JP2003346080A - Character recognition method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a character recognition method making it possible to enhance the overall processing capability of a character recognition system which requires human verification, by making use of the results of character recognition for input images and their certainty. <P>SOLUTION: The method includes performing character recognition for an image on an input form; obtaining similarity as the result of this character recognition; comparing the similarity to the previously registered certainty required of the character recognition; and either providing an output indicating that no human verification is required for the result of the character recognition according to the comparison result, or providing an output urging human verification by presenting candidates for character recognition as options for the result of character recognition according to the comparison result, or providing an output urging manual inputting by presenting a manual new input and a determination for the result of the character recognition according to the comparison result. <P>COPYRIGHT: (C)2004,JPO

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 in a character recognition system that requires manual verification based on a character recognition result from an input image obtained by optically scanning a form.

[0002]

2. Description of the Related Art The accuracy of character recognition is improving year by year, regardless of whether a target is a handwritten character. Part 1
One of the factors is the enrichment of the training data set, and more specifically, the theoretical development of analyzing this data set has made a significant contribution. According to this, in order to improve the discrimination performance with other categories, in a vector set for each category distributed in a vector space, a vector called a support vector, which is present around a boundary of a pattern set, is appropriately selected. Contribute to the improvement of the identification performance. In recent years, this theoretical finding has been generally accepted by being confirmed by computer experiments and the like.

[0003] With such an improvement in the discrimination performance, a certain character image is newly changed to "extremely category" A "" from conventional output information such as "category" A "" or "neither". It has also become possible to provide detailed information such as "A near category" B "" or "neither." The prior art basically considers only the association with the correct category, and is ambiguous in handling information indicating the relationship with other categories and certainty. This is due to insufficient analysis of the data set.

[0004] It is considered that "very likely to be category" A "" has a high appearance frequency and is located near the center of the distribution of the pattern set. Regarding the recognition result of such a pattern, the character recognition world (the world of the high-dimensional vector space) is synchronized with the real world, and is extremely stable. In the future, as the character recognition technology evolves, the number of sets that can obtain such a stable solution is expected to increase more and more. Although character recognition technology is widely used,
For the purpose of minimizing the error rate of character recognition, it is very effective to add such new information to a character recognition system that requires manual verification and reconstruct the system configuration.

[0005]

An advantage of using a character recognition system is that a document image can be identified in a large number at high speed without errors and with high accuracy. For such a character recognition system to be effective, it must be guaranteed that it is superior to manual processing. Here, in order to compare the human and character recognition systems in terms of processing capacity,
The processing capacity is evaluated by the following basic formula.

Ability = speed (average) × precision (average) Although character recognition is overwhelmingly superior in processing speed, accuracy always matters. Normally, 9 characters per individual character
Accuracy of 9.0% or more is required. When operating a system that requires high accuracy (minimum error rate), when the average accuracy of character recognition is not reached, and when handling a large amount of character recognition strings such as word recognition, manual verification is required. I need.

In the present invention, on the premise of such a system that requires manual verification, a method for designing a character recognition system in which the character recognition accuracy can be adjusted is represented by the above-mentioned basic formula, where the capability = α (speed 1 x accuracy) + β (speed 2 x accuracy) + γ
(Speed 3 × Accuracy) (α + β + γ = 1, Speed 1> Speed 2>
> Speed 3) Here, the first, second, and third terms on the right-hand side are referred to as “character recognition processing capability”, “processing capability only for checking character recognition output”, and “complete manual processing capability”, respectively ( The same applies to speed 1, speed 2, and speed 3). If the character recognition accuracy can be adjusted, it is possible to improve the processing capability to the maximum while minimizing errors by appropriately distributing according to the conditions.

This concept is different from the conventional verification based on the result of the “accept-reject” determination. For example, by adjusting the recognition accuracy, for a certain character image, "Absolute" A "", "A" or "B"
I think that it is "C" or "I don't know at all". Although subdividing the output result leaves the final selection by hand, there is a difference in processing speed and accuracy at the level of judgment. It can be easily imagined that the processing capacity of the whole system is improved by distributing these optimally.

