JP2000181995A - Character recognizing device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To enable the recognition of high accuracy with a suitable calculation quantity to the state of a document by measuring the suitability of respective dictionaries based on reliability in any one of the plural dictionaries and the character recognized result and performing the character recognition while using the dictionary selected out of the respective dictionaries corresponding to the measured suitability. SOLUTION: A suitability measuring part 13 holds correlative information defined between the state of the document and an output from a similarity calculating part 11 and estimates the scratch, crush or unknown font of a certain document at the time of operating. The estimated information is held as a data base, this data base is collated after pattern recognition and quality or suitability between a recognition object character and a dictionary is discriminated. Corresponding to this suitability, a dictionary selecting part 14 selects a dictionary 12, the character recognition is performed while using the selected dictionary 12, and the result is synthesized with the similarity calculating part 11. Thus, recognizing processing can be highly accurately performed.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition device for estimating the quality of a character from information such as character recognition of another line.

[0002]

2. Description of the Related Art Paper documents are printed in various fonts, and images are sometimes deteriorated due to repeated copying. On the other hand, the conventional character recognition device uses a fixed recognition method for various qualities and fonts. Therefore, there are drawbacks that a large amount of calculation is used for a high-quality document, and that the recognition accuracy is not improved even if a low-quality document is assumed to be a certain degree of deterioration of the document.

[0003] In addition, if an attempt is made to realize a highly accurate recognition rate for all qualities, and if a variety of patterns are learned more than necessary, a phenomenon called over-learning, in which the recognition accuracy is reduced by the addition of the patterns, occurs. It is known that there is. Therefore, it has been difficult for the conventional method to improve the recognition accuracy for documents of various qualities and unknown fonts.

[0004] Shinichiro Omachi and Hiroki Aso, "High-accuracy recognition of low-quality printed characters considering crushing," IEICE Transactions, (D-II) vol.J79-D-II, No.9, pp. .1535-1542 (199
6). This is only to generate a low-quality dictionary in consideration of collapse and to improve the recognition rate of degraded characters, and does not act adaptively on various types of degradation.

[0005] Shinichiro Omachi and Hiroki Aso, "Print Character Recognition Method Considering Quality", Information Processing Society of Japan, vol.38, N
o.12, pp.2490-2498 (1997). In this method, the quality judgment indicated in advance by the degree of crush is performed as preprocessing, and the recognition method is selected according to the quality. As a result, a processing amount equivalent to performing pattern recognition twice for the pattern is required.

Minako Sawaki, Hiroshi Murase, Norihiro Hagita, “Character Recognition by Automatic Selection of Reference Pattern Based on Deformation Estimation”,
IEICE Technical Report, PRMU97-187 (1997-12), pp.15-22 (Japanese Unexamined Patent Publication No. 7-2491
03) These perform the recognition twice as the initial dictionary and the degradation correspondence dictionary as above.

[0007]

According to the prior art, it is difficult to achieve both the recognition rate and the recognition speed because the same processing as that for a low-quality document is used when a high-quality document is fixedly recognized.

[0008] Further, if an attempt is made to learn a variety of learning patterns more than necessary in order to realize a high recognition rate for all documents, a phenomenon may occur in which patterns that could be recognized before learning cannot be recognized. It was difficult to get a correct reading rate.

SUMMARY OF THE INVENTION The present invention has been made in view of the above problem, and it has been presumed from the recognition result of an already recognized area that the degree of matching with a recognition dictionary in a specific area of a document has been estimated. It is an object of the present invention to perform high-precision recognition with an appropriate amount of calculation for the state of a document by using a dictionary, a classifier, or a classifier or a combination of classifiers.

[0010]

In order to solve these problems, the present invention provides a plurality of dictionaries to be referred to in character recognition, a recognized character recognition result and at least one of its reliability. And measuring means for measuring the fitness of the dictionary, and identification means for performing character recognition using a dictionary selected from the dictionary according to the fitness measured by the measuring means. And

A plurality of identification means for recognizing a character pattern; and a measuring means for measuring a degree of conformity between the identification means and a character based on at least one of the recognized character recognition result and its reliability. Means, and means for selecting or synthesizing the identification means according to the fitness measured by the measurement means.

[0012] Further, a fitness resetting means for resetting the fitness based on the fitness measurement result of an area including a region where the fitness between the dictionary or the identification means and the character and the character is measured or a proximity area, Means for performing character recognition again using the re-set fitness by the fitness re-setting means.

