JPS6365991B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a pattern recognition method, and particularly to an improvement of a character recognition method that employs a pattern matching method (also referred to as a similarity method).

Traditionally, character recognition methods have adopted pattern matching methods for printed characters, while structural analysis methods that use thin lines and center lines for handwritten character recognition (e.g. Sano et al. ``H. -8959 optical character reader"
Hitachi Hyoron, vol. 54, pp. 1077-1082, 1972), or structural analysis methods that focus on white backgrounds and contours have been considered, and some have been manufactured as products.

However, since these structural analysis methods for recognizing handwritten characters utilize topology (connection relationships of branches, presence or absence of loops, etc.), they are vulnerable to deformations such as line breaks and contact, and each It has the disadvantage that a dedicated standard pattern must be prepared for modification. Moreover, these standard patterns are often difficult to create automatically and often rely on human labor, which incurs a great deal of cost in developing character reading devices (hereinafter referred to as OCR).

On the other hand, although the pattern matching method has been successfully applied to printed character recognition, a recognition device using only the pattern matching method for handwritten characters has not been put to practical use.

However, the advantages of the pattern matching method over the disadvantages of the above structural analysis method include: 1) It is possible to calculate the similarity or distance between patterns with any difference in structure; 2) The recognition process is simple and clear because it is numerical rather than algorithmic; 3) Development costs are reduced because standard patterns can be automatically generated. While it is generally recognized that there is a small amount of deformation, conventional pattern matching methods are not able to recognize highly deformed patterns such as handwritten characters.

Attempts have also been made to recognize handwritten characters while taking advantage of the advantages of pattern matching methods. For example, Japanese Patent Application No. 47-50720 ``Graphic Recognition Device'' or Japanese Patent Application No. 53-100377 ``Graphic Recognition Method'' utilize the advantages of pattern matching methods to develop a device or method that enables handwritten character recognition. It shows.

The above method performs a large blurring process to deal with the deformation of handwritten characters, but if the blurring is performed directly at that time, the resolution will deteriorate to the point where it is unrecognizable. Then, the directionality is extracted via the thinning pattern, and the pattern surface is divided for each direction. The introduction of processing was new and unprecedented.

However, in the above method, the connection relationship of the thinning pattern is used to extract the directionality. It has already been generally recognized that line thinning methods have many problems because there is no method for obtaining a thinning pattern that is close to the ideal. for example,
Center lines are disrupted at intersections where crushed characters cannot be dealt with, and extra short branches occur at bend points. Therefore, the method of extracting directionality via a thinning pattern may eliminate the advantages that the pattern matching method originally has.

Therefore, an object of the present invention is to solve the above-mentioned problems when realizing a pattern matching method that introduces directionality and large blurring processing. That is, the present invention provides a method for extracting directionality by subjecting a character pattern to spatial differentiation without determining the center line in order to extract directionality.

First, an overview of the method of the present invention will be explained along with the recognition process. FIG. 1 is a flowchart of the recognition process of this method. If we focus on one character and explain the process,
A character pattern printed, typed, or handwritten on paper is converted by photoelectric conversion into a signal with a value corresponding to reflectance. Photoelectric conversion is usually performed by scanning the paper surface in a certain order, and at this stage, both the characters and other white areas are treated equally, and the position of the characters has not yet been identified. . The signal after photoelectric conversion may be an analog quantity such as voltage, or a digital quantity obtained by quantizing it. Since these quantities require too much processing to pass through to recognition processing as they are, they are binarized so that the character pattern takes a value of 0 at points on the white background and a value of 1 at points on the black line. . The photoelectrically converted signal is affected by the optical system and illumination system, causing shading (uneven ratio between the reflectance and the measured result), small disturbances in the reflectance on the paper, and photoelectric conversion. Noise is mixed in due to the effects of element thermal noise, etc., so direct binarization processing cannot be performed. Therefore, preprocessing such as shading removal and noise removal by filtering is required before binarization. These are process 2 in FIG.

Processing 3 separates each character from the binary pattern obtained in Processing 2 (usually held on a line-by-line basis),
Cut out a frame of a certain size. The extracted pattern is a binary pattern f(i,j) (i=1, 2,
..., m; j=1, 2, ..., n).

The point (i, j) where f(i,j)=0 is a white point on the paper surface, and the point where f(i,j)=1 is a printed part or a black part written in writing.

