JPS60142481A - Character recognizing device - Google Patents

Abstract

PURPOSE:To make it possible to recognize also a character having thin line width with high recognition by adding an isolated bit removing circuit and a weighting/thinning circuit to the titled device. CONSTITUTION:An isolated bit removing circuit 13 is connected among a frame memory 4, a thinning circuit 14 and a thickening circuit 7 and the weighting/ thinning circuit 14 instead of an ordinary thinning circuit, the weighting circuit 15 for controlling the circuit 14 and a slice level setting circuit 16 are also connected. The circuit 13 removes a bit isolated only by one bit and quantized to ''1'' in a picture including an inputted reference character as a bit due to noise or dust independently of the character part. The weighting/thinning circuit 14 is constituted of a weighting circuit 17 and a thinning deciding circuit 18 and controlled by the weight setting circuit 15 and the slice level setting circuit 16 respectively. The circuit 14 weights each bit of a line part of the inputted reference character.

Description

DETAILED DESCRIPTION OF THE INVENTION (1) Technical Field of the Invention The present invention relates to a character recognition device that recognizes digitized character images input from a television camera, digitizer, or the like.

(2) Technical Background With the recent development of digital signal processing technology, it has become possible for computers to recognize characters drawn on a page of paper, industrial products, etc. Along with this, automatic recognition devices have been put into practical use, particularly for industrial characters such as manufacturing numbers and types of industrial products, postal codes, and handwritten characters.

(3) Conventional technology and problems As a conventional character recognition device using digital signal processing,
Decide on the types of characters that should be recognized! It is common that characters are marked and stored in a pattern in advance, and recognition is performed by pattern matching the characters to be recognized and a dictionary.

An example of the configuration of such a character recognition device is shown in FIG.

First, in order to create a dictionary, switch S is connected to the terminal 1 side and the dictionary creation mode is set. Thereafter, the reference of the character to be recognized is manually retrieved from the television camera 1. An image input from a television camera 1 controlled by a camera controller 1-roll unit 2 (CCU) is converted into an electrical signal, and then converted into a binary signal (0 or 1) by an A/D converter 3.
) and converted to a digital signal.

Through the above two processes, the image for one character is divided into fine grid pixels (-pixel-, hereinafter referred to as bits) of, for example, 64 bits x 64 hits, and the part where the character exists for each hit is divided into 1 is assigned, otherwise 0
is allocated and 6 for this one side (64 pills-X 64
Information about the number of hits) is stored in the frame memory 4. In this case each bit is generally very small compared to the character, so
The line formed by the input reference characters is represented on the binarized screen as a set of binarized data having a width of several bits. (The parts other than the lines of characters are 0 except for the parts where noise is placed.) The numbers input to create the dictionary as described above and stored in the frame memory are 0.
First, the line thinning circuit 5 thins the human characters and reference characters on the character image. That is, as described above, since the line width of the input reference character is several bits, this is converted into a line as narrow as 1 bit or 2 pits. As a result, the skeleton structure of the human-powered reference character "ζ" is patterned. The input reference character thinned in this way is written in the thin mask pattern memory 6 as a thin mask pattern. On the other hand, the frame memory 4, and the input reference character is thickened in the thickening circuit 7. That is, for example, a line with a line width of 3 bits is converted into a thick line with a width of 5 bits. Characteristics that cannot be expressed only by the skeletal structure of the reference character are patterned.

The input reference character thus made into a thick line is stored in the thick mask pattern memory 8 as a large mask pattern. By inputting and processing the reference characters character by character in this way, a patterned dictionary (2) is created for each reference character as a thin mask pattern and a large mask pattern.

After the dictionary is created as described above, characters are actually input and character recognition is performed. For this purpose, switch S is connected to terminals 1 and 2 to set the recognition mode.

Then input the characters to be recognized through TV camera 1,
As in the case of creating the dictionary, the A/D converter 36j converts it into a binary digital signal and stores it in the array memory 4. The character image thus input is first matched in the pattern matching circuit 9 with the fine mask pattern of each reference character stored in the fine mask pattern memory 6. Then, as an output of the pattern matching circuit 9, the thin mask pattern and the character portion of the input image are superimposed for each reference character, and the number of bits in the portion where the thin mask pattern does not protrude from the character portion of the input image is determined as the degree of mismatch A. It is output to the mismatch degree determination circuit 11. Similarly, the input character image 1 is matched with the large mask pattern of each reference character stored in the large mask pattern memory 8 in the pattern matching circuit 10. Then, as an output of the pattern mapping circuit 10, the thick mask pattern and the character part of the input image are superimposed for each reference character.