[0013] Further, a measuring means for measuring the fitness of the dictionary based on a plurality of dictionaries to be referred to at the time of recognition, at least one of the recognized character recognition result and its reliability, and this measuring means And a recognition unit for performing character recognition using a dictionary synthesized or transformed in accordance with the degree of conformity measured by the method.

A measuring means for measuring the degree of deformation or quality of the character based on at least one of the recognized character recognition result and its reliability; a large classification means for narrowing down candidates in character recognition; A means for increasing or decreasing the number of character candidates by the large classification means based on the degree of deformation or quality measured by the measurement means; and a detailed identification means for performing character recognition based on the character candidates by the means. It is characterized by doing.

A measuring means for measuring the degree of deformation or quality of a character based on at least one of the recognized character recognition result and its reliability, and a measuring means for measuring the degree of deformation or quality measured by the measuring means. A means for calculating a parameter of similarity calculation; and a means for calculating similarity based on the parameter calculated by this means.

Further, a measuring means for measuring the degree of deformation or quality of a character based on at least one of the recognized character recognition result and its reliability, and a plurality of first character candidates for narrowing down the first character candidates at the time of character recognition. The first character candidate output from the large classifying unit and the second character classifying unit outputs a second character candidate based on the degree of deformation or quality measured by the measuring unit.
Means for determining a character candidate.

In the thus configured character recognition device, the integration method of a plurality of large classification means is changed according to the degree of deformation or quality of the character. For this reason, when there is little deformation and the image quality of the document image is good, it can be read reliably, and if the candidate category output from any of the major classification means is identified by detailed identification at the subsequent stage, the number of candidates can be reduced and high speed can be achieved. Can be recognized. On the other hand, when the character deformation is large or the image quality of the document image is inferior, if it is output to any one of the large classification means, the detailed identification can be performed, and the recognition processing can be performed with high accuracy.

[0018]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS A character recognition apparatus according to an embodiment of the present invention will be described below with reference to the drawings.

As an embodiment corresponding to claim 1 of the present invention, recognition is performed line by line as shown in FIG. 1, and the relevance between the dictionary currently used for recognition and the image to be recognized is estimated from the recognition result. Next, an apparatus for recognizing a recognition target to be recognized next using an optimal dictionary will be described. This is different from the prior art,
It is based on the evaluation of character deformation (or quality) in the already read document area.

FIG. 2 is a block diagram showing the configuration of a character recognition apparatus according to an embodiment of the present invention. The similarity calculation unit 11 for obtaining the similarity of a character pattern and a plurality of units referred to in identification. And a dictionary selecting unit 14 for selecting a dictionary according to the estimated degree of matching using the dictionary 12 and the recognized recognition result to measure the degree of matching with the dictionary.

The similarity calculator 11 outputs a recognition result and a similarity for the input pattern. The similarity calculation unit 11 has a configuration in which the dictionary 12 can be switched. A combination of the similarity calculation unit 11 and the dictionary 12 is called a discriminator 15. The following equation shows the definition of the composite similarity.

(Equation 1) The dictionaries and weights are obtained by principal component analysis.

The goodness-of-fit measuring unit 13 holds correlation information defined between the state of the document and the output from the similarity calculating unit 11, and operates based on the similarity tendency when a certain document is recognized during operation. (For example, blurring, crushing, unknown font, etc.).

Generally, when a blurred pattern or a crushed pattern is recognized with respect to a dictionary learned by a normal pattern, the similarity is low. In addition, there is a case where a certain tendency appears in a candidate character as a recognition result. Based on these pieces of information, it is possible to check the matching degree between the recognition target character and the dictionary or various thresholds.

As another method, the thickness of a stroke or the like can be examined structurally, or the blurring or crushing of a character can be examined from the density of black pixels or the number of black connected components. It is also possible to hold such information as a database and collate the database after recognition to determine the quality and compatibility between the recognition target character and the dictionary.

The dictionary selecting section 14 selects the dictionary 12 according to the output from the fitness measuring section 13 (fitness to the dictionary), and synthesizes it with the similarity calculating section 11.

Next, an overall processing flow of an example corresponding to claim 1 will be described with reference to FIG. In the character recognition apparatus according to the first embodiment, a phase of recognizing a pattern (S1
1), a phase for measuring the degree of matching with the dictionary (S1)
2) and three phases of combining the identification units (S13).