As the character recognition method becomes more accurate,
The analysis of the distribution of the pattern set in the feature vector space proceeds similarly. This fact means that as the detailed analysis of the subset that has been loved until now progresses, the number of subsets that can be analyzed with certainty increases as a result. If confidence is introduced for the purpose of measuring subsets with such properties, their verification work is "determined" while referring to the similarity of character recognition results and the correspondence with candidate character strings. , "Choice questions" and "free writing questions". Specifically, "No key input",
This corresponds to the classification into three typical tasks of “arrow & return key” and “normal keyboard input”. Thereby, both the processing speed and the processing accuracy per one image are improved.

Therefore, the present invention provides a character recognition system which requires a manual verification to utilize a character recognition result of an input image and its certainty factor to improve the overall processing performance. It is intended to provide a recognition method.

[0011]

According to the character recognition method of the present invention, a first method for inputting an image on a form on which characters are described is provided.
A second step of performing character recognition on the input image input in the first step to obtain a similarity as a result of the character recognition, and a similarity degree obtained in the second step. A third step of comparing a certainty factor required for the character recognition registered in advance, and an output which does not require a manual verification process on the character recognition result based on a result of the comparison in the third step. An output for presenting a character recognition candidate option to the character recognition result to prompt a manual verification process, or an output for prompting a manual input process by presenting a new manual input and confirmation to the character recognition result. And performing a fourth step.

Further, the character recognition method of the present invention includes a first step of inputting an image on a form in which characters are described in a predetermined form, and the form using the input image input in the first step. Second identifying form above
And a third step of determining whether or not the form of the form identified in the second step is registered in advance. As a result of the determination in the third step, the form of the form is If the form has not been registered, the fourth step of registering the form and the degree of certainty required for character recognition of the form, and the result of the determination in the third step indicates that the form of the form has been registered in advance. A fifth step of performing character recognition on the input image input in the first step on the basis of the registered form and obtaining a similarity as a result of the character recognition; A sixth step of comparing the obtained similarity with the previously registered certainty factor; and An output that does not require manual verification processing on the character recognition result based on the result of the comparison, an output that presents options of character recognition candidates for the character recognition result and prompts manual verification processing, or the character recognition result. A seventh step of performing any one of an output prompting a manual input process by presenting a new input by a user and a determination with respect to the result.

[0013]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 schematically shows a configuration of a character recognition system according to the present embodiment.
In FIG. 1, a reading unit 11 includes a form H as an object to be read.
The scanner reads an image on the form H, and is configured by, for example, a scanner that reads an image on the form H by optical scanning and inputs the image as digital image data.

The form H is, for example, as shown in FIG.
Characters are described in a predetermined form. That is, in this example, an address field (area) C1 in which an address is described, a name field (area) C2 in which a name is described, an age field (area) C3 in which an age is described, and an occupation in which an occupation is described Columns (areas) C4 are provided, and these columns C
Each character in 1 to C4 is recognized.

The image data output from the reading unit 11 is
Sent to form identification unit 12. Form identification unit 12
Identifies the form of the form H based on the image data output from the reading unit 11, and sends the identification result to the control unit 13. The control unit 13 performs overall control and various processes, and an operation unit 14 and a form database 15 are connected to the control unit 13.

The operation unit 14 includes operation keys such as a keyboard for performing various operations and data input, and a display unit for displaying various guidance and input / output data. In the form database 15, the form of the form H handled by this system is registered and set. That is, for example, as shown in FIG. 3, a form number indicating a type of a form, a type of a character dictionary used for character recognition of the form, positions (coordinate values) and item names of columns C1 to C4 of the form, and The degree of certainty (P) required for character recognition in the column is registered and set.

The image data output from the reading unit 11 is sent to a character recognition unit 16. The character recognition unit 16
Based on the form data of the form H sent from the control unit 13, character recognition is performed by the image data output from the reading unit 11, and the recognition result is output. This character recognition will be described later in detail.

