JPH06231307A - Optical character recognition system - Google Patents

Abstract

PURPOSE:To hold recognition precision and to shorten recognition time by using feature values in common by the two processings of character segment and matching. CONSTITUTION:A recognition processor generates a projection table 31 where a black picture element is projected in a direction vertical to a character string, namely, a longitudinal direction in an input image 30 stored in a picture memory and stores it in a work memory. Then, a projection table 51 where the black picture element is projected in a lateral direction is generated in a belt-form image 36 including only one extracted character and stores it in the work memory. Furthermore, the upper/lower positions 54 and 55 of the character are decided, a rectangular image (input pattern) 56 is segmented and is stored in the work memory. The input pattern 56 is matched with a recognition dictionary which is previously registered after the segment of the character completes. At that time, matching using the longitudinal/lateral projection tables 31 and 51 generated at the time of segmenting the character is executed, and it is applied for large classification or the extraction of a candidate character.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for performing character recognition at high speed and with high accuracy by commonly using feature quantity data once obtained in a plurality of processes in optical character recognition.

[0002]

2. Description of the Related Art Generally, in a character recognition device, characteristic amount data necessary for character extraction is obtained, and at the time of matching,
A new feature amount data is newly calculated, and a feature amount suitable for each process is calculated.

Further, since the matching process is performed with the original character rectangle without noise and edge error after the character cutting process, it is necessary to add a process for removing noise and the like.

For example, Japanese Patent Laid-Open No. 4-205690 proposes to use at least two recognition methods in order to improve the recognition rate.

[0005]

In a character recognition device that cuts out a specified character from image data and performs matching with a recognition dictionary, if there are many types of processing or there are many feature quantities calculated in each processing, the recognition rate will be high. Although improved, there is a problem that it takes a lot of processing time.

An object of the present invention is to maintain recognition accuracy,
It is to provide a processing method for performing character recognition at a higher speed.

[0007]

In order to achieve the above object, it is considered to reduce the feature amount calculated in each process.

According to the present invention, there is provided a character recognition processing method in which the table of the characteristic amount created for the character cutting processing is reused as a matching processing means.

In order to reduce the number of types of processing, when a character is cut out, noise can be removed at the same time by using the feature amount of the character. A processing method is provided in which a character is extracted in advance as noise or the like and a character is extracted.

[0010]

According to the present invention, since the feature amount is commonly used in the two processes of the character segmentation and the matching, it is possible to save the labor for creating the feature data dedicated to the matching, so that the processing time can be shortened by effectively utilizing the data. Is possible.

Further, according to the present invention, the character rectangle is extracted by excluding the portion having a small projection value, and the original character rectangle can be obtained without setting the noise removal processing again. Therefore, the processing time can be reduced. It is possible to maintain the recognition accuracy.

[0012]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described in detail below.

FIG. 1 is a flow chart of character recognition processing according to an embodiment of the present invention, FIG. 2 is a functional block diagram of an optical character recognition device to which the present invention is applied, FIG. 3 is an example of character cutout, and FIG. An example of matching is shown.

First, the outline of the character recognition processing will be described with reference to FIG. The optical character recognition device 10 inputs the input table 11 in which the image width, the image height, the slice level, etc. are set from the host device and stores it in the work memory 17 before the recognition process. Next, the control processor 16 activates the scanner 12 based on the input table 11 that has been input, reads an image by photoelectric conversion, and stores it in the image memory 13. After that, when the control processor 16 activates recognition, the recognition processor 14 performs a process of cutting out a circumscribed rectangular image of a specified character from the image data on the image memory 13, and inputs the cut-out image data to the work memory 17. Then, by matching it with the recognition dictionary 15, character recognition is performed, the recognition result is stored in the work memory 17, and the processing is ended.

Next, referring to FIG. 1, FIG. 2, FIG. 3 and FIG.
A procedure for performing character segmentation and matching using a common feature amount will be described in detail.

