JPH03100786A - Method and device for character recognition - Google Patents

Abstract

PURPOSE:To shorten the recognition time by performing the integration processing of value '0' or '1' for each two-dimensional address to obtain the frequency distribution with respect to plural normalized dictionary characters and using a specific address as the information extraction position for character recognition. CONSTITUTION:Binarized normalized pictures 220 having a prescribed size are generated for all dictionary characters 200 as the recognition object, and information '0' or '1' of all dictionary character patterns corresponding to respective addresses on normalized pictures are integrated by a processing 300 to obtain two-dimensional frequency distribution 310. A prescribed number of two-dimensional addresses are extracted from addresses of the most largest quantity of information based on the frequency distribution 310, and information '0' and '1' of normalized pictures obtained on these two-dimensional addresses for inputted recognition object characters are taken as input data of a character recognizing means. Thus, the processing time required for character recognition is shortened in comparison with the method where all of (MXM)-number of picture element information are used as information for character recognition.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a method for automatically recognizing fixed-form characters, and relates to a method and apparatus for recognizing characters such as numbers on car license plates and characters printed on cardboard.

[Conventional technology]

In the conventional character recognition method for fixed-form characters, Introduction to Character Recognition (Published by Telecommunications Association, edited by Shinichi Hashimoto).

As described in Printed Character Recognition Technology (1982), the pattern matching method, stroke analysis method 1 contour extraction method, etc. are known.

A typical pattern matching method is a template matching method. This method is a technique in which a dictionary character pattern (template) to be recognized is prepared and the degree of matching is determined by superimposing it on the recognition target character pattern.

The stroke analysis method is a method of checking whether a group of black dots exists within a designated area from a binary pattern in which black dots are distributed. Contour feature extraction methods include a method of extracting features while tracing the outline of a character, and a method of extracting features from the shape of the outline of a single character.

[Problem to be solved by the invention]

Conventional character recognition methods have a problem in that recognition processing such as preprocessing and matching for recognition takes time when the number of characters to be recognized is large in any process.

An object of the present invention is to provide a character recognition method that shortens recognition time, including preprocessing such as matching processing with dictionary character patterns or calculating feature quantities such as center of gravity and area of characters to be recognized each time. It is in.

[Means to solve the problem]

In order to achieve the above object, the character recognition method according to the present invention is a character recognition method that recognizes and processes fixed-form characters. Create and normalize 1″ of all dictionary character patterns corresponding to each address on the image plane
A two-dimensional frequency distribution is obtained by integrating 0'' or 1'' information, and based on this frequency distribution, a predetermined number of two-dimensional addresses are extracted from the address with the largest amount of information, and the two-dimensional OI of normalized image obtained on address
The I11171 information is configured to be input data to the character recognition means.

A predetermined number of two-dimensional addresses are extracted from addresses that have a value closest to the frequency of 1/2 of the total number of characters in the dictionary.

Further, the character recognition means is formed by a neural network consisting of three layers: an input layer, a middle layer, and an output layer, and the weighting coefficient of the middle layer for recognizing characters to be recognized is determined in advance by learning dictionary character patterns. It is.

Furthermore, the weighting coefficient of the intermediate layer is determined by increasing the number of inputs to the input layer of the neural network by at least 1 based on the two-dimensional frequency distribution data when recognition fails a predetermined number of times for inputting the recognition target character. The structure is designed to allow learning.

The recognition target character is part of a known recognition target character string, and when recognition of the recognition target character fails, a known character that matches the first candidate character or the second candidate character of the recognition target character is selected. This is the configuration for the recognition result.

Alternatively, when automatic character recognition fails, the character to be recognized may be displayed and recognized by a person, and the recognition results may be used to relearn the weighting coefficients of the intermediate layer of the neural network.

Furthermore, the character recognition device to which the character recognition method is applied includes an imaging unit that images the recognition target character pattern and the dictionary character pattern, and an A/D that converts the imaged signal from A/D.
Consists of a conversion unit, an image memory section for A/D converted images, an image processing/recognition section for the images, a dictionary character memory section and an information memory section regarding characters to be recognized, and a display device for displaying character recognition results. It is the composition.

[Effect]

According to the character recognition method of the present invention, in the binarization of the dictionary character pattern and the recognition target character pattern, in a character image obtained by an imaging means such as a television camera, the brightness of the character part and the background part is always constant. Since the brightness of the text and background is standardized, the information is compressed at the same time. Further, character normalization is performed to allow flexible response to changes in character size obtained by an imaging device. Furthermore, the two-dimensional degree distribution obtained by integrating the information of "II O" or "1" obtained as the dictionary character pattern of the recognition target character for all dictionary characters is calculated based on the two-dimensional image address of the normalized size. This serves as an index for determining which addresses have significant information for character recognition.In this way, information only for specific pixel addresses determined on the character pattern image is extracted.

