JPH05182014A - Character recognition method - Google Patents

Abstract

PURPOSE:To permit a user to precisely recognize the certainty of a recognized result, to efficiently execute correction work and to easily judge the cause of errorneous recognition. CONSTITUTION:A certainty decision part 9 synthetically decides the certainty of the recognized result with information obtained in respective processing stages 4-7 till a final recognized result is obtained as evidence. A display 11 and a printing device 12 add information on the certainty to the recognized result so as to display or print it. A candidate whose certainty is low is removed in the halfway processing stages if need or a processing is interrupted. Furthermore, certainty till the certain processing stage is reflected on the preceding or succeeding processing.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition method using a plurality of processes, and more particularly to a character recognition method suitable for facilitating evaluation and correction of recognition results by a user.

[0002]

2. Description of the Related Art Basically, a conventional character recognition method is to compare an input pattern with a pattern dictionary and output one having a small difference or a large similarity as a recognition result. However, it is virtually impossible to achieve a recognition rate of 100%. In that case, naturally, the recognition result includes a wrong character, so that the user needs to correct it.

In order to improve the efficiency of the correction work, it is preferable that the apparatus user can easily evaluate how reliable the recognition result is.

In order to achieve this object, an optical character recognition device has been proposed which changes the display color of the recognition result according to the degree of similarity with the pattern dictionary (for example, Japanese Patent Publication No. 61-64).
30 publication).

Further, when a character pattern or a feature amount is stored during recognition and the recognition result is corrected, the dictionary for recognition is learned (updated, added) using the stored pattern or feature amount. A method of performing registration) has been devised.

[0006]

In recent years, the typefaces and character types used for reading originals have increased remarkably, and low-quality copy originals and the like are also increasing for reading. Therefore, only pattern matching is necessary. It has been found that there is a limit to the improvement of recognition rate, and post-processing technology such as clipping technology and language processing is being actively researched. In such a situation, if the certainty of the recognition result is judged only by the conventional pattern matching similarity, the risk of misjudgment is high.

Further, there are a method of correcting the recognition result by interrupting the recognition processing for each unit, and a method of correcting the recognition result after the recognition processing of the entire surface of the document is completed. In either method, conventionally, when the dictionary learning is performed at the time of correcting the recognition result, the pattern or the characteristic amount of all the targeted characters from the start to the interruption or the end of the recognition process is saved. Therefore, in the case of the method of recognizing the entire surface of the original all at once, an extremely large amount of patterns or feature quantities must be stored,
It requires a large storage space such as a RAM or a magnetic disk, and most of the stored information is not used for dictionary learning, which is extremely uneconomical. On the other hand, according to the method in which the recognition process is interrupted and corrected in a certain unit, it is possible to reduce the storage space for storing the pattern or the feature amount by reducing the unit of the process, but it is possible to reduce the size of each unit in which the process is small. Very inefficient to break into.

An object of the present invention is to provide a character recognition method which more accurately determines the certainty factor of a recognition result and gives it to the user.

Another object of the present invention is to reduce the time waste caused by continuing the processing when it is clear that the desired recognition result cannot be obtained, and to prompt the user to take necessary measures. It is to provide a character recognition method that can be adopted in.

Another object of the present invention is character recognition which can improve processing efficiency and recognition rate by eliminating inappropriate candidates at the earliest possible point in the processing process or changing processing methods and parameters. To provide a method.

Another object of the present invention is to provide a character recognition method that allows a user to easily determine the cause of erroneous recognition and necessary treatment.

Still another object of the present invention is that a large storage space is not required for storing a character pattern or feature amount for dictionary learning even when the unit of recognition processing is large, such as the entire surface of an original, and It is to provide a character recognition method for efficiently learning a dictionary.

[0013]

According to a first aspect of the present invention, in order to recognize a character, a plurality of processes such as pattern matching process, character combination selection process, rule process, language process, etc. are performed and obtained by each process. It is characterized in that the certainty factor of the recognition result is comprehensively determined based on the obtained information.

According to a second aspect of the present invention, in order to identify a character, a plurality of processes such as pattern matching process, character combination selection process, rule process, and language process are sequentially performed, and the process is performed for each process of each stage. The confidence level of the candidate is obtained based on the information obtained in step 1, and the operation of updating the confidence level obtained up to the process in the previous stage based on the obtained confidence level is repeated until the final process, so that the final recognition result Characterized by determining the final certainty factor.

The invention of claim 3 is characterized in that the information of the determined certainty factor is added to the recognition result and displayed or printed.

According to a fourth or fifth aspect of the present invention, a visual condition such as color, brightness, decoration, or typeface of the recognition result displayed or printed is changed according to the certainty factor, or a character or a symbol indicating the certainty factor. Is displayed or printed in association with the recognition result.

