuracy are not particularly limited, and any of conventional methods such as those exemplified in Patent Documents 6 to 11 and methods to be developed in the future may be used. .

The selection unit 16 controls the output of the character recognition result based on the recognition accuracy calculated by the recognition accuracy calculation unit 14 for the character recognition result (text code) of the recognition processing unit 12 . That is, if the recognition accuracy is equal to or higher than a certain threshold, the selection unit 16 outputs the character recognition result as the final character recognition result of the information processing apparatus itself. If the recognition accuracy is equal to or higher than the threshold, the recognition by the recognition processing unit 12 is trusted to be accurate.

On the other hand, if the recognition accuracy is less than the threshold, the selection unit 16 passes the character recognition result and the corresponding input image data to the confirmation processing unit 18, and confirms whether the character recognition result is correct. Run.

In one example, the confirmation processing unit 18 presents the input image data and the character recognition result to a person in charge of confirmation, and has the character recognition result confirm whether or not the character string in the input image data is correct. The person in charge of confirmation may operate a terminal connected to the information processing apparatus via a network such as the Internet. is sent to the terminal (for example, a web page), and the confirmation person's input for the screen information is accepted. If the person in charge of confirmation determines that the information is correct, he or she makes an input to the confirmation processing unit 18 indicating that fact. Output as the final character recognition result. Also, at this time, the confirmation processing unit 18 accumulates in the accumulation unit 20 confirmation result information indicating that the character recognition result of the recognition processing unit 12 is correct.

When the person in charge of confirmation determines that the character recognition result received from the selection unit 16 is not correct as a character string in the input image data, the person in charge of confirmation inputs to the confirmation processing unit 18 to correct the character recognition result. In response to this, the confirmation processing unit 18 outputs the corrected character recognition result as the final character recognition result of the information processing apparatus itself. Also, at this time, the confirmation processing unit 18 accumulates in the accumulation unit 20 confirmation result information indicating that the character recognition result of the recognition processing unit 12 is incorrect.

In the above, the case where a human confirms the character recognition result of the OCR 10 is exemplified. The confirmation may be performed using a pay high-precision OCR service on the Internet operated by a management entity other than the user. In this case, the confirmation processing unit 18 causes the other OCR to recognize the input image data, receives the recognition result, and outputs the received recognition result as the final character recognition result of the information processing apparatus itself. Further, the confirmation processing unit 18 compares the character recognition result of the recognition processing unit 12 received from the selection unit 16 with the recognition result received from another OCR, and if both match, the recognition processing unit 12 Confirmation result information indicating that the character recognition result is correct is accumulated in the accumulation unit 20, and confirmation result information indicating that the character recognition result is incorrect if both do not match.

In this manner, the confirmation processing unit 18 accumulates in the accumulation unit 20 confirmation result information indicating whether the character recognition result of the recognition processing unit 12 is correct or incorrect. Here, the confirmation processing unit 18 determines whether the character recognition result of the recognition processing unit 12 is correct or incorrect when the recognition accuracy corresponding to the character recognition result is less than the threshold. Therefore, the confirmation result information accumulated in the accumulation unit 20 is the correct/incorrect judgment result for the character recognition result whose recognition accuracy is less than the threshold.

The low-accuracy region accuracy rate calculation unit 22 calculates the low-accuracy region, i.e. The accuracy rate of the recognition processing unit 12 is calculated for the recognition accuracy range less than the threshold. For example, the accuracy rate may be calculated by dividing the number of confirmation result information indicating the correct answer by the total number of confirmation result information for which the accuracy rate is calculated.

The high-accuracy area accuracy rate estimating unit 24 calculates the accuracy rate of the recognition processing unit 12 for the high-accuracy area, that is, the recognition accuracy range equal to or higher than the threshold, based on the accuracy rate in the low-accuracy area calculated by the low-accuracy area accuracy rate calculating unit 22. presume. An example of estimation performed by the high-precision region accuracy rate estimation unit 24 will be described below.

A first example will be described with reference to FIG.

