JP2000293626A - Method and device for recognizing character and storage medium - Google Patents

Abstract

PROBLEM TO BE SOLVED: To perform performance improvement optimized to an object that is actually processed by inputting a correct character string, acquiring knowledge for obtaining a correct preprocessing result with the correct character string as a key and updating knowledge to be used in a preprocessing process by the knowledge when a reject or an error is outputted as a final processing result. SOLUTION: A character segmenting part 24 uses knowledge stored in a preprocessing knowledge storing part 22 and performs image division of a detected character string area image as a character in each an optimum block. An individual character recognizing part 25 recognizes an individual divided image as a character. A knowledge processing part 26 adapts knowledge about a preliminarily given character string, constructs it as an appropriate character string to the recognition results of an individual character and outputs it as a final processed result. A correction processing part 27 performs correction manually when a reject or an error is outputted as the final processed result. A learning mechanism part 20 updates this knowledge to learn stored knowledge.

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 and a character recognition device (so-called OCR; Optical Character Recognition) for optically reading characters described in a processing object and performing character recognition.
acter Reader), in particular, processes a free-form recognition target and, from among many character strings and patterns included in the image to be recognized, recognizes a recognition target area that includes a character string to be actually subjected to character recognition. The present invention relates to a character recognition method and apparatus for detecting and performing character and character string recognition.

[0002]

2. Description of the Related Art An optical character recognizing apparatus generally called an OCR reads an object on which a handwritten character is written or a printed character is printed by an image input device (scanner) and obtains an image to be recognized. Recognize the character strings contained in the image and recognize the characters. Since the image input device itself is a general-purpose technology and can also perform character recognition from existing image data to be recognized, the characterizing optical character recognition device extracts a character string from the image to be recognized. To recognize characters. In this sense, the present specification deals with a technique for recognizing characters included in an image from an image without being bound by the phrase “optical”.

[0003] An entry frame whose position and size are predetermined is
If characters are entered and printed one by one,
What is necessary is just to perform individual character recognition for every character. However, when recognizing a free-form, that is, a character written or printed without defining a format, for example, when extracting a character string to be an address from a mail and recognizing a character in the character string, First, processing for extracting individual characters to be recognized from the recognition target image (referred to as preprocessing) is required. If free-form character recognition is used to separate mail, a block-like area in which characters and character strings of the destination address (or name) are written instead of the sender's address as preprocessing. It is necessary to extract (address description area) as a recognition target area, extract a character string image for each line from that area, and cut out characters one by one from each character string image. Furthermore, as preprocessing,
To make it easier to recognize individual characters in the subsequent process, processing to compensate for blurred characters, processing to remove noise due to stains and dirt on the mail surface, processing to correct skewed character images, underlining Processing for detecting (underlining) and removing it, processing for equalizing (normalizing) the size of characters, and the like are executed as necessary. In character extraction, one character that should be written as "Ki" can be cut out as "word" and "me", or two characters that should be written as "Mihara" can be extracted as "source". It is important to prevent erroneous clipping such as clipping with one character corresponding to "."

It is known that it is effective to provide a learning function to individual character recognition after character segmentation in order to deal with differences in writing habits and fonts (fonts). Many proposals have been made for a learning function in individual character recognition in a character recognition device. For example, JP-A-8-180141, "Character Recognition System", JP-A-5-94565, "Handwritten Character Recognition System",
Japanese Unexamined Patent Publication No. H05-054196 "License plate recognition device" discloses learning on individual character recognition itself.

[0005] It is also known to use a knowledge base in character recognition. For example, JP-A-10-1987
In Japanese Patent No. 64, “Character string recognition device and knowledge database learning method”, when a recognition candidate character string is an unregistered word, the recognition candidate character string is registered in a knowledge database, and the character recognition processing is performed. A technology that enables automatic learning of a knowledge database to be applied is disclosed. Also,
Japanese Patent Laying-Open No. 7-6203 discloses a character recognition device that performs learning for a specific format called a form. Japanese Patent No. 2751865 (JP-A-8-287188) discloses an application of a knowledge base to character extraction in preprocessing.

FIG. 12 is a block diagram showing a configuration of a conventional character recognition device. Here, a character recognition device used for a mail sorting device will be described. The character recognition device includes an image input unit 91 that captures a processing target (mail) as binary or multi-valued recognition target image data, and a preprocessing that previously stores knowledge given as a format database or processing parameters. Knowledge storage unit 92
And a character for detecting a destination name description area (recognition target area) from the recognition target image data acquired by the image input section 91 and extracting a character string image by using the knowledge stored in the preprocessing knowledge storage section 92 A character segmentation unit 94 that divides a detected character string image into blocks that are considered to be optimal as characters using the knowledge similarly stored in the preprocessing knowledge storage unit 92; An individual character recognizing unit 95 that recognizes each divided image as a character, and a predetermined character string knowledge (such as place name information) is applied to the recognition result of the individual character to construct an appropriate character string. , A knowledge processing unit 9 for outputting as a final processing result
6 is provided. If an appropriate character string is not found, that is, if it is determined that a correct recognition result cannot be obtained, the knowledge processing unit 96 outputs a reject as a final processing result. In addition, the character recognition device manually calibrates the rejection or error (error) as a final processing result, and corrects the correct destination of the postal matter (for example, the postal code (7 digits) and the street address building). There is provided a calibration processing unit 97 for manually inputting (room information) to the mail sorting apparatus.

