JPH01211090A - Recognizing system for character in information picture - Google Patents

Abstract

PURPOSE:To eliminate the ambiguity of character recognizing processing and to consistently and uniquely extract a character pattern from a scenic picture by updating the degree of similarity by the character recognizing processing and selecting the character pattern instead of adopting the character pattern obtained from a decision by the character recognizing processing as it is. CONSTITUTION:A degree of similarity updating circuit 30 is composed of a directly related pattern calculating circuit 100, a directly related pattern storing memory 101, a degree of similarity calculating circuit 102, a memory 103 for category parameter, a memory 104 for updating degree of similarity, an indirectly related pattern calculating circuit 105, a memory 106 for storing indirectly related pattern, and a character pattern deciding circuit 107. Here, the character pattern is extracted not by using the value of the degree of similarity from the character recognizing processing as it is but by comparing basic parameters between mutually related patterns and updating the degree of similarity, and the uniquely determined processed result of the character pattern is outputted to a memory 31 for recognized result. Thus, the ambiguity of the character recognizing processing can be eliminated, and the character pattern in the scenic picture can be determined consistently and uniquely.

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention relates to a character recognition device (hereinafter referred to as OCR: 0ptic
The present invention relates to a method for expanding the conventional method of reading only characters existing on paper to include characters existing in real space such as signs and signs.

In particular, it relates to a system that can recognize characters without prior knowledge of the character size, density of character lines, format, illumination conditions, etc. of characters existing in real space.

(Prior art) The prior art closest to the Kome invention is JP-A-62-
No. 173578 ``Character recognition method in scene images'' is available. The principle is shown in FIG. In addition, (1) to (4)
indicates each step.

In the same figure, first, the space in which characters exist is captured as a scene image using a TV camera, etc. (1) Next, character line candidate regions (hereinafter referred to as (2). It is not sufficient to cut out these regions ε alone. That is, there are separated characters (such as "old J, ri") that are composed of multiple areas. Therefore, we consider combinations of areas that are close to each other and extract them as character pattern candidates (3). Furthermore, whether or not these character pattern candidates correspond to characters is determined using character recognition processing (4). If it is determined to be a character, the category with the highest degree of similarity is output as the recognition result.

(Problem to be Solved by the Invention) In this way, conventionally, character recognition processing is used to extract character patterns from character pattern candidates, so character pattern candidates that do not correspond to the true character pattern may occur by chance. The degree of similarity to one of the categories becomes high, and errors often occur where the character pattern is extracted as a character pattern. This is also due to the fact that the current development policy for character recognition processing is to prevent input pattern jets as much as possible, and to output one of the categories as a character recognition result as much as possible.

Since character pattern candidates are determined from combinations of regions as described above, it is possible that character pattern candidates that do not correspond to characters are incorrectly extracted due to the ambiguity of character recognition processing as described above, and that this originally holds true for character patterns extracted from scene images. There was a problem in that a character pattern that contradicted the following two principles was extracted.

(1) A region corresponds to either one character pattern or none at all.

(2) If a certain area corresponds to any character pattern, the hole in that area or the area surrounding the area does not correspond to any character pattern.

If a pattern contradicting either of these two exists, there is a drawback that characters in the scene image cannot be uniquely recognized.

The purpose of this invention is to remove inconsistencies between character patterns caused by ambiguity in character recognition processing and the extraction method of character pattern candidates, and to uniquely extract character patterns from scene images.
The purpose of this invention is to provide a method for obtaining the recognition result.

[Means to solve the problem]

The character recognition method for scene images according to the present invention updates the degree of similarity by comparing the certainty of categories of high similarity patterns extracted by character recognition processing, and updates the degree of similarity in each group of mutually related patterns. The pattern with the highest degree of similarity is extracted as a character pattern, and the category giving that degree of similarity is output as a character recognition result.

[Effect]

This invention repeatedly updates the similarity of character recognition results between character patterns that are contradictory in nature based on the mutual relationship of the character patterns, and finally reaches the maximum similarity in each of the related pattern groups. The character pattern with the new category is output as a uniquely determined character pattern.