According to a sixth aspect of the present invention, in the first or second aspect of the present invention, the number of recognition result characters having a certainty factor of a certain level or less is counted, and when the count value exceeds a predetermined threshold value, a warning is issued. Alternatively, the processing is interrupted.

The invention of claim 7 is characterized in that, in the invention of claim 2, a candidate whose certainty factor becomes a certain level or less in a certain process is excluded from candidates of the subsequent process.

The invention of claim 8 is characterized in that, in the invention of claim 1 or 2, information of the certainty factor for each process is stored separately from the final certainty factor.

A ninth aspect of the present invention is characterized in that a message is output based on the information of the certainty factor for each processing stored for the recognition result character.

The invention of claim 10 is the invention of claim 1 or 2.
In the invention, the certainty factor up to a certain process is reflected in the process before or after the process, or the process before and after the process.

According to the invention of claim 11, in the invention of claim 1 or 2, the pattern or the characteristic amount of the recognition target character is stored only when the certainty factor determined for the recognition target character is a predetermined level. Is characterized by.

According to the twelfth aspect of the present invention, when the recognition result is modified, the pattern is stored using the stored corresponding pattern or feature amount only when the certainty factor for the modified recognition result is a predetermined level. It is characterized in that dictionary learning for matching processing is executed.

[0024]

Confidence indicates the degree of certainty of a character as a recognition result and is represented by a numerical value of 0% to 100%, or the numerical value is quantized in several steps. Let For example, it is represented by the following three ranks A, B, and C. Rank A: The recognition result is correct. B rank: The recognition result is suspicious (I do not know whether it is correct or incorrect). C rank: The recognition result is incorrect.

According to the present invention, the certainty factor of the final recognition result is comprehensively determined on the basis of the information obtained by a plurality of stages of processing such as pattern matching processing, character combination selection processing, rule processing and language processing. For example, from the pattern matching process, the evaluation value of the first candidate or the difference between the evaluation values of the first candidate and the second candidate, from the character combination selection process, the evaluation value when determining multiple characters or a single character, from the rule process, Information indicating whether or not a rule has been applied and corrected, and information indicating the result of language correction from language processing, and using this information as evidence, for example, Dempster
The reliability is comprehensively judged by using the probability theory of Shafer (Dempster & Shafer).

The determination of such a certainty factor is carried out in the final processing stage by collecting the information obtained in the processing stages up to that time, or by determining the candidates based on the information obtained in each processing stage. The certainty factor is obtained, and the operation of updating the certainty factor obtained up to the preprocessing stage is repeated until the final processing stage.

The certainty factor thus determined can be more accurate than the case where only the degree of similarity with the pattern dictionary is used, and the risk of a completely inappropriate certainty factor is greatly reduced. be able to.

By changing the visual condition such as color and brightness in displaying or printing the recognition result or displaying or printing the character or the symbol corresponding to the recognition result in accordance with the certainty factor. The user can easily recognize the certainty factor of the recognition result, quickly and accurately find the character that needs correction, and efficiently perform the correction work.

Furthermore, when the certainty factor is less than a certain level, for example, when the certainty factors of the B rank and the C rank exceed a predetermined threshold value, a warning (for example, a warning sound) is generated, so that the device user can It can be determined that there are many mistakes in the document currently being read, and the necessary processing such as stopping the processing or changing the recognition condition and activating can be taken quickly. Further, in such a case, it is often wasteful to continue the processing, but by automatically suspending the processing, such time waste can be reduced.

Further, based on the certainty factor in a certain processing stage up to the final processing stage, by removing candidates having a low possibility of correct answer from candidates at that stage, the processing amount of the subsequent processing stage is reduced accordingly. The processing efficiency can be improved.

Further, by storing the information of the certainty factor for each processing step separately from the final certainty factor, the information can be utilized for the maintenance of the character recognition device, and erroneous recognition is performed. When correcting a character, a message is output to advise the user of the cause of misrecognition, necessary measures, etc., based on the information of the certainty factor for each related processing step, and the correction work by the device user or Device maintenance can be facilitated.

Further, the certainty factor up to a certain process is reflected in the process before or after the process or before and after the process, and the method and parameters of the process before or after are optimized to improve the recognition rate, In addition, the efficiency of correction work can be improved.

Further, the correction of the recognition result is carried out mainly on the character having a low certainty factor, and the necessity of dictionary learning is high for the character having the low certainty factor. According to the present invention, since the pattern or the characteristic amount is stored only for the character whose certainty factor is lower than a predetermined level and the necessity of dictionary learning is high, the information of the necessary pattern or the characteristic amount is stored using a limited storage space. it can. Further, at the time of correction, only a character having a low degree of certainty and a high degree of necessity for dictionary learning can be efficiently used for dictionary learning by using the saved pattern or feature amount.