Handwritten characters are usually written on a printed sheet of paper in a predetermined frame called a form. In most cases, the frame is about 7 mm in height and 5 mm in width, and if the scanning pitch is 0.11 mm in height and width, the cut out pattern will be approximately
The size is 64 x 48 bits (m = 64, n = 48). Since handwritten characters are written in various sizes and positions within this range, these are normalized to a fixed position and size so that pattern matching works effectively.
This process is process 4 in FIG. Processing 2, 3, 4
will be collectively referred to as preprocessing 10. FIG. 2 illustrates an example of position and size normalization.

The position and size are normalized as follows. As shown in Fig. 2, the cutout pattern f(i,j)
Check the vertical character's top coordinate i _s and bottom coordinate
i _e , the horizontal left end coordinate j _s , and the right end coordinate j _e are determined.
Assuming that the size normalization coefficients, that is, the expansion/contraction ratios are G _i and G _j in the vertical and horizontal directions, respectively, these are determined as follows. Assuming that the character after normalization has a maximum size of 28 bits × 28 bits, G _i = max ((i _e −i _s +1) / 28, 0.25) (1) G _j = max ((j _e −j _s +1)/28, 0.25) (2) Here, in equation (1), max is (i _e −i _s +
1) This means taking the larger value of /28 and 0.25. Furthermore, the purpose of taking the larger value is to prevent an extremely large enlargement in the normalization process shown by equation (5) when the size of the cutout pattern is extremely small.

In the case of equation (1), the value of G _i will never be less than 0.25, and therefore, in equation (5), expansion by a factor of four (the reciprocal of 0.25) or more will not be performed. Equation (1) is about normalization in the vertical direction, whereas equation (2) is about normalization in the horizontal direction, and has the same meaning as equation (1).

The starting point (IS,
JS) as follows.

IS=i _s −G _i ×2 (3) JS=j _s −G _j ×2 (4) This allows the normalized pattern to be
It can be centered like this.

At this time, the normalized pattern is obtained using the following formula.

g (i, j) = f (IS + G _i (i-1) , JS + G _j (j-1)) (5) Here, in the case of Figure 2, i and j are 1 to 32
Takes an integer. As can be seen from equation (5), the normalized pattern g(i, j) is a sample of the cutout pattern f.

In FIG. 1, the next process is directionality extraction (process 5). Directional extraction is performed using patterns g(i, j), (i=1, 2,..., after position and size normalization).
m ₁ ; J=1, 2,..., n ₁ ).
Pattern g(i,j) is a pattern of 8 faces in each direction.
h〓(i,j)(θ=1,2,...8;i=1,2,
…, m ₂ ; j=1, 2, …, n ₂ ).
θ is obtained by quantizing all directions (360°) at intervals of 45°, and represents the 8 directions shown in FIG.

Directivity can be detected by applying spatial differentiation. True differentiation is not possible in quantized space, but it can be obtained approximately. Now, a quadratic surface is approximated around a certain point (i, j) of pattern g (i, j). That is, we set g〓(ξ+i, η+j)=a ₁ ξ ² +2a ₂ ξη +a ₃ η ² +a ₄ ξ+a ₅ η+a ₆ (6). The partial differential values at points (i, j) in the vertical and horizontal directions are respectively ∂g〓/∂ξ〓 ₌ 〓 ₌₀ =a ₄ (7) ∂g〓/∂η〓 ₌ 〓 ₌₀ =a ₅ (8). Therefore, the point (i,j)
The coefficients a ₄ and a ₅ of equation (6) can be found by the least squares method from nine points around . After all, a ₄ = 1/3 {g(i+1,j+1)+g(i+1,j
)g(i,1,j-1)} -1/3{g(i+1,j+1)+g(i-1,j
)+g(i-1,j-1)}(9) a ₅ =1/3{g(i+1,j+1)+g(i,j+1
)+g(i-1,j+1)}-1/3{g(i+1,j-1)+g(i,j-1
)+g(i-1,j-1)}(10).

Here g(i,j) is a binary pattern, so
The resulting spatial differential vectors (a ₄ , a ₅ ) take only values within a certain range: −1≦a ₄ ≦1 (11) −1≦a ₅ ≦1 (12), and therefore each vector (a ₄ ,
If the normal direction or the direction of the contour perpendicular to it is calculated in advance for a ₅ ), the direction and strength of change can be obtained by table lookup immediately after calculating equations (9) and (10). be able to.