This method differs from the conventional technology in that the character pattern is selected by updating the degree of similarity through the character recognition process, rather than directly adopting the character pattern obtained by the determination through the character recognition process.

〔Example〕

FIG. 1 is a block diagram showing an example of a device for carrying out the present invention, in which 21 is an image input section, 22 is an image memory, 23 is a character line candidate area detection circuit, 24 is an area division result storage memory, 25 is a character pattern candidate extraction circuit, 26
27 is a character pattern candidate storage memory, 27 is a character recognition processing circuit, 28 is a character pattern determination circuit, 29 is a high similarity pattern storage memory, 3o is a similarity update circuit, and 31 is a practical memory for character pattern extraction and recognition. be.

In this operation, first, a real space including characters is captured from an image input unit 21 such as a TV camera as a scene image having shading, and is stored in the image memory 22.

Next, the image memory 2 is detected by the character line candidate area detection circuit 23.
The image No. 2 is divided into regions to detect character line candidate regions, and the results are stored in the memory 24.

Here, an example of processing by the character line candidate area detection circuit 23 is as shown in FIG. In other words, Figure 2 (
The image memory 22 is divided into NXM subblocks each consisting of nxm pixels as shown in a), and the threshold value for binarization in each subblock is set as shown in FIG. 2(b).
), and then allocate a threshold value to the center pixel of each subblock and perform linear interpolation to obtain the result shown in Figure 2 (C
) determine the threshold value for each pixel. Binarization is performed by comparing the threshold value of each pixel with the corresponding pixel of the original image (contents of the image memory 22), and the result is shown in FIG.
) in a binary (white, black) image, a labeling process is performed on 8 consecutive white or black pixels, and the image is divided into regions. Here, in order to detect areas that are considered to have a high possibility of corresponding to character lines, as shown in Figure 3, the density difference inside each area and the circumscribed rectangular frame surrounding it is evaluated and character line candidates are detected. Detect areas.

In FIG. 3, 131 is a region of interest, and 132 is the region of interest 1.
31 is a circumscribed rectangular frame, and 133 is a range included in the circumscribed rectangular frame 132 among the areas surrounding the attention area 131. The density difference is evaluated based on the average density of the region of interest 131 and range 133.

L2 is calculated, and if the density difference ΔL defined by ΔL=lL, -L2l satisfies ΔL>θL (θ is a threshold value), it is detected as a character line candidate region, and if it does not, it is detected as a region of interest 1.
31 into each area. This result is stored in the area division result storage memory 24.

The character pattern candidate extraction circuit 25 extracts character pattern candidates from the region division results that are the contents of the region division result storage memory 24. An example of the contents of this process is shown in FIG.

In FIGS. 4(a) to (d), 61 is the character "ke", 62 is the character "iJ, 63.64 is the area constituting the character 61,
65 and 66 are areas constituting the character 62, 67 is a search range, and 68 to 71 are combination patterns. Assuming that we are currently paying attention to the region 65, in the search range 67, the average density is IL compared to the average density L0 of the region 65.
-L.

1≦θ. (θ is a threshold value) or the value (white or black) in the above-mentioned binarization result is B, and the value of the area 65 is BO, then an area where B=BO is searched. Fourth
In the figure regions 64 and 66 can be found. Areas 64, 65.
Four combination patterns 68 to 71 are extracted from the 66 combinations. By the way, the aspect ratio r of the circumscribed rectangular frame 132 of the pattern corresponding to the character is θl < r < θh (θ1.θ
Since ゎ is in the range of the lower limit and upper limit of r), when it is determined whether this formula is satisfied, combination pattern 71 in Fig. 4 is rejected because it is horizontally long, and combination patterns 68 to 68 are rejected.
70 is extracted as a character pattern candidate and stored in the character pattern candidate storage memory 26 together with the identifier of the area forming each character pattern candidate.