A mail sorting device using this character recognition device sorts mails by the output of the knowledge processing section 96,
When a rejection or other error is detected as an output of the knowledge processing unit 96, the mail is classified based on the calibration result of the calibration processing unit 97. In this conventional character recognition device, when the character is cut out by the character cutout unit 94,
The method described in the aforementioned Japanese Patent No. 2751865 can be used. Further, as the processing in the individual character recognition unit 95 and the knowledge processing unit 96, a character recognition method having a learning function and performing knowledge processing can be applied.

[0008]

In the conventional character recognition apparatus shown in FIG. 12, preprocessing by knowledge processing is executed, and individual character recognition which is a knowledge processing and has a learning function is performed. However, learning is not performed on preprocessing (detection of a recognition target area or character cutout), which is particularly important processing in free-form character recognition. This is largely due to the fact that conventional character recognition technologies have been developed for fixed forms, and have not paid much attention to the detection of the recognition target area and the extraction of characters. However, as the demand for a character recognition device for free format has increased, performance problems at the preprocessing stage have increased, and learning for optimizing those processes for operation targets has become essential.

Here, the performance of the character recognition device will be described. At present, character recognition by a character recognition device has not yet reached the level of a human, and therefore, improvement of its reading performance has become a technically significant problem. The performance is, in the simplest case, how correctly the recognition has been made. Specific indexes indicating the performance include, for example, the correct answer rate of the "address detection area detection processing" in the preprocessing,
The correct answer rate in the “character extraction processing” (the correct answer content rate in the case of multiple candidate processing), the individual character recognition correct read rate, the correct answer rate in the knowledge processing, and the like are included.

[0010] A problem common to character recognition for free format is that it is difficult to optimize and commercialize a variation or description of a processing target to be processed in actual operation in advance. For example, as a character recognition device for a free format which is currently in practical use, there is a character recognition device for a destination described in a mail as described above. In the performance, various pictures,
Detecting a destination character string from among texts, logos, etc., has become a major issue, and cannot be handled by pre-processing prepared in advance, for example, detection processing of a destination name description area (recognition target area) and character cutout processing. Since there are many mails and the addresses to be described mainly differ depending on the area where the mail is operated, it is necessary to improve the performance of each character recognition device by learning and optimizing according to the actual operation.

However, a method of optimizing these processes for individual operation situations by automatic learning and improving the performance has not been studied so far.

Accordingly, an object of the present invention is to enable learning of preprocessing such as a recognition target area detection process and a character segmentation process in a free-form character recognition process, and to improve performance optimized for a target to be actually processed. A character recognition method and apparatus are provided.

[0013]

In general, in free-form character recognition, rejects and errors are caused by failures in detection of a recognition target area, extraction of a character string area, and character extraction processing in preprocessing. . On the other hand, even if it is finally rejected, it is rare that no character string candidate can be generated at all, and the candidate is created but eventually rejected because it is not reliable. In many cases, candidates are generated, and at the stage of determining one of them, none of them can be determined and rejected. The same is true for erroneous recognition. In many cases, an error is finally selected from a plurality of candidates.

[0014] As a cause that the correct candidate cannot be left as the final candidate, there is a problem that it is difficult to optimize and commercialize a variation or description of a processing target to be processed in actual operation in advance.
Many methods for optimizing the recognition process in individual character recognition to the operational situation have already been published, but if the pre-processing can be optimized for the actual operational situation, the character recognition performance for free format will be remarkably improved. Can be improved.

Therefore, in the present invention, the knowledge used in the candidate detection and preselection stages in the preprocessing is optimized for the actual operation by using the information obtained when the rejected or misrecognized ones are manually calibrated. By doing so, the performance is improved.

That is, the character recognition method of the present invention is a character recognition method for performing character recognition on a recognition target image in which characters are described in a free format. Applying pre-given knowledge about a character string to a pre-processing step of cutting out a block as a divided image, an individual character recognition step of performing individual character recognition on each of the divided images, and an individual character recognition result. Performs a character string construction step of constructing an appropriate character string and outputting it as the final processing result, and a proofreading processing step of inputting a correct character string when a rejection or error is output as the final processing result. In such a case, knowledge for obtaining a correct preprocessing result is acquired using the correct character string as a key, and the knowledge used in the preprocessing step is obtained based on the acquired knowledge. ; And a learning step of updating the.

The character recognition apparatus of the present invention is a character recognition apparatus for performing character recognition on an image to be recognized in which characters are described in a free format. From the storage means and the image to be recognized,
A character string area detecting means for detecting a desired character string area using knowledge previously given to the preprocessing knowledge storage means, and a character string detected using the knowledge stored in the preprocessing knowledge storage means. A character cutout unit that obtains a divided image by dividing an area image into blocks that are considered to be optimal as characters, an individual character recognizing unit that recognizes each divided image as a character, and a recognition result of an individual character are given in advance. Knowledge processing means for applying knowledge about the given character string, constructing it as an appropriate character string, and outputting it as a final processing result, and rejecting or outputting an error as a final processing result; and Learning means for learning the knowledge in the preprocessing knowledge storage means, and when the calibration processing is performed by the calibration processing means, the correct preprocessing result is obtained by using the correct data as a key. Obtained knowledge acquired for, updating the knowledge in the preprocessing knowledge storing means accumulate knowledge preprocessing knowledge storage means based on the acquired knowledge.