Actually, there is no need to divide by 3 in equations (9) and (10). Here, a ₄ ′=3・a ₄ (13) a ₅ ′=3・a ₅ (14) are determined as spatial differentials.

The strength of the change is determined in proportion to √( ₄ ′ ² + ₅ ′ ² ). In reality, it is quantized into integers and within a certain width. Letting the quantized contour be the direction θ and the strength of the quantized change be v, these can be expressed from (a ₄ ′, a ₅ ′) as shown in the following equation.

θ=θ(a ₄ ′, a ₅ ′) (15) v=v(a ₄ ′, a ₅ ′) (16) Equations (15) and (16) can be realized in the form of a table.

Figure 4 shows an example of spatial differentiation, Figure 5 shows the direction θ,
The figure shows the numerical value of strength v. For example, point (i,
If the pattern of 9 points around j) is as shown in Figure 4 a, then from equations (9), (10), (13), and (14),
a ₄ ′=2, a ₅ ′=2. By subtracting the table shown in FIG. 5 from these values, θ=2, and by subtracting the table shown in FIG. 6, v=6, indicating the direction and strength.

(θ, v) is obtained for all points of the normalized pattern g(i, j), and the directional feature pattern {h〓(i′, j′)|θ=1,2, …
，
8;i'=1,2,..., _m2 ;j'=1,2,..., _n2 }
Create. That is, when (θ, v) is found for point (i, j), h〓(i′, j′)=h〓(i′, j′)+v (17). Here, i′=[m ₂ /m ₁・(i−1)]+1 (18) j′=[n ₂ /n ₁・(j−1)]+1 (19) Also, [ ] is a Gaussian symbol. be.

The directional pattern h〓 obtained in this way is subjected to a blurring process. This is process 6 in FIG. The blurring process uses a two-dimensional point spread function (point spread function).
function) into the directional pattern h〓. The point spread function is a spatial low-pass filter. Now, the point spread function is b(k,l)
When (k=1, 2,..., m ₃ ;=1, 2,..., n ₃ ), the blurring process can be expressed by the following equation.

(h〓〓(i,j)= _n3 〓 ^k=1 _o2 〓 〓 ^l=1 b(k,l)・h〓(i−k+m ₃ +1/2, j−l+
n ₃ +1/2) (20) The point spread function b(k,l) is a Gaussian function appropriately approximated by an integer value. Here, 0
The range having values other than the above is defined as 3 bits in the vertical and horizontal directions centered on the origin, that is, m ₃ =5, n ₃ =5. function b
Specific values of (k, l) are shown in FIG. The degree of blurring is set by a parameter σ that specifies the degree of spread of the Gaussian function. The optimal degree of blurring depends on the type of character and the size of the directional pattern (m ₂ , n _{2 )}
Depends on etc. FIG. 7 shows the results obtained for the optimum σ=1 for m ₂ =n ₂ =8.

The blurring process (process 6 in FIG. 1) is also applied to the pattern h ₀ (i', j') that is simply compressed from the size-normalized pattern g (i, j). The pattern h ₀ (i', j') before the blurring process is obtained by applying the process shown by the following equation to each point of g(i, j). That is, h ₀ (i', j') = h ₀ (i', j') + g (i, j) (21) where, i' = [m ₂ /m ₁ (i-1)] + 1 (22 ) j′=[m ₂ /m ₁ (j-1)]+1 (23) Also, [ ] is a Gaussian symbol.

The blurring process for h ₀ (i′, j′) is h〓(i′, j
′)
^In _{exactly the} ^same _{way as for}
It is performed according to ₃ + 1/2 (24).

The pattern h〓〓 from which the directionality has been extracted is called a first-order feature pattern, and the pattern _h〓〓 which is blurred as is is called a zero-th order feature pattern.

The above processes and processes 5 and 6 are combined into process 20.
In other words, this is called feature extraction. (See Figure 1) Figure 8 shows the actual handwritten character pattern "0",
A pattern obtained by performing feature extraction on “1” and “2” is illustrated. In FIG. 8, h〓 ₀ at the right end indicates a pattern obtained by blurring the compressed pattern, and h〓 ₁ to h〓 ₈ indicate patterns obtained by performing blurring processing on the feature patterns for each direction.