Next, in such a configuration, FIGS.
The flow of the character recognition process will be described with reference to the flowchart shown in FIG. The form H is set on the reading unit 11, and an image on the form H is input (step S1). The input image on the form H is sent to the form identification unit 12.
The form identification unit 12 identifies the form of the form H as described above (step S2), and outputs the identification result to the control unit 1
Send to 3. The control unit 13 checks whether or not the identified form of the form H is registered in the form database 15 based on the identification result from the form identification unit 12 (step S3).

As a result of this check, if the form of the form H is not registered in the form database 15, the form is registered in the form database 15 based on the form identification result of the form H. That is, first, the type of character dictionary used for character recognition of the form is selected (step S4). Next, the position (coordinate value) of the address column C1
Is set (step S5). Next, the item name of the address column C1 is set (step S6). Next, a certainty factor (P1) required for character recognition of the address column C1 is set (step S7). Next, it is checked whether or not there is a next column (step S8). In this example, since there is still a column, the process returns to step S5, and the setting process for the next name column C2 is repeated.

When the setting process for the occupation column C4, which is the last column, is completed, the processing capability of the present system is obtained by performing a predetermined calculation based on the above-described settings.
The obtained processing capacity is displayed on the operation unit 14 (step S9). The operator visually checks the display content of the operation unit 14 and determines whether or not the obtained processing capability is sufficient (step S10).

As a result of the determination in step S10, if the requested processing capability cannot be permitted, the operation unit 14 instructs the fact, and returns to step S5 to repeat the setting process. As a result of the determination in step S10, if the obtained processing capability can be permitted, the setting is set in steps S4 to S7 by instructing the operation unit 14 to that effect.
Is registered in the form database 15 (step S11), and then the process returns to step S1 to start the recognition process again from the beginning.

As a result of the check in step S3, if the form of the form H is registered in the form database 15, the form data corresponding to the form identification result of the form H is obtained from the form database 15 (step S12). By transmitting the obtained form data to the character recognition unit 16, character recognition based on the form data is performed.

First, based on the acquired form data, the character image areas of the character recognition columns C1 to C4 are divided from the image on the form H input by the reading unit 11 (step S13). Next, character recognition is performed on each of the divided character image regions using a character dictionary registered in the form, and a similarity is output as the recognition result (step S14). In this example, the character recognition is performed sequentially from the address column C1.

Next, the degree of certainty P1 of the address field C1 is compared with the degree of similarity S1 of the address field C1 obtained in step S14 (step S15). If [P1 ≦ S1], it is determined that the degree of certainty is high. It is assumed that the character recognition determination processing does not require the manual verification processing (step S16). That is, in this case, the similarity S1 is determined as the final character recognition result, and the determined character recognition result is displayed on the operation unit 14. In this case, since the degree of certainty is high and verification is not required, the display proceeds to the next character image at high speed. Therefore, the operator hardly needs to be aware of the recognition result of the character image. When the character recognition determination process is performed, the address column C of the processing status table 17 shown in FIG.
The number C of character recognition confirmation processes of 1 is incremented by "+1".

As a result of the comparison in step S15, [P1> S
If 1 ≧ P1-5%], the verification processing requires manual verification processing (step S1).
7). That is, in this case, since manual selection and confirmation are required, as shown in FIG. 8A, a highlighted display indicating the current character position is performed on the operation unit 14 for the character image. At the same time, as shown in FIG. 8B, which character candidate is listed as an option is displayed, and the default determination is displayed as the first candidate.
At first glance, if there is no problem with the default value, the operator performs the fixed input operation as it is. When the verification processing is performed, the processing status table 17 shown in FIG.
The number of verify processings V in the address column C1 is set to “+1”.