In the phase S11 of recognizing a pattern, a pattern to be recognized is input to a classifier, and a similarity and a recognition result are output. Phase S1 for measuring the relevance to the dictionary
In step 2, the relevance between the recognized area and the dictionary currently used is measured based on the similarity at the time of recognition. In phase S13 of combining the recognition units, a dictionary that is considered to be most suitable is selected and combined with the similarity calculation unit to form a classifier.

By repeating the phases S11 to S13, dictionaries are dynamically switched, and recognition adaptive to quality and character deformation can be performed.

Further, in the above description, the degree of conformity is obtained in units of one line as shown in FIG. 1, but the degree of conformity may be obtained in units of characters or partial images, and a dictionary may be selected.

FIG. 4 is a block diagram showing the configuration of a character recognition apparatus according to a second embodiment of the present invention. A plurality of discriminators 21 for identifying a character pattern and a dictionary using a recognition result of another pattern are used. And a recognizing unit integrating unit 23 that determines a combination method of classifiers from the degree of conformity and a plurality of identification results.

The discriminator 21 and the fitness measurer 22 are shown in FIG.
In FIG. 4, there are a plurality of discriminators 21. Each discriminator 21 may handle the same feature, but by using different features, the characteristics of each feature can be effectively utilized.

The discriminator integrating unit 23 determines a combination of a plurality of discriminators according to the fitness measured by the fitness measuring unit 22, and performs recognition based on the combination. As shown in FIG. 5, for the combination, C that couples the recognition units in series
Not only 1 but also connecting in parallel, the results can be integrated with specific weighting, C2, or branch C3 or feedback can be performed under specific conditions.

FIG. 6 is a flow chart of an embodiment according to the second aspect. The difference from FIG. 3 is that the entire discriminator is selected instead of selecting a dictionary. As a matter of course, a mechanism for selecting a dictionary to be handled according to a situation such as a dictionary matching degree may be provided.

FIG. 7 is a conceptual diagram of an embodiment according to claim 3. In the embodiments corresponding to the first and second aspects, in order to measure the degree of matching of the dictionary, a more accurate result is generally obtained by applying a wider range of measurement results. For example, if it is determined that the quality is globally low and only a part of the local area is estimated to be high quality, the recognition tends to increase in the number of erroneous recognitions. In this case, the accuracy of character recognition can be improved by performing recognition again using the global quality measurement result.

As one embodiment corresponding to claim 4 of the present invention, a character recognition apparatus using dictionary synthesis corresponding to a plurality of transformations will be described. FIG. 8 is a conceptual explanatory diagram relating to the measurement of the degree of matching between the recognition target and the dictionary.

When a recognition dictionary is generated, a standard dictionary generated from a standard pattern set, two dictionaries generated by subjecting the standard dictionary to thickening / thinning processing, and two dictionaries tilted to the left and right are used. Prepare a dictionary in advance. Recognize input patterns with multiple dictionaries. If the recognition results are the same, character deformation information is obtained according to the similarity,
Accumulate.

If there is a difference in the recognition result by the five dictionaries, the target character pattern is registered as a mark to be recognized again. When the recognition of a certain area is completed, the tendency of deformation in the area is obtained from the stored deformation information, and a dictionary is synthesized based on the obtained information.

In the dictionary synthesizing method, a difference between the transformed dictionary and the standard dictionary is added by the weight of the transformed information. Since the dictionary obtained by the above method is the dictionary closest to the tendency of deformation in the area, highly accurate recognition can be performed.

FIG. 9 shows an execution procedure of an example corresponding to claim 4. The difference from the average pattern is output in addition to the result of character recognition, and the dictionary is transformed based on the tendency. By using the dictionary modified in this way, it is possible to follow the tendency of deterioration of the document and the tendency of the font, and it is possible to perform recognition with high accuracy.

FIG. 10 shows an example of claim 4, in which the degree of crushing is measured in units of characters, and the dictionary of the next pattern is transformed with the degree of crushing calculated for the next pattern to perform recognition. I have.