The next process in recognition is hierarchical pattern matching,
(FIG. 1, Process 7). The pattern matching method (similarity method) is based on the feature extracted unknown pattern {h〓〓
(i, j) | θ = 0, 1, 2, ..., 8} and the standard pattern {H 〓 (i, j; k, ω) | θ = 0, 1, ...,
8; k=1, 2, . . . , K}, recognition is performed by determining the degree of similarity. Here, k is a serial number given to each standard pattern, K is the total number of standard patterns, and ω is a number indicating a category and is a function of k. That is, ω=ω(k) (25) In the field of handwritten characters, the types and amounts of character shape deformation are large, and the number of standard patterns that must be prepared for each category is large. Therefore, the number of categories is around 100 (numbers, alphabets, katakana, symbols)
In this case, the number K of standard patterns whose similarity must be calculated becomes considerably large.

On the other hand, since the recognition speed needs to be high, it is necessary to minimize the number of times similarity calculations are performed. It is known to introduce hierarchy to solve this problem. This method has a two-layer structure, and the 0th-order feature h〓 ₀ is used in the first layer to maximize the number of candidates.
The candidates are narrowed down to _Nc , and only the narrowed down candidates are brought to the final judgment using the primary feature amount {h〓〓|θ=1, 2,...,8}.

First, the processing in the first layer will be explained.

There are three parameters to be specified in the first layer: N _c : maximum number of candidates δ ₁ : absolute threshold ε ₁ : relative threshold, and the number of candidates can be controlled. The input to the first layer is the 0th-order feature h〓 ₀ created from the unknown pattern, and the output is the number of candidate characters N and the set of standard pattern numbers corresponding to each candidate {λ(ν)|
ν=1, 2,...,N}. The procedure for obtaining a candidate character set will be explained.

(S1.1) k = 1, N = 0, ρ _nax = 0, ρ _nio = 1
shall be.

(S1, 2) Calculate the similarity ρ ₁ (h〓 ₀ , H ₀ ) according to the following equation.

ρ ₁ (h〓 ₀ , H ₀ )=α ₀・β _0n3 〓 ⁱ⁼¹ _o2 〓 ^j=1 h〓 ₀ (i, j)・H ₀ (i,
j; k, ω) (26) However, α ₀ = [ _n3 〓 ⁱ⁼¹ _o3 〓 ^j=1 h ² ₀ (i, j)] - 1/2 (27) β ₀ = [ _n3 〓 ⁱ⁼¹ _o3 〓 ^j=1 H ² ₀ (i, j; k, ω)〕−1/2 (28) (S ₁ , 3ρ ₁ (h〓 ₀ , H ₀ )＜max(δ ₁ , ρ _nax −ε ₁ )
If so, go to (S1.7). Otherwise, go to (S1.4).

(S1.4) If N′>0, go to (S1.5).
When N'=0, N'=1 ρ _nax = ρ ₁ (h〓 ₀ , H ₀ ) ρ _nio = ρ ₁ (h〓 ₀ , H ₀ ) λ(N)=k, φ(N') =ρ ₁ (h〓 ₀ , H ₀ ) and go to (S1.7).

(S1.5) When N′=N _c , go to (S1.6).

When N′<N _c , N′=N′+1 λ(N′)=k, φ(N′)=ρ ₁ (h〓 ₀ , H ₀ ), and further ρ _nio >ρ ₁ (h〓 ₀ , H ₀ ), ρ _nio = ρ ₁ (h〓 ₀ , H ₀ ), and when ρ _nax <ρ ₁ (h〓 ₀ , H ₀ ), ρ _nax = ρ ₁ (h〓 ₀ , H ₀ ). Go to (S1.7).

(S1.6) When ρ _nio > ρ ₁ (h〓 ₀ , H ₀ ), (S1.7)
go to Otherwise, ρ _nio =φ(n)
_Find ^{n *} _that ^becomes _{_ c} }
Find the minimum value among and set it as ρ _nio . Furthermore, when ρ _nax <ρ ₁ (h〓 ₀ , H ₀ ), ρ _nax = ρ ₁ (h〓 ₀ , H ₀ ).

(S1.7) Let k=k+1. When k≦K, return to (S1.2). When k>K, proceed to the next step.

(S1.8) φ in {φ(ν) | ν=1,...N′}
Let λ(ν)=0 for all ν such that (ν)<ρ _nax −ε ₁ . Further, if the number of ν is n ₀ , then N=N'-n ₀ . Furthermore, we discard the element λ(ν)=0 from {λ(ν)|ν=1,...,N'} and rearrange the element again {λ(ν)|ν=1,...,N} shall be.