[0034]

Embodiments of the present invention will be described in detail below with reference to the drawings.

FIG. 1 is a schematic block diagram of a character recognition apparatus according to the first embodiment of the present invention. In FIG. 1, a scanner 1 reads an image of a document and inputs it as binary data, and a cut-out processing section is a part for cutting out a character image from the input image. The normalization processing unit 3 is a unit that performs normalization processing for correcting the distortion of the cut out character image. The pattern matching processing unit 4 extracts the characteristic amount of the normalized character image and compares it with a pattern dictionary to calculate the similarity or dissimilarity, and selects one or more recognition candidates with high similarity or low dissimilarity. It is a processing unit that determines. For example, in the case of Japanese, the character combination selection processing unit 5 is a processing unit that selects an optimum character combination such as whether the full-width character is one character or the half-width character is two characters. The rule processing unit 6 includes information on the relative position of characters in a line (related to the distinction between “.” And “.”, Etc.) and character size information (for normalization, such as C and c). According to the rules such as character case information (for example, if there is a kanji number "1" in the katakana character string, this is corrected to the katakana long syllabary "-"), pattern matching is performed. This is a part for correcting the recognition result candidates. The language processing unit 8 is a unit that corrects an error by morphologically analyzing the recognition result character string after correction by rule processing and matching it with a word dictionary. The result of this language processing is the final recognition result, which is subject to correction by the user.

The certainty factor determination unit 9 evaluates the first candidate from the pattern matching processing unit 4 (for example, the value obtained by dividing the distance from the dictionary by the total number of contours) or the evaluation value of the first candidate and the second candidate. , The evaluation value for determining the character combination from the character combination selection processing unit 5, the information of the rule applied for the correction from the rule processing unit 6, and the information representing the result of the language correction from the language processing unit 7, This is a part of comprehensively determining the certainty factor of the final recognition result using the Dempster-Schaefer probability theory using these information as evidence. Specifically, the certainty factor is quantized into A, B, and C ranks as described later.

The output control section 8 is a section for controlling the display of the final recognition result on the display 11 or the printing by the printer 12, and adds the information of the certainty factor given by the certainty factor determining section 9 to the recognition result. Then display 11
Alternatively, the visual condition of the recognition result to be displayed or printed is changed according to the certainty factor in order to be output from the printing device 12.

Specifically, a character having a certainty factor of A rank is black, and a character of B and C ranks has a black background with a white background, the brightness of the character is changed according to the certainty factor, and the background color is black. Change the color of the character or background according to the certainty, such as the A rank character is white, the B rank character is yellow (color that calls attention), and the C rank character is red (color that indicates a warning). Let

Alternatively, the background color is black, the A rank character is white, the B rank character is single underlined or lightly shaded, and the C rank character is double underlined or dark shaded. like,
Depending on the certainty, the decoration of the letters is changed, the background color is black, the A rank letters are Mincho, the B rank letters are Gothic, and the C rank letters are italic. Change the typeface of the letters.

Alternatively, the background color of the image corresponding to the character is white, the image corresponding to the A rank result is black, the image corresponding to the B rank result is yellow, and the image corresponding to the C rank result is red. Then, the color of the image corresponding to the character of the recognition result is changed according to the certainty factor.

Alternatively, a character or a symbol indicating the certainty factor is output in association with the recognition result, such that B is displayed below the character of rank B and C is displayed below the character of rank C to be displayed or printed.

The device control unit 10 controls the entire device, and the recognition result character having a certainty factor of a certain level or less, for example, B.
The number of rank and C rank characters is counted, and when the count value exceeds a predetermined threshold value, the process is interrupted and the device user is warned by sounding a buzzer attached to the display 11.

Next, the processing of the certainty factor determining unit 9 will be described in detail. FIG. 2 shows the flow of processing in the certainty factor determination unit 9.

As described above, the certainty factor determination unit 9 uses the information gathered from the process stages of the pattern matching process, the character combination selection process, the rule process, and the language process as evidence and uses the Dempster-Schaefer probability theory to synthesize the information. The degree of certainty. FIG. 3 shows an outline of the processing.

The Dempster-Schaefer probability theory has a feature that it can represent ignorance and collect and synthesize fragmentary information (independent evidence).

[0046]

[Table 1] As described above, the probability is divided into areas of “credibility”, “ignorance”, and “untrust”, and “credibility” is called a lower bound probability, and “credit + ignorance” is called an upper bound probability.