As a result of the comparison in step S15, [S1 <P
1-5%], a manual input process that requires a new manual input and confirmation is performed regardless of the value of the similarity S1 (step S18). That is, in this case, since a new input by hand is required, the character image is highlighted on the operation unit 14 so as to indicate the current character position in the same manner as described above. Basically, a new input screen is displayed at the same time. In the case of manual input processing, the number I of manual input processing in the address column C1 of the processing status table 17 shown in FIG. 7 is set to “+1”.

Next, the recognition processing section is advanced by one (step S).
19) It is checked whether or not all the fields have been completed (step S20). If not all the fields have been completed, the process returns to step S14 to perform the character recognition processing for the next field (name field C2).

The processing status table 17 shown in FIG.
Is provided in the control unit 13, and stores the number C of character recognition confirmation processing, the number V of verification processing, and the number I of manual input processing within a fixed period for each of the columns C1 to C4. The example of FIG. 7 illustrates a case where the period is 10 days and 1000 forms are processed.

As a result of the check in step S20, if all the columns are completed, it is checked whether or not all the columns have been subjected to character recognition determination processing (step S21). As a result of this check, when all the fields are the character recognition confirmation processing, a determination result of the character recognition confirmation classification is output to the form H (step S22), and the process proceeds to step S23.

If the result of the check in step S21 is that all the fields have not been subjected to character recognition confirmation processing, it is checked whether or not manual input processing has been performed (step S24). As a result of this check, if there is a manual input process, a determination result of a manual input category is output to the form H (step S).
25), and proceed to step S23.

If the result of the check in the step S24 shows that there is no manual input process, a judgment result of the verifying section is output to the form H (step S26), and the process proceeds to a step S23.

In step S23, the processing status table 17
, The ratio of the number of manually input processes I to the number of processed forms is calculated for each of the columns C1 to C4. Specifically, according to the example in which the period of FIG. 7 is from April 1 to April 10, the ratio of the number of manually input processes I in the address column C1 is 100/4.
000 (= 2.5%), the ratio of the number of manually input processes I in the name column C2 is 50/4000 (= 1.25%), and the ratio of the number of manually input processes I in the age column C3 is 50/4000 (= 1). .25
%), And the ratio of the number of manually input processes I in the occupation column C4 is 10/40.
00 (= 0.25%).

Next, it is checked whether or not the ratio of the calculated number of manual input processes I is equal to or greater than a predetermined value (for example, 10%). As a result of this check, if the calculated ratio of the number I of manual input processing is not equal to or greater than the predetermined value, the process proceeds to step S1.
To start the character recognition process for the next document, and if it is equal to or more than the predetermined value, output an abnormality warning to the operation unit 14 (step S27). Then, return to step S1 to start the character recognition process for the next document. I do.

Note that the character recognition decision section output in step S22, the manual input section output in step S25,
Each determination result of the verification classification output in step S26 is used for the classification processing of the form.

Further, it is possible to judge the capability of the present system and reset the certainty factor. That is, when determining the capability of the present system, a capability determination button (not shown) provided on the operation unit 14 is pressed. Then, the control unit 13 refers to the processing status table 17 to obtain the ratio of the total number of character recognition determination processes C, the number of verification processes V, and the number of manual input processes I of the total of each column to the number of forms processed.

More specifically, according to the example in which the period of FIG. 7 is from April 1 to April 10, the ratio of the number of character recognition confirmed processes C of the total of each column is 3500/4000 (= 87.5). %), And the ratio of the total number of verify processes V in each column is 290/4000.
(= 7.25%), the ratio of the number of manually input processes I in each column is 210/4000 (= 5.25%).

Further, the ratio of the number of manually input processes I to the number of processed forms is calculated for each of the columns C1 to C4. This can be obtained by a process similar to the process in step S23 described above. The ratios thus determined are displayed on the operation unit 14. Here, the operator resets the certainty factor for each of the columns C1 to C4 based on the content displayed on the operation unit 14.

[0038]

As described above in detail, according to the present invention, in a character recognition system that requires manual verification, the overall processing capability is improved by utilizing the character recognition result of an input image and its certainty factor. A character recognition method that can be improved can be provided.