The above is the first layer processing, and the output is N and {λ
(ν)|ν=1,...,N}.

The processing of the second layer is performed in response to the output of the first layer. The output of the second layer is the maximum and dimensional similarity between the primary feature of the unknown pattern {h〓〓{i, j) | θ = 1, 2, ..., 8} and the standard pattern of the candidate character, and the standard This is the pattern number. Explain the steps.

(S2.1) Let ν=1, ρ _nax =0, ρ _nax2 =0.

(S2.2) Find the similarity ρ ₂ (h〓〓, H〓) using the following formula with k=λ(ν).

ρ ₂ (h〓〓, H〓)=1/8 ₈ 〓 〓 ⁼¹ aθ{α〓. β〓 _n3 〓 ⁱ⁼¹ _o3 〓 ^j=1 h〓〓(i,j)・H〓(i,j;k,ω} (29) However, here α〓＝〓[ _n3 〓 ⁱ⁼¹ _o3 〓 ^j=1 h〓 ₂ (i, j)〕 ^-1/2 (30) β〓=[ _n3 〓 ⁱ⁼¹ _o3 〓 ^j=1 H〓 ₂ (i, j; k, ω)] ^-1/2 (31) a〓≧0, ₈ 〓 〓 ⁼¹ a〓=1 (32)

(S2.3) When ρ ₂ (h〓〓, H〓)＜ρ _nax2 , (S2.6
). Otherwise, proceed to the next step.

(S2.4) When ρ ₂ (h〓〓, H〓)＜ρ _nax (S2.5)
go to Otherwise, set ρ _nax2 = P _nax k _nax2 = k _nax ρ _nax = ρ ₂ (h〓〓, H〓) k _nax = k and go to (S2.6).

(S2.5) ρ _nax2 = ρ ₂ (h〓〓, H〓) k _nax2 = k and go to (S2.6).

(S2.6) Let ν=ν+1. When ν>N, the process ends. Otherwise, return to (S2.2).

The above process is hierarchical pattern matching (FIG. 1, process 7). The amount of processing required to calculate the similarity per standard pattern in the first layer is 1/8 of that in the second layer, and the introduction of hierarchy can speed up the process by that much.

The final process is the final judgment performed using the maximum and dimensional similarities, ρ _nax and P _nax 2, which are the outputs of hierarchical pattern matching, and their corresponding standard pattern numbers k _nax and k _nax2 , and the output of the results. It is.

The final judgment is made as follows. Here, two parameters δ ₂ : absolute threshold and δ ₂ : relative threshold are used.

(Decision logic) If ρ _nax ≧max (δ ₂ , ρ _nax2 +ε ₂ ) (33) is satisfied, it is determined that the category to which the unknown pattern belongs is ω(k _nax ).

If Equation (33) is not satisfied and ρ _nax <δ ₂ (34), it is unrecognizable of the first kind, and if ρ _nax <ρ _nax +ε ₂ (35), it is unrecognizable of the second kind. The corresponding code is output in a predetermined format.

Finally, we will explain how to selectively and automatically determine a〓 in equation (29) from the blur pattern of the directional feature pattern of the input pattern and the standard pattern under the conditions given by equation (32). do.

From the blurred pattern of the directional feature pattern of the input pattern and the standard pattern, α〓, β〓 are calculated according to equations (30) and (31), and based on this, the 8-dimensional vector X for the input pattern is calculated. → and an eight-dimensional vector S→ for the standard pattern.

X→=(1/α ₁ , 1/α ₂ ,…, 1/α ₈ ) (36) S→=(1/β ₁ , 1/β ₂ ,…, 1/β ₈ ) (37) X→ The method of calculating the weight vector W→=(a ₁ , a ₂ , ..., a ₈ ) (38) from S→ and S→ is based on the type of pattern to be recognized,
Although it is determined by the nature of the object to be recognized, such as the number and quality, recognition experiments have confirmed that the following method is effective.

According to the method, it is possible to further compress the size of the less important weights among the weights for each direction, so it is possible to accurately classify input patterns as belonging to the same group as a group of standard patterns that have similar shapes. can do.

In other words, it can be said that this method is highly effective when applied to the pattern classification step in a normal hierarchical pattern matching method.