Furthermore, by applying a neural network as a character recognition means, by completing learning on dictionary character patterns in advance, in the character recognition stage, "" on a predetermined specific two-dimensional image address 07'1
By inputting the 411j data to the input layer of the 21-ral net, characters are recognized by executing only the product-sum operation with the weighting coefficient in the intermediate layer.

〔Example〕

An embodiment of the present invention will be described with reference to FIG. In FIG. 1, a grayscale image of dictionary characters 200, a dictionary character 210 after binarization, a dictionary character 220 after normalization, . Process for determining the frequency distribution on the normalized two-dimensional image address for the dictionary character pattern of all recognition target characters 300. The resulting frequency distribution 310 is shown.

The binarized and normalized dictionary character pattern 220 can be created, for example, pixel by pixel on a computer memory using an external data input means such as a keyboard. Once input into the image memory by the imaging means as a gray scale image (in the case of 8 bits, it can be expressed in 0 to 255 gradations), it is converted to 1 (Qlj, tt1'' by appropriate binarization processing).
'Convert into information. This binarization process is shown in Figures 4 and 5.
This can be realized by the method shown in the figure. In the method shown in FIG. 4, when the frequency distribution of the grayscale values of a grayscale image in a predetermined region including the character to be binarized is determined, the distribution is as shown in the figure on the right. In this shading frequency distribution, the smaller peak A of the shading level g corresponds to the character part, and the peak B of the larger shading level g to the right of it corresponds to the bright part corresponding to the background. . Therefore, in the binarization process, the algorithm of Equation (1) can be used with the gray level of go, which is the valley between the two peaks, as the threshold.

Here, gt: gradation level of each pixel in the binarized region However, as in the gradation frequency distribution shown in FIG. 4, the two peaks A and B may not clearly occur. In such a case, as shown in FIG.
A method may also be adopted in which the minimum value g ll1in is determined and the median value thereof is used as the threshold value.

Character normalization is an essential process for determining the position of an image within a character to be extracted as input information for single character recognition.The character size before normalization is defined as Nx pixels in the X direction and Ny pixels in the Y direction. Then, the font size after normalization is set to MXM
1. Normally, M is selected as the minimum value that can represent the character pattern as an image, so it is only necessary to reduce it by the magnification of M / N x and M / N y in both the X and Y directions. For each two-dimensional address of n dictionary characters normalized in this way, ago
A frequency distribution 310 is obtained by integrating the values of "or 1". This frequency distribution 310 can be said to indicate at which position on the normalized two-dimensional image there is a large amount of information as a character pattern. Therefore, if specific m addresses among these MXM pixel addresses are used as information extraction positions for character recognition, the character The processing time required for recognition can be reduced. The method for selecting these specific m addresses is that if there are n characters to be recognized, they should be extracted sequentially starting from the address that gives the value closest to the frequency of n/2 in the frequency distribution of the dictionary character pattern. It is reasonable.

The reason for this is that an address that is always "0'" or "1" for any character pattern can be regarded as a position where there is no information about a single character pattern. Considering this point, the information of “NO” and 11171 is 5
Addresses that occur with a probability of 0% contain the most information. FIG. 2 shows a concrete flow of the embodiment of FIG. 1, which has been described at least once, corresponding to the point where such an address gives a frequency of n/2.

In Figure 2, ph(x*j) is a binary character pattern after normalization processing, and h(up J) is "On a 1" of n dictionary character patterns from 1 to n.
It represents the frequency distribution obtained by integrating information for each two-dimensional address (x, j). Processing block 50 is a process for initializing a variable h(i,j) for determining the frequency distribution for all dictionary character patterns Pk(x,j) (k=1 to n).

FIG. 3 shows an embodiment in which the present invention is applied to character recognition using a neural network. In this figure, an imaging device 400 for capturing images of dictionary character patterns and recognition target character patterns, and a dictionary character image capture processing unit 41
0. Binarization processing unit 420 for character images captured by the imaging device 400. Normalization processing unit 43 for binary character patterns.
0. # O # # or 111 for all dictionary character patterns
Frequency distribution calculation unit 44 obtained by IT integration processing
0, a processing unit 450 that sequentially extracts m two-dimensional addresses from the frequency closest to n/2 based on the frequency distribution from the frequency distribution calculation unit 440; a recognition target character image capture processing unit 460; Processing unit 4 that extracts “0” or #1 jj information corresponding to m addresses from the normalized dictionary character pattern or recognition target character pattern
70, a learning processing unit 480 of a weighting coefficient W1□ (WIJ represents a weighting coefficient of a connection between neurons 11) of an intermediate layer that connects the input layer and output layer of a neural network, and a neural network 490 for character recognition are shown. .