Let the recognition accuracy be a real number from 0 to 1, let U be the representative value of the low accuracy range, and V be the representative value of the high accuracy range. When the median value of each region is used as the representative value of each region, and the threshold used by the selection unit 16 is T, U=T/2 and V=(T+1)/2. In the example of FIG. 2, the accuracy rate (recognition rate) when the recognition accuracy is 1 is 1, and the accuracy rate α in the low-accuracy area calculated by the low-accuracy area accuracy calculation unit 22 is the representative value U , the accuracy rate δ at the representative value V in the high-accuracy region is estimated by linear interpolation. That is, the high-probability region accuracy rate estimating unit 24 obtains the accuracy rate δ using the following equation (1).

In the above, as the representative values U and V of the low-accuracy region and the high-accuracy region, the median values of the respective regions themselves are used, but this is only an example. Alternatively, U and V may be representative values of the frequency distribution of recognition accuracy (or the probability density function obtained from this) in each region. That is, the recognition accuracy calculation unit 14 accumulates the recognition accuracy obtained for each piece of input image data, and uses this accumulated information to calculate the frequency (occurrence frequency) of the recognition accuracy belonging to each recognition accuracy interval. Then, the representative values of the high-accuracy region and the low-accuracy region can be obtained from the frequency distribution (histogram) that can be generated by this. Since only the information in the low-accuracy region is accumulated in the accumulation unit 20, the output of the recognition accuracy calculation unit 14 is separately accumulated in order to obtain the distribution of the recognition accuracy over the entire range. As the representative value of the frequency distribution, for example, an average value, a median value, or a mode value may be used.

Also, using the probability density function p(x) of the recognition accuracy, the representative values U and V as average values may be obtained using the following equation (2).

Here, the probability density function p(x) can be obtained as follows.

That is, as shown in FIG. 3, the recognition accuracy x is first divided into a plurality of intervals. Let Z be the number of sections and W be the width of the section. Let the index of each interval be k. k is an integer of 1 or more and Z or less. The central value of interval k (that is, the value obtained by adding the lower limit and upper limit of the interval and dividing by 2) is defined as the interval representative value x _k . The recognition accuracy calculation unit 14 accumulates the recognition accuracy obtained for each piece of input image data, and from this accumulated information, the occurrence frequency (frequency) Yk of the recognition accuracy within each section _k is obtained. Assuming that the number of pieces of input image data (that is, the number of recognition accuracies) is N, the probability density value p(x) in the section representative value is obtained by the following equation.
p( _{xk)=Yk /} NW

This is a discrete probability density function. A continuous function obtained by interpolating this by a known interpolation method may be used as the probability density function p(x).

An improved estimation method of the high-precision range accuracy rate estimator 24 described with reference to FIG. 2 will now be described with reference to FIG.

In the example of FIG. 2, the accuracy rate in the high-accuracy area is calculated using the accuracy rate in the entire low-accuracy area. However, the accuracy rate in the region where the recognition accuracy is very low has little relevance to the accuracy rate in the high accuracy region. Therefore, in this improved method, the accuracy rate of the high-accuracy area is estimated based on the accuracy rate of only the area close to the threshold value T, not the entire low-accuracy area.

That is, a region lower limit value S that satisfies 1<S<T is determined in advance, and the low-accuracy region accuracy rate calculation unit 22 determines that the recognition accuracy x of the confirmation result information accumulated in the accumulation unit 20 is S≦x≦T. Accuracy rate α is calculated only from those that satisfy How to determine the value of S is not particularly limited. For example, a fixed ratio of less than 1 to the threshold value T may be defined as S. In addition, the data (confirmation result information) in the storage unit 20 are sequentially selected in the direction in which the value of the recognition accuracy x decreases from the threshold T, and the number of selected data is the total number of data below the threshold T. The lower limit value S may be the recognition accuracy x when it reaches a predetermined ratio.