Specifically, first, a hypothesis H1 that this character is correct and a hypothesis H2 that this character is incorrect are prepared, and a pseudo-possibility distribution (for each hypothesis, the hypothesis assigned to each hypothesis Distribution of the probability that is true)
Ask for. In the present embodiment, as shown in FIG. 2, the pseudo-possibility distribution regarding the evaluation value of the first candidate, which is the evidence collected by the pattern matching process, and the pseudo-potential distribution regarding the difference between the evaluation values of the first candidate and the second candidate. Possibility distribution, Pseudo-possibility distribution regarding the difference in evaluation value at the time of path selection, which is evidence collected by the path selection process, Pseudo-possibility distribution regarding evidence content collected from the rule process, Pseudo-possibility related to correction contents collected from the language process Generate a sex distribution.

When the evidence is numerical information, this pseudo-possibility distribution is obtained by a linear equation such as y = -x / 3 + 100 (y is pseudo-possibility distribution, x is evidence of numerical information). If the evaluation value is 100, the probability 67% of the hypothesis H1 and the probability 33% of the hypothesis H2 are calculated by this formula.

If the evidence is not numerical information, it is given as a heuristic. For example, in the case of evidence that a certain rule is not satisfied, the probability of the hypothesis H1 is 20% and the probability of the hypothesis H2 is 80%.

From the pseudo-possibility distribution for each piece of evidence thus collected, the basic probability assignment is next obtained. That is, the pseudo possibility distribution is assigned to each probability of “correct”, “ignorance”, and “error”.

For example, the probability of hypothesis H1 is 67%, the hypothesis H is
If the probability of 2 is 33%,

[0052]

[Table 2] Like

For example, the probability of hypothesis H1 is 20%,
If the probability of hypothesis H2 is 80%,

[0054]

[Table 3] Like

Next, the basic probabilities are synthesized by the Dempster integration rule. This is a way to integrate the basic probabilities inferred from independent evidence,

[Equation 1] However, m1 and m2 are always expressed by the true hypothesis that the basic probability H is H1 or H2, and H1H2 H1H2 is H1 or H2.

Next, the lower bound probability and the upper bound probability are obtained. Take the following jury model as an example.

[0057]

[Table 4] In this example, the guilty lower bound probability is 15% and the upper bound probability is 74.
% (= 15 + 59) Innocent lower bound probability is 26%, Upper bound probability is 85% (= 26 + 59
) Becomes.

Finally, the confidence level is determined from the lower bound probability and the upper bound probability. Although various methods are conceivable, as the simplest method, for example, the ranks of A, B, and C are determined from the upper bound probability of the hypothesis H2 by the following threshold values. The upper bound probability of H2 is 0 to 19 correct (A rank) The upper bound probability of H2 is 20 to 79 Suspicious (B rank) The upper bound probability of H2 is 80 to 100 False (C rank)

The rank of the certainty factor may be determined in two steps of the lower bound probability and the upper bound probability.

FIG. 4 shows a specific example of the recognition result, and FIG. 5 shows a specific process for determining the certainty factor for it. That is, in the case of this example, the certainty factor of the final recognition result "ru" is ranked as correct.

Further, FIG. 6 shows an example in which the recognition result to which the certainty factor is added is displayed on the display 11, and FIG. 7 shows an example in which the printer 12 prints on the paper. Further, FIG. 8 shows an example in which the certainty factor is displayed in association with the recognition result, and FIG. 9 shows an example in which the certainty factor is also printed on paper.

FIG. 10 is a schematic block diagram of a character recognition apparatus according to the second embodiment of the present invention.

In FIG. 10, a scanner 1, a cutout processing unit 2, a normalization processing unit 3, a pattern matching processing unit 4, a character combination selection processing unit 5, a rule processing unit 6, a language processing unit 7, an output control unit 8 and a display. 11, printing device 1
2 is the same part as the corresponding part shown in FIG.

The certainty factor determining unit 13 is a unit for obtaining the certainty factor using the probability theory of Dempster-Schafer, but there is a slight difference from the certainty factor determining unit 9 shown in FIG. That is, in the certainty factor determination unit 13, the evaluation value of the first candidate from the pattern matching processing unit 4 (for example, the value obtained by dividing the distance from the dictionary by the total number of contours) or the evaluation value of the first candidate and the second candidate. Difference information, evaluation value information when the character combination is selected from the character combination selection processing unit 5, rule information applied for correction from the rule processing unit 6, and language processing result from the language processing unit 7. Information such as character pitch and character rectangle width is collected from the cutout processing unit 2, and the certainty factor about the candidate at that processing stage is determined by using the information collected at each processing stage as evidence.
The certainty factor up to the preprocessing stage is updated by the determined certainty factor. By repeating this operation up to the final processing stage, that is, the language processing stage, the final certainty factor for the final recognition result is determined, and this is determined by A, B, and
Quantize into 3 ranks of C.