Next, the processing contents of this embodiment will be explained in detail.

Regarding the method of determining m information extraction addresses on the dictionary character pattern and the recognition target character pattern, see Figures 1 and 2.
Since the process has been explained using the flow shown in the figure, the explanation will be omitted here.

The neural network shown in this embodiment simulates the function of the human brain, and is basically composed of an input layer, a middle layer, and an output layer. The intermediate layer generally has a multilayer structure, and each layer is connected using a weighting coefficient representing the strength of connection. In the case of humans, the results of learning over a long period of time for various patterns of information in the input layer are retained as weight coefficients in the middle layer, and the relationship between a certain input information pattern and output pattern is The neural network of this embodiment is for realizing the function of so-called character recognition, in which character patterns entering the input layer are associated with dictionary character patterns in the output layer.

First, a method of learning the weighting coefficient W ta of the intermediate layer of the neural network and a character recognition method using Wl- determined as a result of the learning will be described.

The learning of the weighting coefficient Wi J corresponding to the intermediate layer is carried out by learning the weighting coefficient Wi ``Data is given to m inputs of the neural network, and the weighting coefficient W t J is sequentially learned from the resulting response 0 in the output layer using equation (2).

ΔW I J (k÷1) = -ηδJO amount + αΔWIJ
(k)... (2) where δJ: error in the output layer (difference between the teaching pattern and the response pattern) η, α: learning parameters: repeated steps of learning. This continues until the output corresponding to the dictionary character is obtained for the 0"1" pattern caused by the input. This learning process is executed for n dictionary characters to learn the weighting coefficient Wi- of the intermediate layer.Normally, the value of the weighting coefficient WIJ needs to be set appropriately before learning, but the value is There is a method to set it using random numbers. When the configuration of the neural network is completed in this way, after capturing an image of characters to be recognized by the imaging means 400, after passing through a binarization process 420 and a normalization process 430, m characters extracted from this image are 0”1
``Characters can be recognized from the values in the output layer obtained by inputting patterns into the input layer of a neural network.

In this case, the recognition result is the character corresponding to the response O1 that outputs the largest value among the n responses of the output layer.

Here, a recognition method using a neural network will be explained using a specific example.

In the case of character recognition, the number of responses 01 in the output layer is the same as the number n of dictionary characters, and each response 01 is associated with each dictionary character.

Therefore, when learning the weighting coefficient Ws- of the intermediate layer of the neural network, for example, when inputting information on dictionary character pattern A to the input layer (in the case of learning for recognition target character A)
, 02 is “1” within a predetermined tolerance range
The learning is repeated until all other 02 to On become "o". Similarly, when learning target character B,
Enter the dictionary character pattern of B, 0. is ``1'' and all others are 0''.

Input the character pattern to be recognized into the input layer, and the response of the output layer obtained is 01 = 0.02, 02 = 0.95°Oa.
When O-0-On = O-0 and the response of Ol shows the maximum value, the recognition result is B. The success or failure of character recognition is determined, for example, when the determination value is 0.8, and when all the values of responses 01 to On are 0.8 or less, recognition is failed.

FIG. 6 shows an embodiment of a character recognition device to which the character recognition method according to the present invention is applied.

In FIG. 6, an imaging unit 400 and an A/D (Anal.

-gue to Digital) conversion unit 500
, image memory 510, image processing/recognition unit 520, dictionary character pattern memory 530. An information memory 5401 display device 550 regarding characters to be recognized is shown.

The A/D conversion unit 500 performs A/D conversion of the video signal from the television camera 400, and the period of A/D conversion is determined by the number of pixels forming one image. When one image is composed of 512 x 512 pixels, this A/
The period of D conversion is approximately 0.13 μs. The A/D converted video signal is stored in the image memory 510 as an image composed of 512×512 pixels. This image is read by the image processing/recognition unit 520 and undergoes preprocessing such as binarization and normalization for single character recognition.

The dictionary character pattern memory 530 stores binarized and normalized dictionary character patterns for single character recognition. As shown in FIG. 3, these dictionary characters can be stored in the dictionary character pattern memory 530 after the dictionary characters imaged by the television camera 400 are binarized and normalized by the image processing/recognition unit 520. However, from a keyboard (not shown) connected to the image processing/recognition unit 520, a password corresponding to each dictionary character pattern is input.
I It data can also be keyed in. The neural network for character recognition shown in FIG. 3 can be stored in the memory in the image processing/recognition unit 520 as a computer program. Therefore, one image processing
The recognition unit 520 can be configured by a microcomputer. Learning of the neural network for character recognition is performed by the image processing/recognition unit 520 using dictionary character patterns stored in the dictionary character pattern memory 530. Furthermore, determination of m pixel addresses on the character pattern image to be input to the input layer of the neural network is also executed by the image processing/recognition unit 520 once the dictionary character pattern has been stored in the dictionary character pattern memory 530. Ru.