The high-accuracy region accuracy rate estimating unit 24 obtains the representative value U of the recognition accuracy in the region of the recognition accuracy from S to T by the same method as in the above embodiment. Then, assuming that the accuracy rate α of that area is the value at the representative value U, the accuracy rate δ of the high-accuracy area is calculated using the above equation (1).

In this improved method, the accuracy rate of the high-accuracy area is estimated from the accuracy rate of the area close to the high-accuracy area among the low-accuracy areas. rate can be estimated more accurately.

A further modification will be described with reference to FIG.

In this modification, as shown in FIG. 5, the low-accuracy region correct rate calculation unit 22 divides the low-accuracy region into N (N is an integer equal to or greater than 2) small regions, and for each small region, the storage unit 20 is calculated from the confirmation result information corresponding to the recognition accuracy belonging to the small area. Although the low-accuracy region is divided into four small regions in the example of FIG. 5, this is only an example. Then, the low-accuracy area accuracy rate calculation unit 22 calculates the accuracy rate α of the small area at the representative value x of the small area (for example, the accuracy of the middle between the upper limit and the lower limit of the small area) shown).

The high-accuracy region accuracy rate estimating unit 24 estimates the function α(x) by a known technique such as polynomial approximation or curve fitting under the assumption that the accuracy rate α is a function α(x) of the recognition accuracy x. Then, using this function α(x), the accuracy rate δ in the high-accuracy region is estimated by the following equation (3).

Further, the high-accuracy area accuracy rate estimating unit 24 may estimate the accuracy rate δ in the high-accuracy area using the following equation (4) instead of equation (3).

In Equation (4), p(x) is the probability density function p(x) described above. Conversely, Equation (3) is an equation when it is assumed that the probability density function p(x) has a uniform distribution.

Expression (3) or (4) is for obtaining the accuracy rate for the entire range from the threshold value T to 1 in the high-accuracy region, that is, the recognition accuracy x. Generalizing this, the high-accuracy region accuracy rate estimating unit 24 calculates the accuracy rate for the range of T ₁ ≤ x ≤ T ₂ (where T ≤ T ₁ < T ₂ ) in the high-accuracy region by the following equation (5) can be estimated.

A further modification will be described with reference to FIG.

FIG. 6 shows an example of the internal configuration of the confirmation processing unit 18, the storage unit 20, the low-accuracy area accuracy calculation unit 22, and the high-accuracy area accuracy estimation unit 24 in the information processing apparatus of this modification. there is The information processing apparatus of this modified example further comprises an OCR 10 and a selector 16 similar to those shown in FIG.

If the recognition accuracy calculated by the recognition accuracy calculation unit 14 for the input image data is less than the threshold, the selection unit 16 instructs the confirmation processing unit 18 to execute processing. At this time, the selection unit 16 inputs the input image data and the character recognition result of the recognition processing unit 12 for the input image data to the confirmation processing unit 18 . The character recognition result is passed to matching section 184 and the input image data is passed to manual input section 182 .

The manual input unit 182 presents an image indicated by the passed input image data to a human input person, and receives input of a character string read from the image by the input person. The manual input unit 182 can be regarded as a character recognition unit using a human as a character recognition engine. An input person who performs character recognition may be in a remote position with respect to this information processing apparatus via a network such as the Internet. An image indicated by the input image data is provided in the form of, for example, a web page, and the character string of the recognition result input by the user is received via the network. The character string received by the manual input unit 182 from the input person is input to the matching unit 184 .

The matching unit (X) 184 matches (matches) the character recognition result of the recognition processing unit 12 of the OCR 10 with the character string received from the input person by the manual input unit 182 to match (match) the two. Determines whether or not (non-match). If the two match, the matching unit 184 outputs the matching determination result as the final character recognition result of the information processing apparatus. If the two do not match, the matching unit 184 causes the manual input unit 186 to execute processing. The matching unit 184 also accumulates a matching result (that is, a value indicating “match” or “non-match”) X in the storage unit 20 . The value of match result X is a binary value that indicates whether the match is a non-match. In the following, as an example, for convenience of calculation, the value of the matching result X is set to "1" in case of matching and "0" in case of non-matching (the same applies to the matching sections 188A and 188B described later). . The identification information i of the input image data (for example, a serial number assigned to each piece of input data in order) is associated with the matching result X accumulated in the storage unit 20, and which input image data the matching result corresponds to. is identifiable.