First, the recognition processor 14 creates a projection table 31 in which black pixels are projected in a direction perpendicular to a character string, that is, a vertical direction in the input image 30 (see FIG. 3) stored in the image memory 13. Work memory 1
7 (step 22 in FIG. 1). Since the presence of black pixels for each line can be known from the projection value, a certain threshold value 3
2 can be set, and the line 33 having a projection value higher than that can be regarded as a character portion. Thus, the change point 34 from the threshold value 32 or less to the threshold value 32 or more of the projection value is the left position of the character, and the change point 35 from the threshold value 32 or more to the threshold value 32 or less is the right position of the character. It is possible to determine and extract the band-shaped image 36 that circumscribes the left and right of the character (step 23 in FIG. 1). Next, in the extracted strip-shaped image 36 containing only one character, a projection table 51 projected in the horizontal direction is created and stored in the work memory 17.
Based on this, the vertical positions 54 and 55 of the character are determined (step 24 in FIG. 1), and a rectangular image (input pattern) 56 is cut out so as to circumscribe the upper and lower sides of the character, as in the case of determining the horizontal position. To store. (FIG. 1, step 25). By the above processing, the circumscribed rectangle of the character is extracted, and the character segmentation (20) is completed.

Thereafter, the input pattern 56 is matched with the recognition dictionary 15 registered in advance. At that time, the vertical / horizontal projection table 3 created when the characters were cut out
Matching using 1 and 51 is performed to perform a large classification (step 26 in FIG. 1) or extraction of candidate characters (step 2 in FIG. 1).
Apply to 7) etc. This saves the trouble of preparing a new feature amount at the time of matching.

Next, a matching method using the vertical / horizontal projection tables 31 and 51 will be described with reference to FIG.
As the feature quantity, the portions judged to be characters in the character segmentation of the vertical projection table 31 and the horizontal projection table 51, that is, the portions 33 and 53 of the lines whose projection values are equal to or greater than the threshold value are used. And the character rectangle image
(Input pattern) 56 and dictionary character image (standard pattern)
In order to obtain the similarity with 46, the projection tables 31, 5
1, the distance calculation shown in Expression 41 is performed, and the one with a short distance is considered to have a high degree of similarity.

Distance = Σ | input pattern (i) -standard pattern (i) | (Equation 41) Then, for example, the large classification is performed according to the similarity of the vertical projection table 31 (step 26 in FIG. 1), and the large classification is performed. Next, the similarity of the horizontal projection table 51 is obtained for the characters in the created dictionary, and the candidate characters are further narrowed down (step 27 in FIG. 1). After that, for example, the right edge center position of the input pattern and the right edge center position of the standard pattern are aligned and overlapped, and one character is extracted by a pattern matching method or the like that calculates the degree of similarity by counting the number of matching black pixel dots. It is decided (step 28 in FIG. 1). This completes recognition of one character
(21).

If the above-mentioned method is adopted, it is possible to effectively use the feature amount used in the character extraction even in matching.

On the other hand, when the character is cut out, if the threshold value is set to the minimum projection values 32 and 52 that can be regarded as non-noise portions, the projection values 37, 5 where only a few dots appear.
Since 7 can be removed in advance and taken out, the noise 38a contacting the character and the error portion 38b due to the binarization of the character edge can be simultaneously removed. Therefore, it is not necessary to perform the noise removal process again at the time of recognition, and there is an effect that the processing time can be further shortened while maintaining the accuracy.

[0022]

According to the present invention, in an optical character recognition device for recognizing characters in a form image, the recognition time can be shortened while maintaining the recognition accuracy.

[0023]

[Brief description of drawings]

FIG. 1 is a diagram showing a flow of character recognition processing according to an embodiment of the present invention.

FIG. 2 is a functional block diagram of an optical character recognition device to which the present invention is applied.

FIG. 3 is a diagram illustrating a process of cutting out a character from a form image image in units of one character.

FIG. 4 is a diagram showing a process of performing matching using a vertical / horizontal projection table.

[Explanation of symbols]

 10 ………… Optical character recognition device 12 ………… Scanner 13 ………… Image memory 14 ………… Recognition processor 15 ………… Recognition dictionary 16 ………… Control processor 17 ………… Work Memory 31 ………… Vertical projection table 51 ………… Horizontal projection table 38a ………… Noise in contact with characters 38b ………… Error due to binarization of character edges

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Hiroaki Haneda No. 1 Ikegami, Haruoka-cho, Owariasahi-shi, Aichi Hitachi Asahi Electronics Co., Ltd.

Claims (2)

[Claims] 1. An optical character recognition method in which image data of a form is stored in an image memory and a specified character of the image data stored in the image memory is cut out for recognition, and a feature amount data created at the time of character cutting processing is used. -The optical character recognition method characterized in that the processing time until recognition is shortened by reusing the data in the character recognition processing to reduce the number of times the feature amount is calculated. 2. The optical character recognition method according to claim 1, wherein at the same time as character extraction, an error associated with binarization of a character edge and noise that comes into contact with a character are removed by using the feature amount data. Characteristic optical character recognition method.