Contrary to the above, the number of bits in the part where the character part does not protrude from the thick mask pattern is determined as the mismatch level B by the mismatch level determination circuit 11
is output to. Then, the mismatch degree determination circuit 11 calculates the sum of the mismatch degrees A and B with respect to the input image for each reference character, and assumes that the reference character with the smallest sum (unit: number of bits) is the input character.
The character information is output to the result output device 12.

In the above-described character recognition device, a new dictionary can be created even if the type of character to be recognized changes, so character recognition can be performed with great flexibility. Furthermore, since recognition is performed using two types of dictionaries, a thin mask pattern and a large mask pattern, it is possible to handle characters with different line widths, and a high recognition rate can be obtained. By the way, in FIG. 1, the specific processing operation of the thickening circuit 7 is realized by keeping the line portion (1 portion) of the input character image constant in the width direction of the line and making it thicker (1-number). On the other hand, the specific processing operation of the line thinning circuit 5 is not as simple as in the case of thickening, so it will be explained using an example shown in FIG.

FIG. 2(a) is an example in which a part of a line formed by an input reference character is represented as a set of 3 pino[·hiso 1-] quantized to a width of 0. In this case, only Hiso 1-(Φ is quantized to 0, and the line width is thinner only in this part. In order to thin such a 3-bit width line to a 1-bit width line, looks at the surrounding 8 bits in the up, down, left, right and diagonal directions for each hit in the line part (the part of Hiso I~ that is quantized to 1), and if all of the 3 surrounding 8 bits are 1, then Thinning is performed by assigning 1 to each bit, and setting that bit to 0 if there is at least one O.

Fig. 2 (In the case of al) 2 For example, looking at the bit ■, the left and right 2 hits, the hit and the bottom 3 hits are 1, but the upper 3 hits to 1 are 0, so the bit 1-■ is set to 0. In other words, the bit ■ is discarded. However, when looking at the hit ■, the bits are all 1 in the vertical, horizontal, and diagonal surroundings, so the pitch I is discarded.
・1 is assigned to ■ and that bit is left. The same operation is performed for all the bits in the other line parts.The 3-bit width line in ta+ in Figure 2 is then thinned into a 1-bit width line as indicated by the diagonal lines in the same figure. . However, according to this method of line thinning, bits ■, ■, and ■ should be left (assigned 1), but since the bit and node ■ are O, it is necessary to look at the eight people around each person. Three bits are discarded by
As shown in the figure, there was a problem that when the line was thinned to one hit width, that part was cut off. Such a problem occurs not only when thinning a 3-bit line to a 1-bit width, but also when the line width of an input reference character becomes thinner in the middle. Furthermore, if you try to thin a 2-bit width line using the thinning method described in Figure 2 (bl), each line will have 8 lines around it, or all lines will be 1 line.
Since there is no bit that becomes 0, there is a problem in that all hits are assigned O and erased. This inconvenient thinning of the lines results in a problem in that characters are more likely to be misrecognized.

It is also shown in FIG. J, Una According to conventional character recognition devices, if noise components are included in the image containing the input reference character, parts other than one character will be made into K111 lines or thick lines, and they will be replaced with a thin mask pattern or a large mask. When used as a pattern.

Pattern matching during recognition, 4! In matching I with a thin mask pattern, the noise gB increases the output value of the degree of mismatch A.

There is a problem in that the recognition of characters with a large line width of ♀■1 that is highly dependent on pattern matching with a thin mask pattern deteriorates.

(4) Purpose of the Invention In order to eliminate the above-mentioned problems, the present invention provides an isolated hint in a character recognition device constituted by a conventional image input device, a Δ/D converter, an array, a memory, a pattern matching circuit, etc. To provide a character recognition device capable of recognizing characters with a high recognition rate even when characters with narrow line widths are inscribed by adding a removal circuit, a weight-increasing line thinning circuit, and the like.