[Brief description of the drawings]

FIG. 1 is a block diagram schematically showing a configuration of a character recognition system according to an embodiment of the present invention.

FIG. 2 is a diagram showing an example of a form as an object to be read.

FIG. 3 is a diagram showing an example of a form database.

FIG. 4 is a flowchart illustrating the flow of a character recognition process.

FIG. 5 is a flowchart illustrating the flow of a character recognition process.

FIG. 6 is a flowchart illustrating the flow of a character recognition process.

FIG. 7 is a diagram showing an example of a processing status table.

FIG. 8 is a view showing an example of a screen for selecting and confirming a character recognition result during a verification process.

[Explanation of symbols]

H: form (reading object), 11: reading unit, 12: form identification unit, 13: control unit, 14: operation unit, 15: form database, 16: character recognition unit, 17: processing status table.

   ────────────────────────────────────────────────── ─── Continuation of front page    (72) Inventor Hideo Horiuchi             70, Yanagicho, Saiwai-ku, Kawasaki City, Kanagawa Prefecture             Toshiba Yanagicho Office (72) Inventor Naoki Natori             70, Yanagicho, Saiwai-ku, Kawasaki City, Kanagawa Prefecture             Toshiba Yanagicho Office (72) Inventor Akihiko Nakao             70, Yanagicho, Saiwai-ku, Kawasaki City, Kanagawa Prefecture             Toshiba Yanagicho Office (72) Inventor Takuma Akagi             70, Yanagicho, Saiwai-ku, Kawasaki City, Kanagawa Prefecture             Toshiba Yanagicho Office (72) Inventor Noriyuki Hamamura             70, Yanagicho, Saiwai-ku, Kawasaki City, Kanagawa Prefecture             Toshiba Yanagicho Office F term (reference) 5B029 BB02 CC26                 5B064 AA01 BA01 EA08 EA11 EA12                       EA27 FA04 FA06 FA13

Claims (5)

[Claims] 1. A first step of inputting an image on a form on which characters are described, and performing character recognition on the input image input in the first step, and the similarity as a result of the character recognition And a third step of comparing the similarity obtained in the second step with a certainty factor required for the character recognition registered in advance, and a comparison in the third step. Output that does not require manual verification processing for the character recognition result, output that presents options of character recognition candidates for the character recognition result to prompt manual verification processing, or the character recognition result A fourth step of presenting a new input and confirmation by hand and performing an output prompting a manual input process. Character recognition method. 2. The form according to claim 1, wherein the form has a plurality of character description fields, and the processing from the second step to the fourth step is performed for each of the plurality of character description fields. Character recognition method. 3. A first step of inputting an image on a form on which characters are described in a predetermined form, and a second step of identifying the form on the form based on the input image input in the first step. And a third step of determining whether the form of the form identified in the second step is registered in advance. As a result of the determination in the third step, the form of the form is If the form has not been registered, the fourth step of registering the form and the degree of certainty required for character recognition of the form; and if the form of the form has been registered in advance as a result of the determination in the third step, Performing character recognition on the input image input in the first step based on the registered form, A fifth step of obtaining a similarity as a character recognition result; a sixth step of comparing the similarity obtained by the fifth step with the previously registered certainty factor; Based on the result of the comparison, an output that does not require a manual verification process for the character recognition result, an output that presents a character recognition candidate option for the character recognition result to prompt a manual verification process, or the character recognition result 7. A character recognition method comprising: a seventh step of performing one of a manual input process and an output prompting a manual input process by presenting a manual input and a determination to a result. 4. The form of the form has a plurality of character description fields, and the processing from the fourth step to the seventh step is performed for each of the plurality of character description fields. The character recognition method described. 5. The method according to claim 1, further comprising an eighth step of obtaining and outputting a processing capability based on the form and the certainty factor registered in the fourth step, wherein the processing capability output by the eighth step cannot be satisfied. 4. The character recognition method according to claim 3, wherein the registration of the form and the certainty factor in the fourth step is performed again by giving an instruction to that effect.