That is, according to the present invention, the character string area detecting means converts the recognition target given as a database or processing parameter in the preprocessing knowledge storage means from a recognition target of a free format such as an address written on an envelope. By
Detect the desired character string area. Next, the character string portion obtained by the character string region detecting means is similarly image-divided into blocks each of which is considered to be optimal as a character by the character extracting means, using the knowledge stored in the preprocessing knowledge storage means. You. Next, the character recognition is performed by the individual character recognition means, and the address stored in the knowledge processing means,
It is constructed as the most appropriate character string based on knowledge such as name,
Output as the final processing result. In character recognition by a character recognition device, a character or a character string that cannot be correctly recognized by the OCR in the proofreading processing means is manually input to create a correct recognition result because the character may be rejected or may include an error. .

According to the present invention, a series of processes from the pre-processing is repeatedly and re-executed by using the correct data obtained by the calibration processing means as a key, thereby acquiring knowledge, parameters, and the like for obtaining a pre-processing result of the correct solution. Then, by giving the function to accumulate and update them in the preprocessing knowledge storage means to the character recognition device, the preprocessing itself is optimal for the actual operation situation and the input image while using the existing technology. The present invention provides a character recognition method and apparatus capable of realizing preprocessing.

[0020]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Next, a preferred embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing the configuration of the character recognition device according to the first embodiment of the present invention.

Here, a case where free-form character recognition is performed for the purpose of sorting mail items will be described as an example. FIG. 2 shows a sealed letter 10 on which a stamp is affixed, a company logo (a kind of company emblem) is printed, and an address 11 and a sender 12 are described. Here, from a recognition target image obtained by image-inputting the sealed letter, a block-shaped area (address enclosed by a broken line in the drawing) in which a destination name is described as a recognition target area which is a target area for character recognition. The name description area 13) is detected, and then the destination name description area 1 is detected.
3, a description will be given assuming that a character string area of each line of the destination name is extracted, character extraction processing is performed on the character string area, and processing of recognizing the described character string is performed. In FIG. 2, the address description area 13 is clearly indicated by a broken line, but such a broken line (area) is not described in the actual letter 10, and this block area will be described later. As a result of the preprocessing, the character is recognized and extracted for the first time in the character recognition device. In the example shown in FIG. 2, the character string area of each line of the destination name is a character string “〒”.
183-0036 ", an area containing the character string" 1-10 Nissincho, Fuchu-shi, Tokyo ", an area containing the character string" ○ △ ○ △ KK ", and a character string" Mr. . Are four areas in total.

The character recognition apparatus shown in FIG. 1 reads an object to be processed (here, sealed letter 10) by a photoelectric conversion element such as a CCD (Charge Coupled Device) sensor or the like, and recognizes the image to be recognized as binary or multi-valued digital image data. Using an image input unit 21 for prefetching, a preprocessing knowledge storage unit 22 for preliminarily storing knowledge used for preprocessing in the form of a format database or processing parameters, and the knowledge stored in the preprocessing knowledge storage unit 22 Thereby, the character string area detecting unit 2 extracts the address description area 13 as a block-like area from the recognition target image acquired by the image input unit 21 and detects the character string area from the extracted address description area 13.
3, a character cutout unit 24 that uses the knowledge stored in the preprocessing knowledge storage unit 22 to divide the detected character string area image into blocks each of which is considered to be optimal as a character. An individual character recognizing unit 25 that recognizes an image as a character, and a knowledge about a character string (such as place name information) that is given in advance to the recognition result of the individual character is applied to construct an appropriate character string. And a knowledge processing unit 26 that outputs the result. Further, the character recognition device performs a calibration processing unit 27 for manually correcting when a rejection or an error is output as a final processing result from the knowledge processing unit 16 and a learning of the knowledge stored in the preprocessing knowledge storage unit 22. And a learning mechanism unit 20 for updating the knowledge in order to perform the updating.

In this character recognition device, information can be mutually exchanged from the character string area detecting section 23 to the knowledge processing section 26. In particular, the knowledge processing unit 2
6 is fed back to the character string area detection unit 23 and the character cutout unit 24, and when the proofreading processing is performed in the proofreading processing unit 28, the content of the proofreading processing is performed by the character string area detection unit 23 and the character cutout This is fed back to the unit 24. The learning mechanism unit 29 includes the character string area detection unit 2
The learning of the knowledge in the preprocessing knowledge storage unit 22 is executed in accordance with 3 and the content of the feedback to the character cutout unit 24.

The knowledge processing unit 26 includes, for example, a place name dictionary as a knowledge base, and performs knowledge processing based on the recognition result of the individual character output by the individual character recognition unit 25.
Output the final recognized character (string). By performing the knowledge processing, even if the individual character recognizing unit 25 erroneously recognizes the character of “king” as “ball”, the place name “Hachioji-shi” (Hachioji-shi is located in the western part of Tokyo. There is no place name "Hachitamago City" even though there is a city, so it can be correctly recognized as "Hachioji City".