According to the method, it is possible to further emphasize the size of the weight with high importance among the weights for each direction, so the differences in the structure of each part between the input pattern and the standard pattern can be compared in detail to find out the essential differences. Based on
It is possible to accurately identify any of the standard patterns.

In other words, it can be said that this method is highly effective when applied to the pattern identification step in a normal hierarchical pattern matching method.

Next, the meaning of the weight vector W→ will be explained using FIG. In Figure 15, A, B, C,
D indicates an input pattern or a standard pattern, and at the same time, element 1/α〓 of the 8-dimensional vector This is what is shown. Assume now that B is an input pattern and A, B, C, and D are standard patterns.

In calculating the similarity between B and B, Equation (39) and (4
0) No matter which formula is used, of a〓, mainly a ₂ and
a ₆ is used, but when formula (40) is used to calculate the similarity between B and C, in addition to a ₂ and a ₆ , a ₃ and a ₇ also have large values, so the 3 and 7 directions are The difference in the patterns appears prominently in the similarity value obtained by equation (29).

In other words, 3 and 7 correspond to the different parts of B and C.
The magnitude of a ₃ and a ₇ in the direction is more emphasized, so a ₃
As the values of and a ₇ increase, the values of a ₂ and a ₆ , which correspond to the similar parts of B and C, become smaller, and moreover, a ₁
Since and a ₅ are coefficients for the correlation values of the different parts, the similarity in the 3 and 7 directions is close to 0, and the similarity in the 2 and 6 directions is close to 0.
As a result, the similarity of the entire pattern including the direction is sufficiently smaller than the similarity between B and B, and as a result, it is possible to more accurately determine that input pattern B is not C (B≠C).

On the other hand, when formula (39) is used to calculate the similarity between B and C, the smaller value of a ₃ and a ₇ is selected, so the value of similarity is mainly determined by the values of a ₂ and a ₆ . However, the 2nd and 6th directions correspond to similar parts of the pattern.

Therefore, the degree of similarity between the two patterns and the six-direction pattern becomes larger than the value obtained by the usual method, and the accuracy with which B and C are classified as patterns belonging to the same set is improved.

As another example, when formula (39) is used to calculate the similarity when the input pattern is A and the standard pattern is C, the similarity value is determined only by the values of a ₂ and a ₆ , but 2 and The six directions correspond to similar portions of the pattern.

Therefore, the degree of similarity between both patterns in the 2nd and 6th directions is greater than the value obtained by the normal method, and the accuracy with which patterns A and C are classified as belonging to the same set is improved. .

As is clear from the above explanation, the weight vector W →
The similarity value changes greatly depending on how you choose, so
By automatically selecting appropriate weights for each direction according to the unique characteristics of each direction of the input pattern and standard pattern, highly accurate classification and identification is possible.

In addition, the methods of selecting the weight vector are (39) and (4
It is also possible to obtain equation (41) by calculating the difference between the numerators of equation (0).

According to equation (41), the degree of similarity can be determined by extracting the weights for each direction regarding only the different portions between the input pattern and the standard pattern.

Although the character recognition method according to the present invention has been described above along the flow of recognition processing, the present invention is not limited to characters, but can also be applied to recognition of general patterns such as figures.

Hereinafter, the present invention will be explained in detail with reference to Examples.

FIG. 9 shows a block diagram of a character recognition device according to an embodiment of the present invention. The device of this embodiment is roughly divided into a microprocessor 10 that controls the entire system.
0, a preprocessing unit (PPU) 200 that performs preprocessing such as clipping, and a scanning mechanism 30 that observes characters.
0, a feature extraction unit (FEU) 400, a recognition processing unit (RPU) 500 that performs pattern matching, and an output device 600 that outputs recognition results and the like.

PPU, FEU, and RPU are composed of microprocessors with the same configuration. The microprocessor 100 basically includes MPUs (Microprocessing Units) 201 and 40 that perform basic calculations and control.
1,501; Memory MU (Memory Unit) 20
2,402,502; and an input/output control unit IOC-A (Input-
Output Control-A) 203, 403, 503
It consists of Special logic circuits required for each subsystem are connected to these basic elements.

Subsystem PPU (Preprocessing Unit) 20
The special logic circuits connected to 0 are the extraction auxiliary circuit SGU230 and the position and size normalization circuit SNU270. Furthermore, an input/output control circuit (type B) IOC-B 204 for controlling lower-level devices is connected to the PPU. Same IOC-
A scanning mechanism 300 is connected to B. The scanning mechanism 300 is a PHU (Paper
Handling Unit) 302 and SCU (Scanning Unit) 30 that scans the pattern on the form and converts it photoelectrically.
It consists of 1.