(5) Structure of the Invention The present invention provides means for detecting binary data of bits surrounding a bean node to be detected, means for applying predetermined weighting to the surrounding bits, and weighting based on the binary data of the two surrounding bits. and thinning means for outputting the bit to be detected as a mask pattern hit when the sum of the weights exceeds a first set value. A device, a means for detecting binary data of people around the 1st to 4th bits to be detected, and a means for detecting binary data of people around the 1st stage and 4th round without subjecting the surrounding bits to a predetermined m a hand/leg for detecting the sum of weights, a thinning means for outputting a bit that should not be detected as a hint for a master pattern when the sum of the weights exceeds a first set value, and the peripheral pick. The present invention provides a character recognition device characterized in that it has means for erasing detected pinot 1- as an independent bit when the sum of two digits is less than a predetermined second set value.

(6) Examples of the invention Embodiments of the present invention will be described in detail below.

FIG. 3 is an overall configuration diagram of a character recognition device according to the present invention.

What is different between the present invention and the conventional example shown in FIG.

An isolated weight removal circuit 13 is connected between the frame memory, the thinning circuit, and the thickening circuit. 14, and the weight setting circuit 15 and slice level setting circuit 16 that control it are connected.Other configurations and operations are the same as in the conventional example shown in FIG.
The explanation will be omitted.

First, the isolated bit removal circuit 13 removes the isolated bit 1, which is quantized to 1, from the image containing the input reference character, assuming that it is not a 9-character part but is caused by noise or dirt. This is a circuit that can cause problems. The specific operation is to sequentially look at the bits that are quantized to 1 in the input image, and if the 8 bits around that bit are all 0, that bit is removed as an isolated bit due to noise or dirt. . This circuit can prevent parts other than characters from being erroneously made into thinner or thicker lines due to noise or dust.

Next, the thick line making circuit 7 is the same as the conventional example, so it is omitted, and the one with 1 weight is the thin line setting circuit 141 and the weight setting circuit 15. and the operation of the slice level setting circuit 16 will be explained. FIG. 4 is a detailed configuration diagram of these three circuit parts. In other words, the weighting is the circuit 1 when the wiring circuit 14 is
7 and a thinning determination circuit 18, and are controlled by a weight setting circuit 15 and a slice level setting circuit 16, respectively.

The weighting and thinning circuit 14 is a circuit for performing weighting &J on each bit of the line portion of the input reference character to convert it into Itnl lines. A specific example of the weight setting is shown in the fifth section.
As shown in the figure.

First, in Figure 5 (al is the bit in the shaded part (the bit quantized to 1), look at the surrounding 8 bits, and if Hiso I.p is quantized to 1, the score is 2.
If it is quantized to 0, no point is added.

Below, the weighted scores for hits q, r, s, t, u, V, and w are set as 3, 2.3, 2.3, and 2.3 points, depending on the distance from the hatched part. 2 Calculate the total score for the surrounding 8 bits and g.

The result is used as the weight for Hiso 1 in the shaded area.

Figure 5 (bl is a setting example in which the right side and lower side of the hint to be noticed (hatched bit) is set to give directionality.In other words, the weight of the right side and lower side is higher. Therefore, the hints on the right and bottom sides are quantized to 1, and the bits in the shaded area have a larger weight.

Figure 6 ta+ is 1 using the weight setting example of Figure 5 (bl).
This is an example in which weighting is applied to each of the pips I in the 1-width line portion of the 3 pins that are missing at a location (piston o). For example, if we look at bit ■, Figure 5 (bl bit
- Since only the bits corresponding to u, v, w are quantized to 1, the weight for u', v, w is 5.4.
．． The 5 points are added to give 4 points.

Also, regarding bit ■, Fig. 5 (bl's secret i)
Since all of the surrounding eight hits corresponding to p = w are quantized to 1, the weights for bits p to W are all added up, resulting in a maximum score of 28 points.

Also missing is the inner hit near @,
Regarding [phase], ■, etc., for example, when looking at p]-■, only the bit corresponding to hiso l-p of tb+ in FIG. 5 must be quantized to 0.

If you do a heavy survey, you will get a relatively high score of 25 points. The same applies to bit [phase] and 0.

After weighting as shown in Figure 6 (al), only the bits with a weight of 21 points or more are left as thinning bits (quantized to 1 and other bits are 0).
As a result, the part surrounded by the thick line in the figure becomes the result of thinning, and is output as a thin mask pattern. According to such a method, even if HISOL - ■ is missing, the HIZ and TORTOISE @ parts will not be erased 5 by thinning.