In particular, the knowledge processing section 26 of the character recognition apparatus described here stores information on the format as well as information on the character string as knowledge, and from the knowledge processing section 26, the character area detection section 23 and the character cutout section 24. It has a function to create preprocessing candidates from the top down using knowledge processing, such as direct feedback to. That is, in this character processing device, at the stage of constructing a character string by performing knowledge processing when performing character recognition processing for a free format, it is determined whether or not the preprocessing result given is appropriate and fed back to the preprocessing. The top-down process is used, and the flow of information from the character string area detection unit 23 to the knowledge processing unit 26 is not unique. For example, the most appropriate recognition target area is detected as a result of performing the processing up to the knowledge processing on a plurality of recognition target area candidates, and the final processing is performed again using the optimal parameters, processing, and processing procedure from the preprocessing. Techniques for obtaining results are also used.

In the character recognition apparatus shown in FIG. 1, the correct character string obtained by the proofreading processing unit 27 is given, and a series of processing from the character string area detecting unit 23 is repeatedly executed, so that the correct answer is initially obtained. The learning mechanism unit 28 detects and recognizes the correct character string from the processing target (e.g., postal matter) that could not be obtained, and uses the preprocessing knowledge at the time of obtaining the correct answer, stores the new character string in the preprocessing knowledge storage unit 22. Or the knowledge stored in the preprocessing knowledge storage unit 22 is updated.

Next, the operation of the character recognition apparatus shown in FIG. 1 will be described with reference to the flowchart shown in FIG.

First, the image input unit 21 scans an object to be read (for example, a sealed letter) and captures it as an image to be recognized as binary or multi-valued digital image data (step 101). For this recognition target image, the character string area detection unit 23 uses the knowledge stored in the preprocessing knowledge storage unit 22 to store a block-shaped area (see FIG. The destination name description area 13) is determined (step 102), and a character string area for each line is extracted from the area as a character string area image (step 1).
03). Next, the character cutout unit 24 uses the knowledge stored in the preprocessing knowledge storage unit 22 to divide the image of the character string region image into blocks that are considered to be optimal as one character (step 104). . This step 10
The processing of No. 4 is usually called character extraction processing. Note that the processing of steps 102 to 104 is collectively called preprocessing.

The pre-processing knowledge storage unit 22 preliminarily estimates, for example, an object to be processed as knowledge necessary for the character string area detection unit 23 to detect a block-shaped area which is a destination addressing area. A typical description format (description pattern), processing parameters necessary for selecting the most appropriate format from several formats, or the processing procedure itself are stored. The description format or description pattern is, for example,
For a certain type of envelope, the point at which the address moved from the top left vertex of the envelope by how many centimeters and to the right how many centimeters was defined as
How large and small in the area are described,
Such knowledge data. Since there is a case where the description format cannot be estimated in advance, the preprocessing knowledge storage unit 22 also stores parameters used for estimating and detecting an appropriate candidate area from the character size and the like, the processing procedure itself, and the like. Is preferred.

The preprocessing knowledge storage unit 22 stores, as knowledge used in the character extracting unit 24, parameters such as a character size and a pitch estimating method assumed at the time of character extracting and a processing procedure of the character extracting process. It itself is stored. Here, the parameters used for the preprocessing by the knowledge processing are collectively stored in a single preprocessing knowledge storage unit 22, but in actual operation, the parameters are distributed and stored for each processing. You may make it accumulate. Further, when there are a plurality of processing methods that can be used in the preprocessing, the preprocessing knowledge storage unit 22 may store knowledge for selecting a processing method according to the recognition target image.

In step 104, character cutout unit 2
Next, the individual divided images divided by 4 are subjected to character recognition by the individual character recognition unit 25 (step 10).
5) Further, it is constructed as an appropriate character string by the knowledge processing unit 26 (step 106) and output as a final processing result (step 107).

A final result as the character recognition device (OCR) is obtained by the processing up to the above step 107. The final output result is the detection of the destination addressing area in the step 102 and the step 103 Character area detection, character extraction in step 104, step 10
5 may be rejected or may contain errors due to failure at each stage, such as individual character recognition at 5 and knowledge processing at step 106. Therefore, this character recognition device determines whether or not a character includes a reject or an error (step 108), and, like a general character recognition device,
Providing the proofreading processor 27, for characters or character strings that could not be correctly recognized in the processing up to step 107, manually inputting the correct characters and character strings as proofreading processing and creating correct recognition results (Step 109). If it is determined in step 108 that neither a rejection nor an error is included, the process ends.

The calibration processing section 27 generally includes a display device (for example, a CRT or the like) for displaying an image read in step 101 and an intermediate result in each processing until output of a final result.
It is composed of an input device such as a keyboard for inputting correct characters and character strings. Then, the calibration processing unit 27
As a specific method of the proofreading process, there are generally a method of directly inputting a character or a character string read by a human or a method of selecting a correct one from a plurality of candidates indicated by a character recognition device. It is a target.

In the character recognition processing of the conventional character recognition apparatus, the processing ends when a correct recognition result is obtained, including manual correction (correction processing). When a correct result cannot be obtained by the recognition device alone and the correct result is input and corrected manually, the correct answer data input manually in the calibration processing unit 27 is fed back, and the individual A series of recognition processes from the character recognition unit 25 are performed (Step 110).

In the re-processing in step 110, in order to save a judgment error or rejection in the pre-processing, the probability of all possible pre-processing candidates is high using the pre-processing knowledge possessed at that stage. The processes up to the individual character recognizing unit 25 and the knowledge processing unit 26 are repeatedly performed in the order of the candidates considered until it is determined in step 111 that the correct answer given by the proofreading processing unit 27 has been obtained. Pre-processing candidates are
In each of the series of processes in steps 102 to 104, a plurality of candidate regions are given by assigning a priority (likelihood) to a region (or a divided image) to be extracted (detected) in each process. , Are the candidate areas listed above. That is, the (plural) regions extracted as candidates for the address description region, the (plural) regions extracted as character string region candidates, and the (multiple image) candidates extracted for one character region (divided image) Area). If only one area is extracted in each of the already executed steps 102 to 104, the pre-processing (steps 102 to 104) is performed again so that a plurality of candidate areas can be found after completion of the calibration processing in step 109.
Step 104) is executed. Alternatively, when the processing of step 110 is repeatedly performed, the preprocessing may be performed while changing the knowledge to be used for each single repetition.

As a result of these operations, if it can be determined in step 111 that the correct answer has been reached, it means that the pre-processing that has obtained the correct answer is the correct pre-processing for the processing target. The learning mechanism unit 28 according to each processing content in the detection unit 23 and the character cutout unit 24
Causes the preprocessing knowledge storage unit 22 to store new preprocessing knowledge corresponding to the correct preprocessing, or updates the knowledge in the preprocessing knowledge storage unit 22 so as to correspond to the correct preprocessing. That is, learning of knowledge used for preprocessing is performed.

The configuration and storage method of the knowledge stored in the preprocessing knowledge storage unit 22 depend on the contents of the existing preprocessing used. For example, a new format may be stored or parameters may be changed. You may. The selection criterion for the processing in each stage of the preprocessing may be changed.

Further, the preprocessing knowledge storage unit 22 sets an appropriate threshold value for each knowledge in order to avoid learning about an irregular format which often occurs in description of a free format but does not normally exist. You may make it hold | maintain. Alternatively, in order to avoid storing a processing object with a low frequency of occurrence as knowledge, if the knowledge to be learned is obtained for a plurality of processing objects, the structure may be reflected as preprocessing knowledge for the first time. Good.

When a similar processing target is given next time by the learning, a correct result can be obtained from the data in the preprocessing knowledge storage unit 22 having the learned new knowledge. By learning many patterns, it becomes possible to realize preprocessing optimized for the operation situation. Further, even when the preprocessing knowledge suitable for the processing target changes over a long period of time, it is possible to automatically automatically retain the optimal preprocessing knowledge.

Hereinafter, the present embodiment will be described with reference to actual examples.

FIG. 4 shows a case where the address and the destination name of the letter written on the sealed letter 10 have two areas in which the address and the destination name are considered to be described ("@ 123-4567 Adachi-ku, Tokyo ..." An area 15 and an area 16) of "@ Konan-ku, Yokohama-234-5678 ..." are shown. These areas 15, 16
One is an address writing area, and the other is an address or name writing area of the sender. In general, it can be determined which is the real addressing area by the description position in the sealed letter 10. However, depending on the case (especially in the case of direct mail), the address in the character string area detecting unit 23 may be determined. Depending on the description area extraction processing, it may be difficult to determine which area is the real destination name description area. Therefore, by applying the character recognition method of the present embodiment to learn which of the areas 15 and 16 is the real address description area as preprocessing knowledge, a similar sealed letter appears thereafter. In this case, the address description area can be detected without fail. In the case of direct mail in which a number of similar sealed letters appear in succession, learning is performed in the first one, and subsequent mails are quickly and without rejection or error based on learning in the first one. Processing.

The knowledge (parameter) for detecting the address description area includes the position of the corresponding area in the envelope (two-dimensional position from one vertex of the envelope as a reference point).
And size can be used.

FIG. 5 shows another example. Here, a case in which individual characters are cut out from an address written in handwritten characters will be described. (a) and (b) are each "Utsunomiya City"
(Utsunomiya City is the name of a city in Tochigi Prefecture) and "Hachioji City". According to the findings of the present inventors,
The four characters (kanji) that make up "Utsunomiya City" tend to be written in approximately the same size, while "Hachioji City"
With regard to, there is a tendency that the character of "king" is described smaller than other characters. Here, when performing the character cutout processing by the character cutout unit 24, assuming that all the characters are described in approximately the same size, "Utsunomiya City" can be cut out correctly, while "Hachioji" For the "city", an error occurs in the extraction processing, and for example, there is a possibility that "eight" and "king" are united and extracted as "all child city". According to the character recognition device of the present embodiment, if “Hachioji” appears and is rejected or an error occurs, “Hachioji” is input in the calibration processing, and the learning of the knowledge of the preprocessing knowledge storage unit 22 is performed. By doing so, when "Hachioji City" comes out thereafter, it will be possible to recognize correctly. If this character recognition device is used for sorting mail, there are many mails addressed to Hachioji city, such as the Tama area in Tokyo, even if the preprocessing knowledge for each region is not prepared in advance. In such an area, the efficiency of sorting mail items can be improved.

FIG. 6 shows a modification of the character recognition device shown in FIG. The character recognition device of the present invention does not need to be integrally configured with an image input unit or an image input device such as a scanner, and can perform character recognition based on image data read in another place in advance. it can. FIG. 6 shows such a character recognition device, which has a configuration in which an image storage unit 31 for storing a recognition target image is provided instead of the image input unit in the character recognition device shown in FIG.
The image storage unit 31 stores recognition target images via, for example, a network or a removable storage medium (magneto-optical disk), and the stored recognition target images are sent from the image storage unit 31 to the character string area detection unit. 23.

FIG. 7 shows another modification of the character recognition apparatus shown in FIG. In recent years, in particular, the number of mails, in particular, direct mail or the like, in which a large number of advertising words are written on the addressing face, has been increasing. In such mail, a large number of block-like areas that are regarded as a set of characters and character strings are extracted during the extraction processing of the address description area, and only knowledge processing using the position and size of the address description area as parameters is performed. In many cases, it is extremely difficult to identify which is the real addressing area. FIG. 8 is a diagram showing an example of such a mail as a character recognition processing target.
The rectangular areas shown by the broken lines in the drawing respectively represent block-like areas regarded as a set of characters and character strings.
Only one of the rectangular areas is a real destination description area.

Conventionally, mails as shown in FIG. 8 have to be manually sorted one by one, resulting in a significant decrease in work efficiency. Therefore, in the character recognition device shown in FIG. 7, the preprocessing knowledge storage unit 22 in the character recognition device shown in FIG.
Of these, the portion storing the written pattern is made independent, and is used as the pattern storage unit 32, so that the pattern storage unit 32 can update and additionally register the pattern. That is, with the arrangement of a plurality of regions regarded as a set of characters and character strings as a pattern, the pattern storage unit 32 stores the pattern and information indicating the destination addressing region (recognition target region) in the pattern. Remember.

In this character recognition apparatus, when extracting the address description area by the character string area detection unit 23, first, an area regarded as a character or a character string set is extracted from the recognition target image. When such a region is detected in a certain number or more, pattern matching between the arrangement of such a region and the pattern stored in the pattern storage unit 32 is performed to determine which pattern the recognition target image belongs to. Is read out from the pattern storage unit 32, and information on the address description area in the pattern is read out,
An addressing area is extracted from the block-shaped area in the recognition target image based on the information. As a result, even when there are many block-shaped areas that may be confused with the address description area, the address description area can be accurately extracted.

Further, in the character recognition apparatus shown in FIG. 7, the pattern storage section 32 receives addition or update of a new pattern by the learning mechanism section 28, similarly to the preprocessing knowledge storage section 22. In other words, when the final recognition result is rejected or erroneous due to the erroneous extraction of the address description area, if the correct address description area is specified by the calibration processing unit 27, the learning mechanism unit 28 stores the pattern based on it. A pattern is added to or updated in the unit 32. By learning the pattern in this way, even for direct mail where it is difficult to extract the address description area,
From the second copy onward, the address description area can be accurately extracted.

Next, a character recognition apparatus according to another embodiment of the present invention will be described with reference to FIG. This character recognition device has substantially the same configuration as that of the character recognition device shown in FIG. 1, but can be operated via a character string region detection unit 23 and a character cutout unit 24 in accordance with the result of input to the calibration processing unit 27. Instead, the preprocessing knowledge storage unit 2 is directly controlled by the learning mechanism unit 29.
2 is performed.

As described above, in the character recognition device, usually,
The data finally output by the knowledge processing unit 26 is the final processing result of the character recognition device. In some cases, a plurality of character string region candidates and character cutout combination candidates have already been obtained by this stage. In many cases, a free-form character recognition apparatus generally uses a method of obtaining a correct result by using each processing result as needed in a top-down process. That is, even if the output result is finally rejected or incorrect, the data to be input to the calibration processing unit 27 often includes a plurality of pieces of processing result information that can be candidates. The character recognition apparatus shown in FIG. 1 employs a method of obtaining correct preprocessing knowledge while reprocessing the entirety from an input image based on a character string input by the proofreading processing unit 27. The unit 26 can also provide the calibration processing unit 27 with a combination of a plurality of processing result candidates that could not be output as a final result and preprocessing information such as a format and a processing parameter corresponding thereto. Then, when the result input correctly by the calibration processing unit 27 is included in the candidates,
It is possible to know preprocessing knowledge to be learned from preprocessing information given according to the candidate without reprocessing the whole. Therefore, in the character recognition device shown in FIG. 9, when the result of the correct answer input in the proofreading processing unit 27 is included in the processing result candidate, the processing result candidate and the pre-processing information are forward By feeding back to the processing knowledge storage unit 22, learning of the preprocessing knowledge storage unit 22 is performed.

FIG. 10 is a flowchart for explaining the character recognition processing in the character recognition device shown in FIG.

As in the processing procedure shown in FIG. 3, after reading the image (step 121), extraction of the address description area (step 122), extraction of the character string area (step 123), and character extraction (step 121) are performed as preprocessing. 124), and then character recognition of individual characters is performed (step 1).
25), the recognition character string is determined by the knowledge processing (step 126), and the final result is output (step 12).
7), it is determined whether or not a reject or error is included (step 128). Here, if neither rejection nor error is included, the processing is terminated.

On the other hand, if it is determined in step 128 that a rejection or an error is included, the calibration processing unit 27 executes calibration processing by selecting from among processing result candidates (step 129), and performs learning. The mechanism unit 29 updates the knowledge in the preprocessing knowledge storage unit 22 according to the selected candidate (step 130), and ends the process.

Since the character recognition apparatus shown in FIG. 9 does not need to be reprocessed as a whole, the time required for learning is short, and it is suitable for learning in an operation state. However,
If the output information created as a candidate tends to be large and the possibility of a correct answer is not very low because of the originally given knowledge, it may not be learned no matter how much processing is performed.

While the character recognition apparatus according to the preferred embodiment of the present invention has been described above, each of the above-described character recognition apparatuses causes a computer such as a workstation or a personal computer to read a computer program for implementing the apparatus. It can also be realized by executing the program. A program for performing character recognition is read into a computer by a recording medium such as a magnetic tape or a CD-ROM. FIG. 12 is a block diagram illustrating a configuration of a computer that executes the above-described character recognition processing.

The computer includes an image input device 50 such as a scanner, a central processing unit (CPU) 51, a hard disk device 52 for storing programs and data,
Main memory 53 and input device 5 such as keyboard and mouse
4, a display device 55 such as a CRT, a magnetic tape or a CD
A reading device 56 for reading a recording medium 57 such as a ROM
It is composed of The image processing device 50, the hard disk device 52, the main memory 53, the input device 54, the display device 55, and the reading device 56
Connected to 1. In this computer, a recording medium 57 storing a program for performing character recognition processing is mounted on a reading device 56, the program is read from the recording medium 57 and stored in the hard disk device 52, and the program stored in the hard disk device 52 is read. Central processing unit 51
Executes the recognition target image captured via the image input device 50 or the hard disk device 5
The character recognition processing based on the above-described processing procedures is performed on the recognition target images stored in the storage device 2 or the like. The character recognition result is output from the central processing unit 51 for use in another device (not shown) (for example, a sorting device) or the like.

The present invention is not limited to the above embodiment. Character recognition method and apparatus of the present invention, character recognition other than character recognition for sorting mail, for example,
The present invention can also be applied to character recognition for processing various slips and character recognition for automatically recognizing and classifying character strings serving as keywords by reading various reports and in-house documents. In addition, the language (character type) for character recognition is not limited to Japanese (numbers, kana, kanji, etc.). For example, it is also possible to sort postal items whose names are written in the Roman alphabet such as English. The invention can be applied.

[0058]

As described above, the present invention processes a free-form recognition target, detects a desired character string candidate area from many character strings, patterns, and the like included in the target image, and performs character recognition. In the character recognition method and the character recognition device for performing character string recognition, learning of knowledge used in preprocessing is automatically performed by using data input in manual calibration processing, so that conventionally difficult and many man-hours are reduced. There is an effect that a required performance improvement method after the start of operation, particularly performance improvement in preprocessing, can be realized.

In particular, since learning is performed based on the actual recognition target, it is possible to improve the performance optimized for the processing target. Further, even when the preprocessing knowledge appropriate for the processing target changes over a long period of time, it is possible to automatically and always retain the optimum preprocessing knowledge. Furthermore, since the performance can be improved by optimizing the existing processing according to the operation situation, the performance can be improved without increasing the hardware or modifying the processing program.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a configuration of a character recognition device according to a first embodiment of the present invention.

FIG. 2 is a diagram illustrating an example of an image to be recognized.

FIG. 3 is a flowchart illustrating a procedure of a character recognition process using the character recognition device illustrated in FIG. 1;

FIG. 4 is a diagram illustrating an example of character recognition.

FIG. 5 is a diagram illustrating an example of character recognition.

FIG. 6 is a block diagram showing a modification of the character recognition device shown in FIG.

FIG. 7 is a block diagram showing another modification of the character recognition device shown in FIG.

FIG. 8 is a diagram illustrating a reading target having a large number of block-shaped regions.

FIG. 9 is a block diagram illustrating a configuration of a character recognition device according to a second embodiment of the present invention.

10 is a flowchart showing a procedure of a character recognition process using the character recognition device shown in FIG.

FIG. 11 is a block diagram showing a computer system suitably used in configuring the character recognition device of the present invention.

FIG. 12 is a block diagram showing a configuration of a conventional character recognition device.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 10 Envelope 11 Address 12 Source 13 Address description area 21 Image input unit 22 Preprocessing knowledge storage unit 23 Character string area detection unit 24 Character cutout unit 25 Individual character recognition unit 26 Knowledge processing unit 27 Calibration processing unit 28, 29 Learning Mechanism unit 31 Image file storage unit 32 Pattern storage unit

Claims (13)

[Claims] 1. A character recognition method for performing character recognition on a recognition target image in which characters are described in free format, wherein a block considered to be optimal as a character is cut out from the recognition target image as a divided image by knowledge processing. A pre-processing step, an individual character recognition step of performing individual character recognition on each of the divided images, and an appropriate character string by applying knowledge given in advance to the character string to a result of the individual character recognition. And a character string construction step of outputting a final character string as a final processing result, and a calibration processing step of inputting a correct character string when a rejection or an error is output as the final processing result. Then, using the correct character string as a key, knowledge for obtaining a correct preprocessing result is acquired, and the acquired Character recognition method characterized by having a learning step of updating the. 2. The pre-processing step includes: a step of extracting a recognition target area, which is an area to be subjected to character recognition, from the recognition target image by a knowledge processing; and a step of extracting each line from the recognition target area by the knowledge processing. Extracting a character string region corresponding to each of the character strings of the following, and obtaining the divided image by image-dividing the image of the character string region for each block considered to be optimal as a character by knowledge processing. The character recognition method according to claim 1, further comprising: 3. A plurality of preprocessing candidates are generated in the preprocessing step, and in the learning step, each of the preprocessing steps is performed until a character string that matches a correct character string input in the calibration processing step is constructed. Performing the individual character recognition step and the character string construction step based on the candidate, and then performing the preprocessing based on knowledge corresponding to a preprocessing candidate when a character string that matches the correct character string is constructed. 3. The character recognition method according to claim 1, wherein knowledge used in the process is updated. 4. A case in which a plurality of pieces of processing result information that can be candidates for a recognition result are output along with the final processing result, and one of the processing result information is selected as the correct character string in the calibration processing step. 3. The character recognition method according to claim 1, wherein in the learning step, knowledge used in the preprocessing step is updated based on preprocessing information corresponding to the selected processing result information. 5. The arrangement of a plurality of regions as candidates for the recognition target region is stored in advance as a pattern, and the pattern of the recognition target region is determined by pattern matching between the stored pattern and a region arrangement extracted from the recognition target image. The character recognition method according to claim 2, wherein extraction is performed. 6. A character recognition apparatus for performing character recognition on a recognition target image in which characters are described in a free format, comprising: preprocessing knowledge storage means for storing knowledge required for preprocessing of character recognition; From the image, using the knowledge previously given to the preprocessing knowledge storage means, a character string area detection means for detecting a desired character string area, using the knowledge stored in the preprocessing knowledge storage means,
A character cutout unit that divides the detected character string region image into blocks that are considered to be optimal as characters to obtain divided images, an individual character recognition unit that recognizes each divided image as a character, and an individual character recognition result. On the other hand, knowledge processing means for applying knowledge about a given character string in advance, constructing it as an appropriate character string, and outputting it as a final processing result, and calibrating when a rejection or error is output as the final processing result Calibration processing means, and when the calibration processing means performs the calibration processing, obtains knowledge for obtaining a correct preprocessing result using the correct data as a key, and stores the preprocessing knowledge storage based on the obtained knowledge. A character recognition apparatus characterized in that knowledge is accumulated in the means and the knowledge in the preprocessing knowledge storage means is updated. 7. The character recognition apparatus according to claim 6, further comprising image input means for taking a recognition target as binary or multi-valued image data and using the recognition target image as the recognition target image. 8. The character string region detecting means extracts a recognition target region, which is a region to be subjected to character recognition, from the recognition target image by knowledge processing, and converts the recognition target region into a character string of each line from the recognition target image. The character recognition device according to claim 6, wherein a corresponding character string region is extracted. 9. A method according to claim 9, further comprising a pattern storage unit for storing in advance a pattern of a plurality of regions to be candidates for the recognition target region as a pattern, wherein the character string region detection unit stores the pattern stored in the pattern storage unit and When the recognition target area is extracted by pattern matching with the area arrangement extracted from the recognition target image, and the calibration processing is performed by the calibration processing means, the correct data is used as a key to obtain a correct preprocessing result. 9. The character recognition apparatus according to claim 8, wherein a pattern in the pattern storage unit is updated based on the obtained pattern. 10. A recording medium readable by a computer, wherein a block considered to be optimal as a character is cut out as a divided image by a knowledge process from a recognition target image in which characters are described in a free format. A processing step, an individual character recognition step of performing individual character recognition on each of the divided images, and applying a given character string to a result of the individual character recognition to obtain an appropriate character string. A character string constructing step of constructing and outputting as a final processing result; a rewriting processing step of inputting a correct character string when a reject or error is output as the final processing result; and , Using the correct character string as a key, acquiring knowledge for obtaining a correct preprocessing result, and using the acquired knowledge in the preprocessing step. A storage medium storing a program for executing a learning step of updating knowledge to be used. 11. The pre-processing step includes: a step of extracting a recognition target area, which is an area to be subjected to character recognition, from the recognition target image by a knowledge processing; and a step of extracting each line from the recognition target area by the knowledge processing. Extracting a character string region corresponding to each of the character strings of the following, and obtaining the divided image by image-dividing the image of the character string region for each block considered to be optimal as a character by knowledge processing. The recording medium according to claim 10, comprising: 12. A plurality of preprocessing candidates are generated in the preprocessing step, and in the learning step, each of the preprocessing steps is performed until a character string that matches a correct character string input in the calibration processing step is constructed. Performing the individual character recognition step and the character string construction step based on the candidate, and then performing the preprocessing based on knowledge corresponding to a preprocessing candidate when a character string that matches the correct character string is constructed. The recording medium according to claim 10, wherein knowledge used in a process is updated. 13. A case in which a plurality of pieces of processing result information that can be candidates for a recognition result are output along with the final processing result, and one of the processing result information is selected as the correct character string in the calibration processing step. 12. The recording medium according to claim 10, wherein in the learning step, knowledge used in the preprocessing step is updated based on preprocessing information corresponding to the selected processing result information.