FIG. 11 shows the concept of calculating the degree of crush. In a space that defines a dictionary for each category,
The average vector of each feature vector in the space can be defined as a standard pattern. Detects the position from the standard pattern,
By investigating the tendency of deterioration at various positions (for example, inclination, blurring, crushing, etc.) in advance, it is possible to know the tendency of deterioration of the input pattern. It is unlikely that the tendency of deterioration will differ significantly for each character or local area in a document. Therefore, by recognizing adjacent recognition target areas assuming the same deterioration tendency, a recognition dictionary adaptive to the deterioration tendency of the document can be formed, and the recognition accuracy can be improved.

As one embodiment corresponding to the fifth aspect of the present invention, a character recognition apparatus for determining the number of character candidates according to a large classification according to the degree of deformation (or quality) of a character will be described. As a configuration of the present invention, for example, a recognition unit having a multi-stage configuration such as large classification / intermediate classification / detailed identification as shown in FIG. 12 is considered. In addition, each classification unit and the detailed identification unit can change internal parameters by using the degree of character deformation (or quality) as input.

FIG. 13 is a diagram for explaining the concept of increasing or decreasing the number of candidates according to the degree of character deformation. By increasing or decreasing the candidate list from the large classification unit to the detailed identification unit according to the estimated degree of deformation of the character, the amount of calculation in the detailed identification unit is increased or decreased.

For example, when the similarity is used for measuring the degree of deformation of a character, the degree of deformation of the character is estimated from the first-place similarity and (first-place similarity-second-place similarity) (see FIG. 14). The number of large classification candidates is determined using the estimated degree of deformation (see FIG. 15).
When the degree of deformation is estimated to be large, the number of character candidates to be recognized in the detailed identification is increased, and when the degree of deformation is estimated to be small, the number of character candidates is reduced.

As one embodiment corresponding to claim 6 of the present invention, a character recognition apparatus for changing a parameter for determining the calculation accuracy (or calculation amount) of a recognition unit according to the degree of deformation of a character will be described. In the present invention, for example, when the compound similarity method is used, the parameters in the following equation

(Equation 2) Generally, a larger value is required as the degree of deformation increases.
When it is determined that the degree of deformation is small, the degree of similarity is once obtained in the middle, and the remainder is calculated only for the top candidates among them.

(Equation 3) In this equation

(Equation 4) , The calculation is temporarily stopped at X, and the similarity at the intermediate stage is obtained. The result is sorted, and if the degree of deformation of the character is small, the rest is not calculated, or the calculation is continued with only a very small number of upper candidates. When the degree of deformation is large, the calculation is not interrupted in the middle stage but is calculated to the end. This allows
For patterns with little deformation, the amount of calculation can be greatly reduced.

As one embodiment corresponding to claim 7 of the present invention, a character recognition apparatus for changing the method of integrating a plurality of large classification part character candidate lists in accordance with the degree of character deformation will be described.

The present invention is applicable to a case where the character deformation degree of the input pattern is estimated to be large in the character candidate list (first character candidate) output from the large classification unit for a plurality of various features as shown in FIG. Takes the OR of the candidates and takes the AND of the candidates if the degree of character deformation is small (second character candidate).

According to this method, due to a synergistic effect using different features, in the case of OR, the leakage of candidates at the time of classification can be further reduced, and in the case of AND, the number of candidates can be greatly reduced.

The processing in the embodiment of the present invention is realized by a computer-executable program, and the present invention can be realized by a computer-readable storage medium.

The storage medium in the present invention includes a magnetic disk, a floppy disk, a hard disk, and an optical disk (CD-ROM, CD-R, DVD).
Etc.), a magneto-optical disk (MO or the like), a semiconductor memory, or any other storage medium that can store a program and that can be read by a computer.

An OS (operation system) running on the computer, database management software, MW (middleware) such as a network, etc., based on instructions of a program installed in the computer from the storage medium realizes this embodiment. May be executed.

Further, the storage medium in the present invention is not limited to a medium independent of a computer, but also includes a storage medium in which a program transmitted via a LAN, the Internet, or the like is downloaded and stored or temporarily stored.

Further, the number of storage media is not limited to one, and a case where the processing in the present embodiment is executed from a plurality of media is also included in the storage medium of the present invention, and the configuration of the medium is any configuration. Good.

The computer according to the present invention is:
The computer executes each process in the present embodiment based on a program stored in a storage medium, and may have any configuration such as an apparatus such as a personal computer or a system in which a plurality of apparatuses are connected to a network. Is also good.

Further, the computer in the present invention is not limited to a personal computer, but also includes an arithmetic processing unit, a microcomputer, and the like included in an information processing device, and generically refers to devices and devices capable of realizing the functions of the present invention by a program. are doing.