The number-of-characters counting unit 14 counts the number of recognition-result characters whose final certainty factor determined by the certainty factor determining unit 13 is a certain level or less, for example, B rank and C rank characters, and the count value is predetermined. A character number excess signal is output when the threshold value is exceeded. FIG. 11 shows the processing flow.

The candidate deletion determination unit 15 compares the certainty factor of each candidate obtained by the certainty factor determination unit 13 with a certain threshold value based on the information obtained in the cut-out processing stage, and selects the candidate with the certainty factor lower than that. Judge as a candidate for deletion and output a deletion signal.
Here, the cutout process is selected as the target process stage, but another process stage may be selected as the target, or two or more process stages may be selected as the target.

The device control unit 10 controls the entire device, interrupts the process when a character number excess signal is output from the character counting unit 14, and sounds a buzzer or the like attached to the display 11 to allow the user of the device to operate. When the deletion signal is output from the candidate deletion determination unit 15, control is performed to remove the corresponding candidate from the candidates for the next processing stage.

Next, the processing of the certainty factor determination unit 13 will be described. FIG. 12 shows the relationship between the processing of each processing unit and the processing of the certainty factor determination unit 13, and FIG. 13 shows the flow of the certainty factor determination process at each processing stage in the certainty factor determination unit 13.

As described above, the certainty factor is determined by using the information obtained from the processing stage as evidence and using the probability theory of Dempster-Schaefer. More specifically, H
1: first candidate, H2: second candidate, ..., Hn: nth candidate,
As described above, the hypothesis H for the predetermined maximum number of candidates is
make i. That is, the probability of each character is obtained by considering whether the final recognition result character is the first candidate or the second candidate.

First, at each processing stage, a pseudo-possibility distribution (a distribution of the possibility that the hypothesis is true assigned to each hypothesis) is obtained for each piece of information gathered there, that is, each piece of evidence. At this time, if the evidence is numerical information, y
As described above, the pseudo possibility (y) is obtained by a linear equation such as = -x / 3 + 100, and the pseudo possibility distribution is heuristically provided if the evidence is not numerical information.

Next, using the pseudo-possibility distribution of the hypotheses H1 to Hn, the sequences of hypotheses H1 and H called consonance sequences.
2 ,. ． ． , Hn, and convert to basic probability allocation. That is, the pseudo possibility distribution is “correct” and “ignorance”.
Allocate to each basic probability of "mistake". The basic probability is the certainty factor, but the certainty factor may be determined from the lower bound probability and the upper bound probability obtained from it.

If there is a plurality of evidences, the basic probability of each hypothesis obtained for each evidence is synthesized by the synthesis formula (Equation 1) using the Dempster integration rule as described above.

When the basic probability (certainty factor) of each hypothesis based on the evidence obtained at each processing stage is obtained in this way,
This is combined (updated) by the basic probability of each hypothesis obtained up to the immediately preceding processing stage and the above-mentioned composition formula (Equation 1). Naturally, the lower bound probability and upper bound probability of each hypothesis can also be obtained.

At the final stage, that is, language processing, the determined certainty factors are ranked. Although various ranking methods can be considered, the simplest method is a method by comparing the upper bound probability with a threshold value as described above.

The above-described certainty factor determination process will be described more specifically according to the flow of the character recognition process.

First, it is assumed that the cutting process results in the cutting result as shown in FIG. The rectangles in the figure represent character rectangles that are character candidates. And A, B, C
When paying attention to the rectangle of, the rectangle A (full-width character) and the rectangles B and C (half-width character) are candidates for the character, and the probability (confidence) described above is obtained from the information (evidence) such as the pitch between characters and the width of the rectangle for each. Degree) is an A rectangle lower bound probability? % Upper bound probability? % B rectangle Lower bound probability? % Upper bound probability? % C rectangle Lower bound probability? % Upper bound probability? % Is required.

In the present embodiment, for example, a rectangle whose lower bound probability or upper bound probability falls below a predetermined threshold is excluded from character candidates in the next processing stage.

It is also possible to determine only one rectangle that is most likely to be the candidate from the certainty factor at this stage.

For each of these rectangles A, B and C (of course, those excluded from the character candidates are not included),
A recognition result candidate group is obtained by the next pattern matching processing (recognition processing), but the evaluation value of the first candidate for each rectangle,
Based on information such as the difference in evaluation value between the first candidate and the second candidate, A rectangle first candidate lower bound probability? % Upper bound probability? % 2nd candidate lower bound probability? % Upper bound probability? % 3rd candidate lower bound probability? % Upper bound probability? % B rectangle First candidate lower bound probability? % Upper bound probability? % 2nd candidate lower bound probability? % Upper bound probability? % 3rd candidate lower bound probability? % Upper bound probability? % C rectangle First candidate lower bound probability? % Upper bound probability? % 2nd candidate lower bound probability? % Upper bound probability? % 3rd candidate lower bound probability? % Upper bound probability? Ask for%. As described above, the certainty factor is updated by combining the certainty factor with the certainty factor obtained in the cutting process in the previous stage.

At this stage, it is possible to reduce the processing time by excluding, from the candidates, the certainty factor in each of the rectangle candidates, which is equal to or less than a predetermined threshold value, as in the case of the clipping process. Furthermore, at this stage, if it is A rectangle,
It is also possible to determine whether the rectangle is a B or C rectangle from the combined certainty factors. Here, the description will proceed assuming that the A rectangle is selected as a character candidate.

In the next rule processing, the certainty factor of each recognition result candidate is obtained from the information such as the character position, the character size, the character type, etc., and is combined with the certainty factor at the pattern matching processing stage to obtain the updated certainty factor. 1 candidate lower bound probability? % Upper bound probability? % 2nd candidate lower bound probability? % Upper bound probability? % 3rd candidate lower bound probability? % Upper bound probability? Ask for%. For example, regarding the position information of characters, if the target image is located near the center of the line, the certainty factor of "." Increases, but the certainty factor of ".""'" Decreases.

In the next language processing, which candidate is the most suitable is determined using a word dictionary or the like, a certainty factor is obtained for the result, and the result is combined with the certainty factor in the preprocessing stage. For example, according to matching with a word dictionary, when a meaningful word is formed by combining the first candidate of the target character and the second candidate of the next character, the first candidate of the target character and the second candidate of the next character are formed.
The confidence level of the candidate increases. Therefore, in order to perform language processing, a character string having a certain length is required.

Thus, for example, the first candidate lower bound probability 80% the upper bound probability 15% the second candidate lower bound probability 33% the upper bound probability 10% the third candidate lower bound probability 12% the upper bound probability 33% Since a certainty factor is obtained, for example, the first candidate with the maximum lower bound probability is output as the final recognition result,
The certainty factor is ranked by the threshold processing as described above. The candidates may be sorted in descending order of certainty at each processing stage, or may be sorted in descending order of final certainty.

Further, the lower bound probability and the upper bound probability in the case of a plurality of pieces of evidence need only be obtained by adding the corresponding basic probabilities as described in the example of the jury model using Table 4 above. .. In addition, since it is the case of character recognition processing, "guilty", "don't know", and "innocence" in the above description of the jury model can be paraphrased as "correct", "don't know whether it is right or wrong", and "wrong". Good.

15 to 17 show specific processing examples of the second embodiment. In this example, there are first to third candidate characters, and the third candidate character is cut off at the stage of the character combination selection processing. That is, as shown in FIG. 16, m (H ₂ ), m (H ₃ ) are both 0 as a result of the combination processing of the basic probabilities up to the character combination selection processing, but H ₂
Is 5 (m (H ₁ H ₂ ) + m (H ₁ H ₂ H ₃ ))
7%, whereas H ₃ contains 5% of m (H ₁ H ₂ H ₃ ).
Since it is (0% after that), the third candidate cuts off. As a result, in the subsequent processing, the first candidate and the second candidate are the processing targets, and the time for the confidence factor processing is saved. That is,
Since there are two hypotheses from the rule processing, m (H ₃ ),
m (H ₂ H ₃ ), m (H ₁ H ₃ ), m (H ₁ H ₂ H ₃ ) are no longer calculated.

Next, a modification of the first and second embodiments will be described. In this modified example, as in the above-described embodiment, the certainty factor determination units 9 and 13 obtain and combine certainty factors for each processing step of cutout processing, pattern matching processing, character combination selection processing, rule processing, and language processing. The final confidence level is determined by determining the confidence level for each processing step, that is, the confidence level (numerical value or rank) determined based on the information obtained at each processing step, in addition to the final confidence level. ) In memory. For the recognition result character on the screen of the display 11, the device control unit 10 reads information on the certainty factor for each processing step regarding the character from the memory, and creates a message (advice) to the user based on the information. And pass it to the output control unit 8 display 1
Display it on the 1st screen. FIG. 18 shows a processing flow chart thereof.

For example, if it is judged from the certainty factor information that the certainty of the cutout process is more suspicious than the other processing steps, it is possible that the character could not be cut out correctly. If the certainty of the pattern matching process (recognition process) is more suspicious than the other processing stages, the message such as "Check the recognized character image." If the certainty of processing (language processing) is more doubtful than other processing steps, "Check by character string" or "Knowledge (rule), check word dictionary". Message is displayed.

By displaying such a message, the user can not only correct the erroneously recognized character, but also easily determine the cause and take necessary measures.

Further, in the first and second embodiments, the optimum method is selected from a plurality of processing methods, parameters, etc. of the processing steps before or after the processing step or before and after the processing step according to the certainty factor up to a certain processing step. You can also choose one
It is effective in improving the recognition rate and improving the efficiency of the final correction work. FIG. 19 shows a processing flow chart thereof.

For example, in the second embodiment, when the certainty factor up to the pattern matching process is low, the device control section 10 changes the cutout position in the determination character cutout process and recognizes again, or the pattern matching process. Each part is controlled so that the feature extraction method, which is the pre-processing in the above, is changed and recognition is performed again. This is because the former is considered to have low confidence in the recognition result because the character cutout position is inappropriate, and the latter is considered to be because the recognized image was not suitable for the previous feature extraction method. In this way, when the information on the certainty factor is fed back to the previous process, the device control unit 10 selects a result having a higher certainty factor than a plurality of recognition results tried by changing the processing method or the parameter as an effective one.

On the contrary, when the certainty factor up to the pattern matching process is high, the device control unit 112 controls so as not to change the result in the subsequent word matching in the language process.
In other words, word matching generally matches a combination of recognition result candidates with a word dictionary prepared in advance, but at that time, the character with high certainty is fixed, and only the candidate for the character with low certainty is fixed. And replace it with the word dictionary. By doing so, the number of word combinations is reduced, which leads to a reduction in processing time and an improvement in recognition rate. In this way, when utilizing the information on the certainty factor in the subsequent processing, it becomes possible to select the parameters, conditions, processing methods, etc. in the subsequent processing.

FIG. 20 is a schematic block diagram showing the third embodiment of the present invention. Reference numeral 200 denotes a part in which a scanner is added to the same structure as the character recognition device main body shown in FIG. 1 or FIG.

Reference numeral 202 denotes a storage memory for storing the feature amount of the character pattern extracted by the pattern matching processing unit in the apparatus main body 200 or the character pattern after being normalized by the normalization processing unit in the apparatus main body 200. However,
Here, it is assumed that the characteristic amount of the character pattern is stored in the storage memory 202. Reference numeral 204 denotes the apparatus main body 200
The confidence level information is received from the confidence level determination unit in
It is a storage control unit that controls the storage memory 202 to store the feature amount of the character pattern corresponding to the rank and the C rank. FIG. 21 shows the flow of processing for saving the characteristic amount (or pattern) by the saving control unit 204.

Although not shown in FIG. 1 or FIG. 10, 206 is a recognition result memory which is naturally provided in the first embodiment or the second embodiment. Here, the recognition result data and the confidence information are stored. Remembered. Reference numeral 11 is a keyboard for the user to input information, and the correct character when the recognition result is corrected is input from the keyboard. A correction processing unit 208 corrects the recognition result data in the recognition result memory 206 according to the input from the keyboard 11.

Reference numeral 212 is a dictionary memory that stores a dictionary used for recognition processing in the pattern matching processing section in the apparatus main body 200, which is naturally also present in the first embodiment or the second embodiment. is there. 214 is
A learning processing unit for learning this dictionary. This learning processing unit 214, when the recognition result is corrected,
When the confidence character rank information of the corrected character and the correct character code are passed from the correction processing unit 208, when the confidence rank is B rank or C rank, the characteristic amount of the corrected character is read from the storage memory 202 and used. And executes the dictionary learning process. FIG. 22 shows the processing flow, and FIG. 23 shows a specific example of the processing. FIG. 23 shows an example in which "To" is corrected to "Bright", which shows that the character pattern of "Bright" is displayed on the screen of the display and the dictionary is learned using the feature amount.

When the character pattern is stored in the storage memory 202, the learning processing unit 214 uses
Read the pattern of modified characters from the save memory 202,
The feature amount is extracted inside the learning processing unit, or the pattern matching processing unit in the apparatus main body 200 is caused to use the obtained feature amount for dictionary learning.

[0094]

As described in detail above, according to the present invention, it is possible to more accurately determine the certainty factor of the recognition result, add the information to the recognition result, and notify the user of the device. The user can easily recognize the certainty factor of the recognition result and can efficiently perform the necessary correction work.If the certainty factor of the recognition result is low and the continuation of the process is wasteful, the user can process it. It becomes possible to quickly take the necessary action such as stopping the process or starting it after changing the recognition condition, and in such a case, it is possible to reduce the waste of time due to the continuation of useless processing. ,
Furthermore, based on information on the certainty factor for each process, a message such as the cause of misrecognition and necessary measures is output, facilitating correction work and device maintenance work by the device user. The recognition rate can be improved or the correction work can be made more efficient by automatically switching the method or parameter of the previous or subsequent processing or the processing before or after the processing depending on the degree. The dictionary learning can be efficiently performed, and the storage space for storing the character pattern or the feature amount necessary for the dictionary learning can be reduced.