Next, the character pattern candidates stored in the character pattern candidate storage memory 26 are processed by the character recognition processing circuit 27 to have a degree of similarity S (CI) (0≦) with the category CI in the character recognition dictionary.
S (CI)≦1, the closer the value is to 1, the more similar they are). Here, the degree of similarity S(CI) is obtained by a calculation method according to the character recognition method. Also, the distance between each category is determined through recognition processing! l (c+)(ora (CI
), the closer the value is to 0, the more similar) is used, but s z ((,) ==e −d L CI
The sono range is changed to 0≦S' (CI)≦ by transformation of 1st class.
Since it can be converted to 1, the following explanation will be made using similarity.

By character pattern determination circuit 2nd day, S(CI)≧T (
CI), (i=t, 2...-, N, N is the number of categories, 'r(at) is the threshold of similarity in category C1), and the category that satisfies this is evaluated. If there is at least one, it is considered a high similarity pattern, including the number of regions, identifiers, and candidate categories that make up the pattern.

The similarity is stored in the high similarity pattern storage memory 29.

Here, the candidate category refers to a category that satisfies the above formula.

If there is no category that satisfies the above formula, the category is rejected as there is no possibility of it corresponding to the character.

Here, there would be no problem if all the high similarity patterns stored in the high similarity pattern storage memory 29 correctly correspond to true characters, but in reality, due to the ambiguity of character recognition processing, true characters It often happens that patterns that do not correspond to the pattern are extracted as high similarity patterns.

Originally, the following principle holds between the region obtained by region segmentation of an image and a character pattern.

(1) Does one area correspond to one character pattern?
Either it does not correspond to any character pattern.

(2) If a certain region corresponds to a character pattern, the adjacent region does not correspond to the character pattern.

However, due to the ambiguity of character recognition processing as described above, patterns that contradict the above (1) and (2) are extracted as high similarity patterns. Furthermore, if a character pattern is attempted to be determined so as to satisfy the above (1) and (2) by simply comparing high similarity values, there is a high probability that an error will occur. Therefore, the similarity update circuit 30 updates the similarity while removing contradictions from the mutual relationships of high similarity patterns, and uniquely determines the character pattern in a manner that does not contradict the above (1) and (2). FIG. 5 shows details of the similarity update circuit 30.

In FIG. 5, 100 is a directly related pattern calculation circuit, 101
102 is a memory for storing directly related patterns, 102 is a similarity calculation circuit, 103 is a memory for category parameters, 104 is a memory for updating similarity, 105 is a circuit for calculating indirect related patterns, 106 is a memory for storing indirect related patterns, 107 is a character pattern This is a judgment circuit.

In the similarity update circuit 30, first, directly related patterns of each high similarity pattern are calculated by the directly related pattern calculation circuit 100 and stored in the directly related pattern storage memory 101. A determination flag is provided in the directly related pattern storage memory 101 corresponding to each pattern, and is turned on in the initial state.

Here, a directly related pattern i of a certain high similarity pattern i
' is a set RI, R2 that satisfies the following conditions.

It is defined as a pattern consisting of a region that is an element of either R3 or R4.

■PCR,, ■R2CP, ■ft3 (H, ■Q c
R4 Here, P is the set of regions constituting i, H is the set of holes in regions that are elements of P, and Q is the region surrounding the elements of P when all the elements of P are holes in other regions. Indicates a set. Examples of these are shown in FIGS. 6(a) and 6(b).

The similarity calculation circuit 102 shown in FIG. 5 calculates the correlation between each high similarity pattern stored in the high similarity pattern storage memory 29 shown in FIG. 1 and the directly related pattern stored in the directly related pattern storage memory 101. From the relationship, the similarity is updated as follows: (S-〇) (CJ) is candidate category C for pattern i.

initial similarity).

(CJ) if S r”'(CJ)<0.5j=
1.2. --------, m: m is the number of candidate categories for pattern i q+・(1/Nt') Σd, lΣr'th' (C, (')
S, 'k)1/ c/ (C') C' is the set of Nc candidate categories of directly related pattern i' di proverb 1/Ne, where rm1'(C+C') is the fitness coefficient can be,
The likelihood of candidate category C of pattern i and candidate category C' of pattern i' is compared, and if the former is more likely, the value is +1, and if the opposite is true, it takes a value close to -1. There are many possible evaluation parameters for the certainty of a category, such as the number of areas in the category.