[0056]

As described above, according to the present invention, the quality is pre-estimated at the time of pattern recognition, and the next pattern is recognized according to the result, whereby the character pattern can be processed with a necessary and sufficient processing amount and with high accuracy. Can be recognized.

[Brief description of the drawings]

FIG. 1 is a conceptual explanatory diagram of an example corresponding to claim 1 of the present invention.

FIG. 2 is a block diagram of a character recognition device according to claim 1 of the present invention.

FIG. 3 is a flowchart of processing of a character recognition device according to claim 1 of the present invention.

FIG. 4 is a block diagram of a character recognition device according to claim 2 of the present invention.

FIG. 5 is a diagram for explaining the concept of integration processing of a character recognition device according to claim 2 of the present invention.

FIG. 6 is a flowchart of a process performed by the character recognition device according to claim 2 of the present invention;

FIG. 7 is a conceptual explanatory view of an example corresponding to claim 3 of the present invention.

FIG. 8 is a conceptual diagram in the case of dictionary modification of an example corresponding to claim 4 of the present invention.

FIG. 9 is a flowchart in the case of dictionary modification of an example corresponding to claim 4 of the present invention.

FIG. 10 is a conceptual explanatory diagram of an example corresponding to claim 4 of the present invention.

FIG. 11 is a conceptual explanatory diagram of an example corresponding to claim 4 of the present invention.

FIG. 12 is an overall configuration diagram of an example corresponding to claim 5 of the present invention.

FIG. 13 is a conceptual explanatory view of an example corresponding to claim 5 of the present invention.

FIG. 14 is a correlation diagram between the similarity and the degree of deformation of the example corresponding to claim 5 of the present invention.

FIG. 15 shows the degree of deformation and the number of candidate characters in an example corresponding to claim 5 of the present invention.

FIG. 16 is a diagram showing integration of a candidate character list in an example corresponding to claim 7 of the present invention.

[Explanation of symbols]

 11: Similarity calculation unit 12, 21: Dictionary 13, 22 ... Fitness measurement unit 14: Dictionary selection unit 15: Classifier 23: Classifier integration unit

Claims (7)

[Claims] 1. A plurality of dictionaries to be referred to in character recognition, a measuring means for measuring the fitness of the dictionary based on at least one of the recognized character recognition result and its reliability, A character recognition unit that performs character recognition using a dictionary selected from the dictionaries in accordance with the degree of conformity measured by the means. 2. A plurality of identification means for recognizing a character pattern, and measurement means for measuring a degree of conformity between the identification means and a character based on at least one of a recognized character recognition result and its reliability. And a means for selecting or synthesizing the identification means according to the degree of conformity measured by the measurement means. 3. A fitness resetting means for resetting fitness based on a fitness measurement result of a region including a region where the fitness of the dictionary or the identification means and the character is measured or a proximity area, 3. The character recognition apparatus according to claim 1, further comprising: means for performing character recognition again by using the re-set relevance by the re-set relevance unit. 4. A plurality of dictionaries to be referred to at the time of recognition, a measuring means for measuring the fitness of the dictionary based on at least one of a recognized character recognition result and its reliability, and a measuring means. A character recognition unit that performs character recognition using a dictionary synthesized or modified in accordance with the degree of conformity measured by the character recognition apparatus. 5. A measuring means for measuring the degree of deformation or quality of a character based on at least one of a recognized character recognition result and its reliability, a large classification means for narrowing down candidates in character recognition, A means for increasing or decreasing the number of character candidates by the large classification means based on the degree of deformation or quality measured by the measurement means; and a detailed identification means for performing character recognition based on the character candidates by the means. A character recognition device. 6. A measuring means for measuring the degree of deformation or quality of a character based on at least one of a recognized character recognition result and its reliability, and a measuring means for measuring the degree of deformation or quality measured by the measuring means. A character recognition device comprising: means for calculating a parameter for similarity calculation; and means for performing similarity calculation based on the parameter calculated by this means. 7. A measuring means for measuring the degree of deformation or quality of a character based on at least one of a recognized character recognition result and its reliability, and a plurality of first character candidates for narrowing down first character candidates in character recognition. Major classification means; and means for determining a second character candidate for the first character candidates output from the plurality of major classification means based on the degree of deformation or quality measured by the measurement means. A character recognition device comprising: