JPH0765124A - Character recognition device - Google Patents

Abstract

PURPOSE:To provide the character recognition device which can evade a scan on the whole input image for character detection and segmentation before recognition and can recognize complicated character detection and segmentation candidates. CONSTITUTION:A memory 21 is stored with a binary image of the entire surface of postal matter which is divided, line by line. An MPU 22 traces an outline from the 1st line of the memory 21 and writes structure feature information on the coordinate values, the number of black dots, and respective outline points of a block connected in black each time the block is found. An MPU 24 performs a recognition processing by matching arithmetic between the structure feature information in the memory 23 and dictionary information of a dictionary 26 and writes the recognition result in the memory 23. An MPU 25 performs a recognition processing by composite similarity arithmetic, based upon the block coordinate values in the memory 23, between the character image in the memory 21 and the dictionary information in a dictionary 27 and writes the recognition result in the memory 23. An MPU 28 stops the MPUs 22, 24, and 25 from operating when a limit time is elapsed, and determines a final recognition result uniquely by referring to a section address data base 29.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention is, for example, a mail address automatic reading / sorting device that optically reads address information on a mail item to classify the mail item. The present invention relates to a character recognition device that recognizes characters of address information.

[0002]

2. Description of the Related Art Generally, for example, in a character recognition device for recognizing a character of address information written by handwriting on a postal matter, redundant character detection cutout candidates are generated by image processing. Then, the character detection / cutout candidate is subjected to recognition processing by pattern recognition by the composite similarity method or structure analysis.

By the way, the number of character detection / cutout candidates increases in proportion to the degree of freedom of character description of an input image. In the case of recognition of free handwritten characters, the character detection / cutout candidates become very redundant, which imposes a time burden on the character recognition. In addition, holding the features calculated in the process of character recognition is a heavy burden.

Further, since the character detection / cutout processing is performed by scanning the entire input image, it is necessary to scan the entire image at least once before starting character recognition. However, the input image is redundant, and a lot of time is allotted to unnecessary image processing. Furthermore, image processing for the entire input image is
When there are many unnecessary areas, redundant character detection cutout candidates are generated that are difficult to recognize.

[0005]

Conventionally, in the case of character detection / cutout, a very redundant character detection / cutout candidate is generated when there are many unnecessary areas because the entire area of the input image is scanned. Even when the recognition process is performed while the character detection / cutout process is performed, the entire image needs to be scanned before the recognition process because of the labeling process necessary for the recognition.
Scanning the entire image means that part of the character detection / cutout processing is performed.

Conventionally, since the character detection / cutout processing and the recognition processing are performed independently, the time cost increases and the character detection / cutout candidates frequently occur. At the time of character detection cutout, since the concept of the recognition target object does not exist, character detection cutout by recognition is also proposed, but there is preprocessing for character detection cutout.

Therefore, according to the present invention, it is possible to avoid scanning the entire input image for character detection cutout before recognition, and it is also possible to recognize a complicated character detection cutout candidate that cannot be expressed by a simple polygon. An object of the present invention is to provide a character recognition device.

[0008]

A character recognition apparatus according to the present invention comprises an image inputting means for inputting an image in which characters are present on an object to be read, and successive contour tracing for the image input by the image inputting means. And a structure feature extracting unit that extracts the structure feature of the image by performing the above operation, and the structure feature extracting unit operates in parallel. When the structure feature extracting unit extracts the structure feature, the structure feature is used. A recognition means for performing recognition processing by performing feature matching calculation is provided.

Further, the character recognition apparatus of the present invention comprises an image input means for inputting an image in which a character is present on an object to be read, and a contour tracing for the image input by the image input means. Structural feature extraction means for extracting structural features of an image, and the structural feature extraction means operate in parallel,
And a recognition unit that performs a recognition process by performing a composite similarity calculation using the image input by the image input unit based on the structure feature when the structure feature extraction unit extracts the structure feature. ing.

Further, the character recognition device of the present invention comprises image input means for inputting an image in which characters on the object to be read exist.
Structural feature extracting means for extracting structural features of the image by sequentially performing contour tracing on the image input by the image inputting means, and operating in parallel with the structural feature extracting means. Once the structural features are extracted,
When the structure feature extraction unit operates in parallel with the first recognition unit that performs the recognition process by performing the structure feature matching operation using the structure feature, and the structure feature extraction unit extracts the structure feature. , Based on its structural features,
Second recognition means for performing recognition processing by performing composite similarity calculation using the image input by the image input means, recognition result of the first recognition means and recognition result of the second recognition means And means for outputting one final recognition result.

[0011]

Since the recognition target object is included in the image held at the image processing stage, the character detection cut-out processing and the recognition processing can be simultaneously performed by directly applying the recognition target concept. It is possible to avoid scanning the entire input image for character detection clipping before recognition.

Further, a part of the recognition processing also serves as character detection / cutout processing, and the target image is completely captured at the stage of character detection / cutout, so that a complicated character detection / cutout candidate that cannot be expressed by a simple polygon is used. Recognition is also possible.

[0013]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings. FIG. 7 schematically shows the configuration of a mailing address automatic reading / sorting machine to which this embodiment is applied. In other words, this automatic mail address sorter and sorter has a supply unit 1 in which a large number of postal products P are set in a batch and standing state, and one postal product P set in the supply unit 1 in order from the front. A take-out section 2 for taking out, a take-in conveyance path 3 for conveying the postal matter P taken out by the take-out section 2, and a reading section 4 for optically reading address information on the postal matter P conveyed by the take-in conveyance path 3 ,
The reading unit 4 is configured by a sorting unit 5 that sorts the postal matter P whose address information has been read based on the reading result.

The sorting section 5 is for the postal matter P that has passed through the reading section 4.
And a plurality of pockets (section boxes) 8 provided along each of the segmented transport paths 7a to 7g. ing.

An operation panel 9 operated by an operator is provided above the supply unit 1. Then, by pressing the start button after designating the reading mode on the operation panel 9, the postal matter P set in the feeding section 1 is taken out one by one by the take-out section 2, and the take-in conveyance path 3 Is conveyed to the reading unit 4. Reading unit 4
Optically reads the address information on the conveyed postal matter P, and the sorting unit 5 operates based on the read result, whereby the sorting information is sorted and accumulated in the pocket 8 corresponding to the sorting designation data. There is.

FIG. 6 schematically shows the structure of the reading unit 4. That is, the photoelectric conversion unit 11 as an image input unit photoelectrically converts the entire image on the mail P.
For example, it is configured by a light source that illuminates the mail P, a self-scanning CCD image sensor that receives reflected light from the mail P and converts it into an electric signal.

The output of the photoelectric conversion unit 11 is sent to the address image processing unit 12. In the address image processing unit 12, first, an address area is detected from the input image from the photoelectric conversion unit 11, a line area is detected from the detected address area, and characters are detected and cut out from the detected line area. The character detection / cutout candidate is sent to the address character recognition unit 13.

The address character recognition unit 13 performs address recognition processing according to the character detection cutout candidates from the address image processing unit 12, and sends the recognition result to the number recognition unit 14. The number recognition unit 14 performs number recognition processing while detecting and cutting out the image area following the address information, evaluates and integrates the recognition results of the address and the number, and finally determines the recognition result uniquely, and the recognition result ( Answer) to the first CPU (Central
To the processing unit) 15.

The first CPU 15 controls the entire reading unit 4, and transfers the recognition result from the numeral recognition unit 14 to the second CPU 16. The second CPU 16 controls the entire sorting machine, and controls the drive unit 17 based on the recognition result to sort and collect the postal matter P in a specific pocket 8.

FIG. 1 shows the configuration of the numeral recognition unit 14 in detail. That is, the binary image of the entire surface of the postal matter P divided line by line is transferred from the address image processing unit 12 to the shared image memory 21. Contour tracking MPU (micro
The processor unit) 22 performs contour tracing from the first line of the image in the shared image memory 21 to find each block connected in black, the coordinate value and the number of black points in the shared image memory 21. , And writes structural feature information such as each contour point in the shared memory 23.

There are a plurality of structure matching MPUs 24 and a plurality of composite similarity MPUs 25, respectively, and when structural feature information is written in the shared memory 23, recognition processing is started. The structure matching MPU 24 performs recognition processing by using the structure feature information in the shared memory 23 and matching operation with the structure dictionary information in the structure dictionary 26, and writes the recognition result in the shared memory 23.

The composite similarity MPU 25 combines the character image in the shared image memory 21 with the dictionary information in the composite similarity dictionary 27 according to the coordinate values of the blocks connected in black in the shared memory 23. The recognition process is performed by calculating the similarity, and the recognition result is stored in the shared memory 2.
Write to 3.

The control MPU 28 constantly monitors the shared memory 23 and limits the recognition that does not lead to the recognition start point. When the time limit has passed, the contour tracking MPU 22, the structure matching MPU 24, and the composite similarity MPU 25 are stopped, and the block address database 29 is stopped.
The final recognition result (answer) is uniquely determined by referring to. The determined recognition result is transferred to the CPU 30 that controls the entire number recognition unit. In CPU30,
The received recognition result is transferred to the outside as the recognition result of the numeral recognition unit 14.

FIG. 2 shows a part of the numeral recognition unit 14, and the numeral recognition unit 14 will be described in more detail below with reference to this drawing. The shared image memory 21 stores the line-detected and cut-out image. The contour tracking MPU 22 starts the process from the line c (the number portion indicating the block information) in the shared image memory 21 according to the recognition result of the address character recognition unit 13. The contour tracking by the contour tracking MPU 22 is performed by raster scanning with the Y axis of the image as the main scan.
Then, when a black block is detected, its contour is tracked,
The structural characteristic information is written in the shared memory 23.

The structure matching MPU 24 and the composite similarity MPU 25 constantly monitor the shared memory 23,
When the structure tracking information is newly written by the contour tracking MPU 22, the processing is started. There is a message queue in the shared memory 23, and the structure matching MPU 24 and the composite similarity MPU 25 poll the head of the queue.

The contour tracking MPU 22 sets a flag in the message queue at the time when the calculation of the contour feature of the black block is completed, and notifies that there is a pattern that can be processed. The structure matching MPU 24 and the composite similarity MPU 25 are
The flag is used to obtain information on a pattern that can be processed from the shared memory 23 and the recognition process is started.

The structure matching MPU 24 and the composite similarity MPU 25 poll the message queue again after the recognition process is completed. Structure matching MPU24
Reads the structure characteristic information from the shared memory 23, further calculates the contour characteristic information in which each contour point is approximated by an arc, performs the matching calculation with the dictionary information of the structure dictionary 26, and writes the result to the shared memory 23.

The composite similarity MPU 25 is shared by the shared memory 23.
Only the corresponding image is acquired from the shared image memory 21 by referring to the coordinate values of the black-connected blocks of the image inside. The image thus obtained is subjected to processing such as normalization and sampling, and is then subjected to composite similarity calculation with the dictionary information of the composite similarity dictionary 27, and the result is written in the shared memory 23.

FIG. 3 shows an example of character detection cutout candidates in the numeral recognition unit 14. The content of the shared memory 23 is added each time the contour tracking MPU 22 detects a black block. The recognition unit 1, the recognition unit 2, the recognition unit 3, and the recognition unit 4 are composed of a structure matching MPU 24 and a composite similarity MPU 25, respectively. Each recognition unit has a shared memory 23.
Each time a is added, character detection cutout is performed and recognized in the range of 1 block to 4 blocks.

At P1 in the figure, since the shared memory 23 has only one block, only the recognition unit 1 operates.
In P2 in the figure, one block is added to the shared memory 23 to make two blocks. Therefore, the recognition unit 2 recognizes the pattern of two blocks including the newly added block, and the block is newly added. Only the block is recognized by the recognition unit 3. In this way, the recognition units 1, 2, 3, and 4 are responsible for different numbers of blocks, and the recognition processing is performed with a maximum of 4 blocks.

The control MPU 28 searches for a recognition result having high recognition accuracy when the recognition process is completed. When the candidate A having high recognition accuracy is selected, other candidates B including the candidate A are excluded. Of the remaining candidates, the unrecognizable candidate C is excluded, and for the recognized candidates that have a plurality of character detection cutout candidates, the recognition result is re-evaluated to uniquely determine the answer.

FIG. 4 shows an example of answer editing in the numeral recognition unit 14. In the figure, the thick character in the frame indicates the recognition target, and the character in () next to it indicates the recognition result. Further, the shaded frame portion is an arc of a graph that is discarded by adopting a candidate with high recognition accuracy. The upper left number in the frame is the arc number.

The recognition candidates are represented by a graph, and the recognition results of the recognition units 1, 2, 3, 4 correspond to arcs on the graph. Graphs typically generate redundant arcs from one node to include redundant candidates. If an arc (arc number is 5) D that can be recognized with high recognition accuracy is found,
Arc number is 4,6,7,9,10,11,12,15
Arcs are removed from the graph.

Similarly, when an arc (arc number 22) E that can be recognized with high recognition accuracy is found, the arc number is 19, 20, 21, 23, 24, 25, 27, 2
The 8,31 arcs are deleted from the graph. Then, the arc F having arc numbers 13 and 16 in which no recognition result exists is deleted from the graph.

As shown in FIG. 5, the extruded graph becomes a final recognition candidate, and the recognition result "2-18" is obtained by searching the graph by a predetermined procedure.
-3 "can be obtained.

In FIG. 2, the structure matching MP
If the recognition results of the U24 and the composite similarity MPU25 are different, or if multiple positives occur in the recognition processing of the structure matching MPU24, the answer is uniquely determined by a predetermined procedure. Is often not uniquely determined.

Since the shared memory 23 holds all the structural feature information of the recognition target area, new information can be obtained by performing a more detailed matching calculation of the necessary candidate structural features. In this way, the shared memory 2
By holding all the structural feature information of the recognition target area in 3, it is possible to perform high-speed feedback when the recognition result cannot be uniquely determined.

As described above, according to the above embodiment,
The following effects can be expected. That is, conventionally, in character recognition, the entire input image is scanned to perform character detection cutout, and then the recognition process is performed. However, since the recognition target object is included in the image held at the image processing stage, the character detection cutout processing and the recognition processing can be performed simultaneously by directly acting the recognition target concept. It is possible to avoid scanning the entire input image for the previous character detection crop.

Further, in the past, since the coordinate designation was rectangular or polygonal, complicated character detection cutout could not be designated. However, in the present embodiment, part of the recognition processing is character detection cutout, and since the target image is completely captured at the stage of character detection cutout, recognition of a complicated character detection cutout candidate that cannot be expressed by a simple polygon. Processing is also possible.

Further, conventionally, all the original images were processed, but by retaining the structural features of the original images as in the present embodiment, it is possible to feed back when the answer is uniquely determined, and the whole image is fed back. See the structural features for.

[0041]

As described above in detail, according to the present invention, it is possible to avoid scanning the entire input image for character detection and clipping before recognition, and moreover, to complicate a complicated character that cannot be represented by a simple polygon. It is possible to provide a character recognition device that can also recognize a detection cutout candidate.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of a numeral recognition unit.

FIG. 2 is a block diagram showing a part of a numeral recognition unit.

FIG. 3 is a diagram showing an example of character detection cutout candidates in a number recognition unit.

FIG. 4 is a diagram showing an example of answer editing in a number recognition unit.

FIG. 5 is a diagram showing a graph of final recognition candidates.

FIG. 6 is a block diagram schematically showing the configuration of a reading unit.

FIG. 7 is a configuration diagram schematically showing a configuration of a mail address automatic reading / sorting machine.

[Explanation of symbols]

P: postal matter (object to be read), 4 ... reading section, 11 ... photoelectric conversion section (image input means), 12 ... address image processing section,
13: Address character recognition unit, 14: Number recognition unit, 21 ...
… Shared image memory, 22 …… Contour tracking MPU, 23 ……
Shared memory, 24 …… Structure mating MPU, 25 ……
Compound similarity MPU, 26 ... Structural dictionary, 27 ... Compound similarity dictionary, 28 ... Control MPU, 29 ... Block address database.

Claims (3)

[Claims] 1. An image input means for inputting an image in which characters are present on an object to be read, and a structure for extracting structural features of the image by successively performing contour tracing on the image input by the image input means. Feature recognition means operates in parallel with the structure feature extraction means, and when the structure feature extraction means extracts a structure feature, a recognition process is performed by performing a structure feature matching operation using the structure feature. A character recognition device comprising means and. 2. An image input means for inputting an image in which characters are present on an object to be read, and a structure for extracting structural features of the image by successively performing contour tracing on the image input by the image input means. The feature extracting means operates in parallel with the structure feature extracting means, and when the structure feature is extracted by the structure feature extracting means, based on the structure feature, a composite image is obtained using the image input by the image input means. A character recognition device comprising: a recognition unit that performs recognition processing by performing similarity calculation. 3. An image input means for inputting an image in which characters are present on an object to be read, and a structure for extracting structural features of the image by successively performing contour tracing on the image input by the image input means. The feature extracting means operates in parallel with the structure feature extracting means, and when the structure feature is extracted by the structure feature extracting means, recognition processing is performed by performing a structure feature matching operation using the structure feature. When the structural feature extracting unit operates in parallel with the recognizing unit No. 1 and the structural feature extracting unit and the structural feature extracting unit extracts the structural feature, the image input by the image input unit is used based on the structural feature. A second recognition unit that performs recognition processing by performing a composite similarity calculation, and a final recognition result by the recognition result of the first recognition unit and the recognition result of the second recognition unit. Character recognition apparatus characterized by comprising means for force, the.