The information memory 540 regarding characters to be recognized includes rules regarding character strings to be recognized (for example, when recognizing multiple character strings, the first letter must be an alphabetic character, etc.) and information about characters to be recognized when they are known in advance (automatic information). Recognition is a means to confirm that there are no errors), and the character string data is stored. In character recognition using a neural network as shown in Figure 3, the response 0 that outputs the largest value is
The character corresponding to 4 (-ax) is the recognition result, but if the maximum value of this response is not greater than a predetermined value or the difference from the second largest value is not greater than a predetermined value, recognition is considered to have failed. It will be done. If such a recognition failure occurs for only one character in a multiple character string, if there is no information about the character string to be recognized in advance, it will be a complete recognition failure and the recognition rate will decrease. . On the other hand, if you know in advance what the character string to be recognized is, select the first or second candidate.
If a candidate character matches a pre-stored recognition target character, the probability that the character will be mistaken as a recognition result is extremely small. In the case of a system that does not require high-speed continuous recognition, it is also possible to display character patterns that have failed in recognition on the display device 550 and leave the recognition to the operator. In such a case, by inputting the recognition result by the operator into the image processing/recognition unit using a means such as a keyboard, it is possible to retrain the neural network using the character pattern that failed to be recognized.

〔Effect of the invention〕

According to the present invention, a recognition target character image is binarized and M
After normalization with XM pixels, it is possible to extract points on m (m < M 6 In addition, if the weight coefficients of the middle layer of the neural network are learned in advance by combining the m pieces of extracted information with the neural network, character recognition can be performed using the m input data of the input layer of the neural network. Since this is achieved only by the product-sum operation of and the weighting coefficient of the intermediate layer, it is also effective in significantly shortening the time required to recognize one character.

[Brief explanation of drawings]

FIG. 1 is a diagram showing one embodiment of the present invention, FIG. 2 is a flowchart showing specific processing of the embodiment of FIG. 1, FIG. 3 is a diagram showing another embodiment of the present invention, and FIG. Figures and Figure 5 are
2 and 3, and FIG. 6 is a diagram showing an embodiment of the character recognition device. 200... Grayscale image of dictionary characters, 210... Binarized image of dictionary characters, 220... Normalized image of dictionary characters. 300...0'' or 111 jl for dictionary characters
Pattern frequency distribution calculation process, 310... Frequency distribution.

Claims (1)

[Claims] 1. In a character recognition method that recognizes fixed-form characters, a binarized normalized image of a predetermined size is created for all dictionary characters to be recognized, and the normalized image is A two-dimensional frequency distribution is obtained by accumulating "0" or "1" information for all dictionary character image data corresponding to each address, and based on this frequency distribution, a predetermined number of two-dimensional frequency distributions are calculated from the address with the largest amount of information. A character recognition method characterized by extracting an address and using information of a normalized image obtained on a two-dimensional address extracted by the means for an input recognition target character as input data of a character recognition means. 2. The character recognition method according to claim 1, wherein a predetermined number of two-dimensional addresses are extracted from addresses having a value closest to a frequency of 1/2 of the total number of characters in the dictionary. 3. The character recognition means is formed by a neural network consisting of three layers: an input layer, an intermediate layer, and an output layer, and the weighting coefficient of the intermediate layer that recognizes the recognition target character is determined in advance by learning dictionary character patterns. The character recognition method according to claim 1, characterized in that: 4. The weighting coefficient of the intermediate layer is such that when recognition fails a predetermined number of times with respect to input characters to be recognized, the number of inputs to the input layer of the neural network is set to at least 4. The character recognition method according to claim 3, wherein the character recognition method is re-learned by incrementing the number by 1. 5. The recognition target character is a part of a known recognition target character string, and when recognition of the recognition target character fails, a known character matching the first or second candidate character of the recognition target character is used. 4. The character recognition method according to claim 3, wherein the recognition result is a character. 6. A claim characterized in that when automatic character recognition fails, the character to be recognized is displayed and recognized, and the recognition result is used to relearn the weighting coefficients of the intermediate layer of the neural network. Character recognition method described in Section 3. 7. A character recognition device to which the character recognition method according to claim 1 is applied, comprising: an imaging unit for imaging a character pattern to be recognized and a dictionary character pattern; an A/D conversion unit for A/D converting the imaging signal; /D-converted image; an image processing/recognition unit for the image; a dictionary character memory unit and an information memory unit regarding characters to be recognized; and a display device for displaying character recognition results. character recognition device.