In the above-mentioned process, the setting of the weight inquiry method as shown in FIG.
In this step, heavy interrogation as shown in FIG. 6+a+ is performed. And in the Sura Is 1 Nohel setting circuit 1G “ζ
In the case of the example shown in Fig. 6 ia), the line thinning determination circuit 18 sets Hm igt according to the recorded value in the line thinning determination circuit 18.
(a process for extracting the large frame part in FIG. 6(.l)).

Figure 5 (l)) is 2.
This is an example in which market I=J is applied to each pitch I of a line portion with a bit width. In this case, the market setting example in Figure 5 (I)) is an example of directional weight setting, so the points for heavy weighting are different between the upper bit 0 and the lower bit ``■''.
For example, the darkness value is set to 1 in the 1 slice level setting circuit 16.
By setting seven points, the part surrounded by the thick line frame in the same figure becomes the result of converting into aa line, and is output as a thin mask pattern. In this way, the slice level setting circuit 16
In A, it is possible to thin line portions with a width of 2 bits by changing the darkness value. Easy line thinning from 3-bit width to 2-bit width 1
It will be possible on 11th. Furthermore, by setting weights other than those shown in FIG. 5 in the weight setting circuit 15 (for example, making the viewing range wider than the surrounding 8 bits), it is possible to thin lines even for lines wider than 3 bits. Become.

As shown in [2], by performing the weighting (=J-J line thinning) according to the present invention, various types of lines, including thinning of the 8 line width part, which was difficult in the conventional method, can be achieved. It becomes possible to perform line thinning.The thin mask pattern thus thinned is stored in the ha master pattern memory 6 (FIG. 3).

By performing pattern matching as in the conventional example, character recognition can be performed with a higher recognition rate than in the conventional example.

(7) Results of the Invention According to the present invention, the isolated dot removal circuit can eliminate unnecessary histograms caused by noise or dust other than characters.2 Especially characters with a line width of 6111. recognition rate can be increased.

In addition to this, by increasing the weight (=J and making the J line thinner), the degree of freedom in making +no lines can be increased compared to the conventional method, and higher recognition can be achieved6
It is possible to create a mask pattern.

[Brief explanation of the drawing]

Fig. 1 is a diagram showing the configuration of a conventional character recognition device. Fig. 2 is an explanatory diagram of conventional thinning processing. Fig. 3 is a configuration diagram of a character recognition device according to the present invention. Fig. 4 is a diagram showing the configuration of a character recognition device according to the present invention. A configuration diagram of a thinning circuit and its peripheral circuits. Figure 5 (al, '(bl) is a pattern diagram of the weight setting, Figure 6 (ta+), (b) is the actual weight adjustment pattern diagram. 1...Television camera 2...Camera controller I-Roll Uniso I-(CCU) 3...A/I) converter
4...Frame memory 5...Thinning circuit 6...Thin mask pattern memory 7...Thick line making circuit 8...Thick mask pattern memory 9.19...Pattern matching circuit 11... Discrepancy degree publication circuit 1
2...Result output device 13...Isolated gap 1-removal circuit 14...Weighting thinning circuit 15...Weight setting circuit 16...Slice level setting circuit 17...Weighting circuit 18... Thinning determination circuit S... Swissochi Fig. 1 and Fig. 2 (0)

Claims (3)

[Claims] (1) A means for detecting binary data of bits surrounding a bit to be detected, a means for applying a predetermined weight to the surrounding bits, and a sum of m based on the binary data of the two surrounding bits. 1. A character recognition device comprising: a means for detecting a weight, and a thinning means for outputting a bit to be detected as a master pattern bit when the sum of the weights exceeds a first set value. (2) The character recognition device according to claim 1, wherein the mask pattern pinot [. is a line thinning hit. (3) Means for detecting binary data of hints around the person to be detected, means for applying a predetermined weight to the surrounding people, and detecting the sum of weights based on the binary data of 9 surrounding bits. a thinning means for outputting undetected bits as mask pattern bits when the sum of the weights exceeds a first set value; A character recognition device comprising means for erasing a detected bit as an isolated hint when the value is equal to or less than a second set value.