When the manual input unit 186 receives the trigger for the case of non-match from the matching unit 184, the input person of the manual input unit 182 presents an image indicated by the input image data to a different second input person. accepts the input of the character string read from the image by the input person. Then, the character string received by the manual input unit 186 from the second input person is output as the final character recognition result of the information processing apparatus for the input image data.

In parallel with the OCR 10 and the manual input unit 182, the manual input unit 186 may perform a process of always accepting a character string input from the second input person for the same input image data. It may be performed only when the determination result is non-coincidence. This reduces the cost for the processing of the manual input unit 186 (for example, the cost for the second input person).

The OCR 10, the manual input unit 182, the matching unit 184, and the manual input unit 186 are recognition mechanisms that perform character recognition on input image data in a low-accuracy area, that is, an area where recognition accuracy is less than a threshold.

On the other hand, the matching units 188A and 188B, the accumulation unit 20, and the low-accuracy region accuracy calculation unit 22, which will be described below, accumulate a large number of determination results performed by the recognition mechanism, and based on the accumulated information, calculate the low-accuracy region , the accuracy rate of the OCR 10 and the manual input unit 182 are calculated respectively. Furthermore, the accuracy rate of the recognition mechanism for the low-accuracy region may be calculated.

That is, first, the matching unit 188A matches the character recognition result of the OCR 10 with the character string received by the manual input unit 186, and stores the matching result (matching result A) in association with the identification information i of the input image data. Store in unit 20 . The matching unit 188B matches the determination result of the manual input unit 182 with the determination result of the manual input unit 186, and stores the matching result (matching result B) in the storage unit 20 in association with the identification information i of the input image data. do.

The accumulation unit 20 accumulates three matching results X _i , A _i , and B _i by the matching units 184 , 188 A, and 188 B for each input data i.

The low-accuracy region correct rate calculation unit 22 uses the matching results _Xi , Ai, and _Bi accumulated in the accumulation unit 20 to calculate the correct answers in the low- _accuracy region of the OCR 10, the manual input unit 182, and the recognition mechanism. Calculate the rate.

A method of calculating the accuracy rate by the low-probability region accuracy rate calculation unit 22 will be described. First, a method of calculating the accuracy rate α of the OCR 12a and the accuracy rate β of the manual input unit 182 will be described.

This calculation method calculates accuracy rates α and β based on the following two premises (a), (b) and (c).
(a) If the matching result X of the matching section 184 is "match", both the recognition results of the OCR 10 and the manual input section 182 are correct.
(b) If the matching result A of the matching section 188A is "match", the recognition result of the OCR 10 is correct.
(c) If the matching result B of the matching unit 188B is “match”, it is the correct input by the inputter accepted by the manual input unit 182 .

That is, here, the OCR recognition result is correct when it matches the character string input to the manual input unit 182 or the manual input unit 186, and the character string input to the manual input unit 182 is recognized by the OCR 10. If the character string matches the result or the character string input to the manual input unit 186, it is regarded as correct, and the accuracy rates α and β are calculated. Based on these premises, the low-probability region accuracy rate calculator 22 calculates the accuracy rates α and β according to the following equation (6).

Here, i is a serial number that is identification information of input image data, and N is the total number of input data. Also, "P|Q" is an operation that gives a value of 1 if P or Q is 1, and gives a value of 0 if both P and Q are 0.

It should be noted that if the matching result of the matching unit 184 is "match", the manual input unit 186 may not be allowed to make the determination. In this case, since the judgment result of the manual input unit 186 cannot be obtained, the matching results of the matching units 188A and 188B using this may both be "0". In this case, the low-probability region accuracy rate calculation unit 22 may calculate the accuracy rate using the following equation (7) instead of the above-described equation (6).