Subsystem FEU (Feature Extracting Unit)
A special logic circuit with 400 performs directional extraction.
FXU420 and BLU1 that performs blurring processing
450 and BLU2 460.

Subsystem RPU (Recognition Processing
Unit) 500 special logic circuits are high-speed memory HMU
An inner product calculation circuit IPU 530 is connected to 520. In addition, the multiplication circuit MPY51 is used to quickly execute multiplication when calculating similarity equations (18) and (21).
There is 0.

The recognition result output device in this embodiment is a floppy disk device 602. 601 is a control circuit FDC for the device. Further, the TCE 620 is a control circuit for connecting this character recognition device to a host computer.

The floppy disk 602 is a medium for outputting recognition results, and is also a secondary storage device for storing system programs, recognition programs, utility programs, and the like. These programs are loaded into the main memory 102 of the main microprocessor 100 as needed and executed by the CPU 101.

Control information from the operator to the entire system is input from the keyboard KBD104. Additionally, the system status is displayed on the CRT 630. The operator is
Various commands can be input from the KBD while viewing the CRT.

Next, the role of each circuit will be explained along with character recognition processing.

When the main power of the entire system is turned on, the state of each circuit is automatically reset to its initial state. Furthermore, the system in the floppy disk 602 is automatically
The program is loaded into the main microprocessor 100, and the CPU 101 automatically starts from a predetermined location. Commands for subsequent processing are input by the operator via the keyboard 103. To perform the recognition process, a command to load the recognition program from the floppy disk can be input from the keyboard. Thereafter, recognition processing is executed until another command is input.

The recognition program first passes through PPU200.
Sends a form feeding command to PHU302, and SCU30
1 to start scanning. These commands are macro commands and are issued to the PPU 200 when the CPU 101 executes a recognition program.
The microprocessor MPU201 of the PPU decodes the same instruction and converts it into a more micro-instruction sequence.
It is transmitted to PHU302 or SCU301.

The image pattern for one line scanned by SCU301 and subjected to photoelectric conversion, preprocessing, and binarization is IOC
-B204 and SGU (cutout auxiliary circuit) 23
0, and after discarding an unnecessary part of the video pattern for one line, it is stored in the memory MU202. The MPU 201 performs the extraction process according to the microprogram within the MPU 201 . Regarding the extraction process using SGU230, please refer to the patent application No. 53-81871.
The details are omitted as they are detailed in the issue.

The binarized pattern cut out from the video pattern for one line has a size of 64×48 (bits) for one character, and is stored again in the MU 202.

Next, normalize the position and size. Position normalization and size normalization are performed simultaneously by circuit SNU 270. Figure 10 shows a block diagram of the circuit SNU. 271 and 272 are G _j and G _i respectively
Each register consists of 5 bits. The values of G _i and G _j range from 0 to 4.00 in steps of 0.25. 277 and 278 are registers for holding the values J=JS+G _j・(j−1) and I=IS+G _i・(i−1), respectively, each consisting of 8 bits, and 0 in 0.25 increments. Takes values from ~63.75. 281 is a memory for holding a 64×48 (bit) cutout pattern f(i,j). Pattern writing to the memory 281 is performed using an address register 279 and a shift register 282 for parallel/serial conversion. 283,284
is a shift register that holds one column of the size normalized pattern g(i,j). Each consists of 16 bits. These outputs are routed to data bus 2 via bus drivers 285 and 286.
05 and is returned to the memory 202 (FIG. 9). The processing procedure will be explained below.

First, pattern f(i,j) is written into memory 281. The first time chart for writing operation
Shown in Figure 1. During a write, multiplexer 280 selects address register 279 by signal S202. Register 279 is first set to 0 by signal CL201. One cutout pattern has 192 words (one word is 16 bits) in the memory 202, and is loaded word by word into the parallel-to-serial conversion shift register 282 by the pulse LD 204. Next, the write pulse W201 to the memory 281 is applied 16 times, and the clocks CK201 and CK
202 and written into the memory 281. Pattern f(i,j) is written into memory 281 by repeating the above operation 192 times.