There are the number of holes and the aspect ratio of the circumscribed rectangle. Specifically, it is calculated as follows.

(1) Directly related pattern i'' in the relationship P CR
Since i′ consists of multiple regions, if c′ (
If C' (candidate category of t') is a single region character, the probability of C' is lower than that of C (candidate category of i). increase degree. If C′ is a multi-basin character, then the rth character (
C, cl) takes a negative value depending on the difference in the number of regions between i' and C'.

Written as a formula, −N'l'l/NR' )) (NR'=2) e is a constant of O<e<1, NI and NR' are each i'
and the number of regions of C'.

(2) Directly related pattern i′i′ in R2CP relationship
and i, C' and C are replaced and I l/NR))
(NH-4) Here, NI and NR are the numbers of regions of i and C.

(3) Directly related pattern i′ in the relationship between R and CH Since the pattern of interest i has a hole, if C does not have a hole, C
The probability of is smaller than C'.

By making the r-th (c, c') negative, the similarity of C decreases. If C has a hole, r++(C.

C') becomes positive depending on the difference in the number of holes in i and C.

That is, II/HR)) (NR≧1) Here, Hl and HR are the numbers of holes i and C.

(4) Nil/HR by replacing directly related patterns i'i and i, C and C' in the relationship of QER4)) (N R'≧1) Here, H1' and HR' are i' and C ′ is the number of holes.

The aspect ratios of the circumscribed rectangular frames of each category are calculated as follows.

r++'(C,C')-Ce/4)(2-I QA I
−QACl) −(e/8)(IQAI − QACI −I QA I−QAC+ ”) where, Q
AC, QAC', QAl, QAI' are C, C',
is the quantized aspect ratio QA value of i, i'. The QA value is a value obtained by quantizing the aspect ratio into several levels. According to this formula, if i and C have similar QA values, i' and C'
If they take different QA values, the r-th (C, C') becomes positive, and the similarity of C increases.

As shown in FIG. 7, the above processing is performed in two stages (the first half is based on the number of regions, and the second half is based on the aspect ratio). Also, an appropriate number of times (
ideally until each similarity converges to a certain value), and the results are stored in the update similarity memory 104. Furthermore, as shown in FIG. 8, each category evaluation parameter is stored in the category parameter 103 in a table format, so that it is appropriately used to obtain the parameters c and c' during the above-described processing.

Note that the QA values in FIG. 8 are quantized as follows.

QA=1 i fAp≦0.5 QA-2if0.5<Ap≦0.8 QA=3if0.8<Ap≦1.25 QA=4if1.25<Ap≦2.0 QA=5ifAp>2.0 Next Next, character patterns are determined, and the high similarity patterns are divided into indirectly related pattern groups by the indirectly related pattern calculation circuit 105 in advance. A certain pattern 1
Put 0 into the initial group. Next, all 10 directly related patterns 11 are placed in that initial group. Similarly,
Insert directly related pattern 1 k+1 for each ik (k>0). Here, the ones that have already been placed are not processed. If there are no more new patterns to add, create a new initial group and put the next pattern that has not been placed in any initial group into this, and do the same. Process. Repeat this until all patterns are placed in one of the groups. The set of regions existing in the pattern of a certain initial group is Rp I The set of holes in the region of any element of Rp is Rh1 The set of regions surrounding the elements of Rp is R1. If any pattern of a different initial group If each group is composed of elements R,,Rh,R,, these initial groups are combined to form an indirectly related pattern group. An example of this is shown in FIG. The processing results of the indirect related pattern calculation circuit 105 are stored in the indirect related pattern storage memory 106. After updating the above-mentioned similarity, the character pattern determination circuit 107 calculates the maximum similarity S l, max-max (S 1' to ' (C+), ..., 31'to' (C,) for each pattern i by the character pattern determination circuit 107.
) is obtained. Next, the pattern with the maximum S,,txax in each indirectly related pattern group is extracted as a character pattern, candidate categories that give the similarity are extracted as character recognition results, and these are used for character pattern extraction and recognition. Similarly, character patterns and recognition results are extracted from all indirectly related pattern groups stored in the memory 31. Thereafter, patterns having at least one of the following areas are deemed not to correspond and are stored in the directly related pattern storage memory 101.
The determination flag for that pattern is turned OFF. That is, (a) an area constituting a pattern determined as a character pattern (b) an area (C) corresponding to the hole in the area of (a)
This is any area surrounding the area in (a). Directly related pattern storage memory 101
If all of the decision graphs are OFF, the process is stopped, and if ON patterns remain, the same process is performed only for the ON patterns.