[Brief description of drawings]

FIG. 1 is a schematic block diagram according to a first embodiment of the present invention.

FIG. 2 is a flowchart of a certainty factor determination process in the first embodiment.

FIG. 3 is a conceptual diagram of confidence factor determination processing in the first embodiment.

FIG. 4 is a specific example of recognition result characters.

FIG. 5 is a diagram showing a specific example of the certainty factor determination processing in the first embodiment in the example of FIG. 4;

FIG. 6 is an example in which a recognition result with a certainty factor is displayed.

FIG. 7 is an example in which a recognition result with a certainty factor is printed.

FIG. 8 is an example in which a certainty factor is displayed in association with a recognition result.

FIG. 9 is an example in which the certainty factor is printed in association with the recognition result.

FIG. 10 is a schematic block diagram according to a second embodiment of the present invention.

FIG. 11 is a flow chart when a warning is issued and the process is interrupted when the certainty factor of the recognition result is low.

FIG. 12 is an overall flow chart of the second embodiment.

FIG. 13 is a flowchart of a certainty factor determination process at each processing stage in the second embodiment.

FIG. 14 is an explanatory diagram related to character cutting processing.

FIG. 15 is a specific example of a certainty factor determination process in the second embodiment.

FIG. 16 is a diagram following FIG. 15;

FIG. 17 is a diagram subsequent to FIG. 16;

FIG. 18 is a flow chart when a message is output according to the certainty factor for each processing stage.

FIG. 19 is a flowchart when the certainty factor up to a certain process is reflected in the previous or subsequent process.

FIG. 20 is a schematic block diagram according to a third embodiment of the present invention.

FIG. 21 is a flowchart of a storage process of a feature amount or a pattern.

FIG. 22 is a flowchart of dictionary learning processing.

FIG. 23 is a specific processing example of dictionary learning.

[Explanation of symbols]

1 Scanner 2 Cutout Processing Part 3 Normalization Processing Part 4 Pattern Matching Processing Part 5 Character Combination Selection Processing Part 6 Rule Processing Part 7 Language Processing Part 8 Output Control Part 9 Confidence Determining Part 10 Device Control Part 11 Display 12 Printing Device 13 Confidence Degree determining unit 14 Character number counting unit 15 Candidate deletion determining unit 200 Device body 202 Storage memory 204 Storage control unit 206 Recognition result memory 208 Correction processing unit 212 Dictionary memory 214 Learning processing unit

Claims (12)

[Claims] 1. In order to recognize a character, a plurality of processes such as pattern matching process, character combination selection process, rule process, and language process are performed, and a comprehensive recognition result is obtained based on the information obtained by each process. A character recognition method characterized by determining a certainty factor. 2. A plurality of processes such as a pattern matching process, a character combination selecting process, a rule process, and a language process are performed in order to identify a character, and the information obtained by the process is performed for each process. The final confidence for the final recognition result is obtained by repeating the operation of updating the confidence level of the candidate based on A character recognition method characterized by determining the degree. 3. The character recognition method according to claim 1, wherein information on the determined certainty factor is added to the recognition result and displayed or printed. 4. The character recognition method according to claim 3, wherein visual conditions such as color, brightness, decoration, and typeface of the recognition result displayed or printed are changed according to the certainty factor. 5. The character recognition method according to claim 3, wherein the character or the symbol indicating the certainty factor is displayed or printed in association with the recognition result. 6. The method according to claim 1, wherein the number of recognition result characters having a certainty factor equal to or lower than a certain level is counted, and when the count value exceeds a predetermined threshold value, a warning is issued or the process is interrupted. Or the character recognition method described in 2. 7. The character recognition method according to claim 2, wherein a candidate whose certainty factor is lower than a certain level in a certain process is excluded from candidates for a subsequent process. 8. The character recognition method according to claim 1 or 2, wherein information on the certainty factor for each process is stored separately from the final certainty factor. 9. The character recognition method according to claim 8, wherein the message is output based on the information of the certainty factor for each processing stored for the recognition result character. 10. The character recognition method according to claim 1, wherein the certainty factor up to a certain process is reflected in the process before or after the process, or the process before and after the process. 11. The character recognition method according to claim 1, wherein the pattern or the characteristic amount of the recognition target character is stored only when the certainty factor determined for the recognition target character is a predetermined level. .. 12. A dictionary for pattern matching processing using a stored corresponding pattern or feature amount only when the certainty factor for the corrected recognition result is a predetermined level when the recognition result is corrected. 12. The character recognition method according to claim 11, wherein learning is performed.