Next, the processing for obtaining the accuracy rate γ of the recognition mechanism (that is, the portion comprising the OCR 10, the manual input section 182, the matching section 184, and the manual input section 186) for the low-accuracy region of this information processing apparatus will be described. Here, it is assumed that the manual input section 182 and the manual input section 186 have the same characteristics. That is, the manual input unit 182 and the manual input unit 186 are considered to have the same accuracy rate statistically.

Assume that the accuracy rates α and β in the low-accuracy region of the OCR 10 and the manual input unit 182 have already been calculated by the method described above. In this example, as described above, the manual input unit 186 can be considered to have the same accuracy rate α as the manual input unit 182 when the number of input data is sufficiently large. Therefore, the low-probability region accuracy rate calculator 22 can calculate the accuracy rate γ using the following equation.
γ=αβ+(1−αβ)α

More specifically, the cases where the judgment mechanism as a whole is correct are (a) cases where the recognition result of the OCR 10 is correct and the input received by the manual input unit 182 is correct, and (b) cases other than this and manual These are two cases where the input unit 186 is correct. The probability of case (a) occurring is αβ, and the probability of case (b) occurring is the product of the probability (1-αβ) other than (a) and the probability α that the manual input unit 186 is correct (1-αβ ) α, the sum of the probabilities of (a) and (b) is the final accuracy rate γ.

The high-accuracy area accuracy rate estimating unit 24 uses the accuracy rate α in the low-accuracy area of the OCR 10 calculated by the low-accuracy area accuracy calculation unit, and uses the method shown in the above-described embodiment or each modification to increase the OCR 10. The accuracy rate is estimated in the accuracy region (that is, the recognition accuracy is greater than or equal to the threshold). Further, when estimating the accuracy rate of the entire system, the above-mentioned γ is used as the accuracy rate in the low-accuracy region, and from this γ, the system in the high-accuracy region is An overall accuracy rate may be estimated.

The confirmation processing unit 18 illustrated in FIG. 6 has a character recognition result ( That is, the accuracy of the output of the confirmation processing unit 18 can be increased, and the accuracy of the accuracy rate of the OCR 10 in the low-accuracy range can be increased.

In the example of FIG. 6, the character recognition result of the OCR 10 is checked by a person, but it may be checked by means other than a person. As confirmation means other than human, for example, a character recognition system that is expected to have a higher accuracy rate of character recognition than OCR 10 may be used. If the use cost of this character recognition system is high and a sufficient accuracy rate can be expected with OCR 10, this mechanism can be used for the purpose of reducing costs by not using the character recognition system.

Both the embodiments and modifications described above recognize character strings in input image data. can be applied to general information processing apparatuses that determine the content of and output the determination result. That is, if the accuracy of determination of determination means (an example of which is an OCR 10) that determines the content of input data, that is, the degree of likelihood that the determination result is correct, is greater than or equal to a threshold value, the determination means In a system that outputs the judgment result as it is, if it is less than the threshold, the judgment result is confirmed by another means, and if it is an error, it is corrected. The methods of the above embodiments and modifications are applicable.

The information processing apparatus exemplified above can be configured as a logic circuit of hardware in one example. As another example, the information processing apparatus may be implemented by causing a built-in computer to execute a program representing the function of each functional module in the system or apparatus. Here, the computer includes, for example, hardware such as a processor such as a CPU, memory (primary storage) such as random access memory (RAM) and read only memory (ROM), HDD controller for controlling HDD (hard disk drive), various It has a circuit configuration in which an I/O (input/output) interface, a network interface for controlling connection with a network such as a local area network, and the like are connected via a bus, for example. Also, for the bus, for example, via an I/O interface, a disk drive for reading from and/or writing to a portable disk recording medium such as a CD or a DVD, a portable nonvolatile recording medium of various standards such as a flash memory, etc. A memory reader/writer for reading from and/or writing to the medium, etc., may also be connected. A program in which the processing contents of each functional module exemplified above is described is stored in a fixed storage device such as a hard disk drive via a recording medium such as a CD or DVD or via a communication means such as a network. Installed. A program stored in a fixed storage device is read out to a RAM and executed by a processor such as a CPU to implement the functional module group illustrated above. Also, the information processing apparatus may be configured by a combination of software and hardware.