As described above, the result of character pattern extraction and recognition processing is obtained in the practical memory 31.

In the above description, the method described above is used to detect character line candidate regions, but other image region division methods can also be applied to the present invention.

〔Effect of the invention〕

As explained above, this invention does not rely on the value of the similarity itself in character recognition processing, but rather compares basic parameters between mutually related patterns and updates the similarity. This method has the advantage that ambiguity in recognition processing can be removed and character patterns in a scene image can be uniquely determined without contradiction.

Therefore, this invention is based on the letter 1 of the car license plate.
It can be widely applied to patrol robots that recognize characters on name tags, sorting by recognizing the destination of packages, automation of work, and character recognition written on documents in book formation.

[Brief explanation of the drawing]

Fig. 1 is a block diagram showing an example of a device implementing the present invention, Fig. 2 is a diagram showing an example of binarization processing, Fig. 3 is a diagram explaining a density difference evaluation method, and Fig. 4 is a text diagram. An explanatory diagram of the principle of extracting pattern candidates, FIG. 5 is a block diagram showing an example of a similarity updating circuit, FIG. 6 is a diagram showing an example of a directly related pattern and an indirectly related pattern group, and FIG. 7 is a diagram showing the similarity. FIG. 8 is a flowchart showing how to obtain a coupling coefficient for updating, FIG. 8 is a diagram showing an example of an evaluation parameter table, and FIG. 9 is an explanatory diagram of the prior art. In the figure, 21 is an image input unit, 22 is an image memory, 23 is a character line candidate area detection circuit, 24 is a region division result storage memory, 25 is a character pattern candidate extraction circuit, 26 is a character pattern candidate storage memory, and 27 is a character pattern candidate storage memory. Character recognition processing circuit, 28 character pattern determination circuit, 2nd high similarity pattern storage memory, 3o similarity update circuit, 31 character pattern extraction and recognition practical memory, 61 character "ke", 62 is the character ``J, 63.64 is the area that makes up the character 61, 65
．． 68 is an area forming the character 62, 67 is a search range, 6
8 to 71 are combination patterns, 100 is a directly related pattern calculation circuit, 101 is a directly related pattern storage memory, 10
2 is a similarity calculation circuit, 103 is a memory for category parameters, 104 is a memory for updated similarity, 105 is an indirect related pattern calculation circuit, 106 is a memory for storing indirect related patterns, and 107 is a character pattern determination circuit. . Figure 1 Figure 2 (a) (b) (C) (d) Figure 3 Figure 4 Figure 5 Standing Figure 6 (a) Figure 6 (b)

Claims (1)

[Claims] A scene containing characters existing on a two-dimensional surface or three-dimensional space is input as a gray scale image, areas that are likely to correspond to character lines are detected as regions, and combinations of regions that exist in the vicinity of each other are used as character pattern candidates. In a character recognition processing method for determining whether or not the character pattern candidate corresponds to a character, the character pattern candidate is extracted as A scene image characterized by updating the degree of similarity, finally extracting a pattern having the maximum degree of similarity among mutually related pattern groups as a character pattern, and outputting a category giving that degree of similarity as a character recognition result. Character recognition method inside.