10 OCR, 12 recognition processing unit, 14 recognition accuracy calculation unit, 16 selection unit, 18 confirmation processing unit, 20 accumulation unit, 22 low-accuracy area accuracy calculation unit, 24 high-accuracy area accuracy estimation unit, 182, 186 manual input Part, 184, 188A, 188B Abutment.

Claims (7)

recognition means for performing character recognition on an input and outputting the recognition results and recognition accuracy of the character recognition;
confirming means for confirming whether the recognition result is correct or incorrect, adopting the recognition result if the recognition result is correct, obtaining the correct recognition result for the input if the recognition result is incorrect, and adopting the obtained recognition result; ,
Output control for outputting the recognition result of the recognition means without intervening the confirmation means for the input whose recognition accuracy is equal to or higher than the threshold, and outputting the recognition result adopted by the confirmation means if the recognition accuracy is less than the threshold. means and
A correct answer rate for calculating a ratio of inputs confirmed as correct answers by the confirmation means among inputs whose recognition accuracy is within a first range of the range less than the threshold, as a correct answer rate of the recognition means in the first range. calculating means;
an estimation means for estimating the accuracy rate of the recognition means in a second range out of the range equal to or greater than the threshold based on the accuracy rate in the first range;
Information processing equipment including. 2. The information processing apparatus according to claim 1, wherein said first range is a range from a value larger than 0 determined according to a predetermined criterion to said threshold value. The estimating means assumes that the accuracy rate calculated by the accuracy rate calculating means corresponds to a first representative value of the recognition accuracy in the first range, and a second representative value of the recognition accuracy in the second range. The accuracy rate corresponding to the representative value of is estimated by linear interpolation between the accuracy rate corresponding to the first representative value and a predetermined maximum accuracy rate at the maximum value that the recognition accuracy can take. 3. The information processing device according to 1 or 2. The accuracy rate calculation means obtains the accuracy rate for each of a plurality of ranges in which the recognition accuracy is less than the threshold;
3. The information processing according to claim 1, wherein said estimation means estimates said accuracy rate in said second range based on a tendency of change of said accuracy rate in each of said plurality of ranges according to said recognition accuracy. Device. The accuracy rate calculation means obtains the accuracy rate for each of a plurality of ranges in which the recognition accuracy is less than the threshold;
The estimation means estimates a function for obtaining the accuracy rate corresponding to the recognition accuracy from the relationship between the accuracy rate and the recognition accuracy for each of the plurality of ranges, and uses the estimated function to determine the second range. 3. The information processing apparatus according to claim 1, wherein the accuracy rate in is estimated. 2. The information processing according to claim 1, wherein said estimation means obtains a probability density function of said recognition accuracy from a distribution of occurrence frequencies of said recognition accuracy, and estimates said accuracy rate in said second range using said probability density function. Device. the computer,
recognition means for performing character recognition on an input and outputting the recognition result and recognition accuracy of the character recognition;
Confirmation means for confirming whether the recognition result is correct or incorrect, adopting the recognition result if the recognition result is correct, obtaining the correct recognition result for the input if the recognition result is incorrect, and adopting the obtained recognition result;
Output control for outputting the recognition result of the recognition means without intervening the confirmation means for the input whose recognition accuracy is equal to or higher than the threshold, and outputting the recognition result adopted by the confirmation means if the recognition accuracy is less than the threshold. means,
A correct answer rate for calculating a ratio of inputs confirmed as correct answers by the confirmation means among inputs whose recognition accuracy is within a first range of the range less than the threshold, as a correct answer rate of the recognition means in the first range. calculating means;
Estimation means for estimating the accuracy rate of the recognition means in a second range out of the range equal to or greater than the threshold based on the accuracy rate in the first range;
A program to function as

Images