Position information i _s of the character part in f (i, j),
i _e , j _s , and j _e are determined by extraction processing. From these quantities, G _i , G _j , IS, and JS are calculated by MPU2 according to equations (1) to (4).
Calculate with 01. Data G _i and G _j (10 bits) are set in registers 272 and 271, respectively. Next, register 278 and register 277 have the initial value IS,
JS via multiplexers 276 and 275, respectively. In this state, the memory 28
The output signal 290 of No. 1 outputs the value of f(IS, JS). By issuing the clock CK203, the value f(IS, JS) is set as g(1,1) in the shift register 283.Next, when the pulse LD203 is output, I=I+G _i is executed by the adder circuit 274. signal 290
becomes f(IS+G _i , JS. When the above operation is repeated 32 times, the registers 283 and 284 have {g(i, 1)|i
= 1, 2, ..., 32}, 32 words will be set. Here, g(i,j) by signal T202
the first word of is placed on the bus 205 to the memory 202,
Subsequently, the second word is placed on the bus 205 and transferred to the memory 202 in response to the signal T201.

Next, while the register I278 is again initialized to the value IS, the addition circuit 273 executes the calculation J=J+G _j using the pulse LD202, and g(i,
The process starts in the second column of j).

If the process is repeated up to the 32nd column, the size normalization will be completed. A time chart of the control pulses in the above process is shown in FIG.

Pattern g(i,j) with size normalization
are temporarily stored in the memory 202 and then transferred to the main memory 102 via the data bus 105. The same pattern is subsystem FEU40
0 memory 402, where directionality extraction and blurring processing are performed. After the results are temporarily stored in the memory 402, they are collectively sent back to the main memory 102 as a characteristic pattern.

Here, the directionality extraction circuit will be explained. A block diagram of the same circuit is shown in FIG. The circuit includes a memory 422 that temporarily stores a binary pattern (32 x 32 bits) whose size has been normalized, an address register 421 for the memory, and a parallel-to-serial converter for transferring data to the memory. shift register 423,
Three shift registers 424, 425, 426 for taking out the data of the 3×3 small area, and a direction signal 44 for receiving the partial binary pattern of the small area.
2, a ROM (read-only memory) 427 that outputs the density signal 441, and a memory (8 x 8 x 4 x 8 bits) 43 that holds characteristic patterns for each direction.
2. An adder 434 that adds the density of each direction to the feature pattern of each direction one after another, a latch 435 that temporarily holds the result of the addition, a maximum value detection circuit 436 that detects the maximum value of the feature pattern of each direction, and the same circuit. It consists of a ROM 437 which inputs the maximum value detected and converts it into a value normalization coefficient, and a multiplication circuit 438 which normalizes the value by multiplying each value of the characteristic pattern by the normalization coefficient.

Next, the operation of the directionality extraction circuit will be explained.

First, the pattern g(i,
j) is transferred from the memory 402 to the memory 422 via the parallel-to-serial conversion shift register 423. The address of memory 422 is specified by address register 421. After transferring pattern g(i,j), address registers 429 and 430 of memory 432 are set to 0 by reset signal CL401.
Further, the internal state of the maximum value detection circuit 436 is reset by the signal CL404. In addition, the contents of address register 421 are also used as signal LD.
Enter 401 to reset to 0.

Next, all contents of the memory 432 are reset to 0 as follows. That is, the selection signal S401 is set to "0" to the multiplexers 431 and 433.
and reset the output of address register 428 to 431 and LOW to 433.
(value 0) is selected. The address register 428 is reset to 0 by outputting the signal CL402, and thereafter the 512 words in the output memory 430 are cleared to 0 by synchronizing the signals CK402 and W401 512 times.
After the above is completed, the signal S401 is set to "1" so that either multiplexer 431 or 433 can be selected.

Based on the above preparations, the directionality extraction operation is started. A waveform diagram of the same operation is shown in FIG. First, 68
Outputs time signals CK404 and CK405 synchronously,
Shift registers 424 (4 bits), 425
(32 bits) and 426 (32 bits), the first two lines and the third line 4 of the pattern in memory 422.
Transfer bits. Thereafter, by repeating the operation 1024 times or less, a directional feature pattern can be created in the memory 432.

That is, shift registers 424, 425,
If 426 contains a certain part of a pattern, eight points around a certain point of interest are input to ROM 427. In the ROM 427, the values of θ and v for all the above eight bit patterns are calculated and written in advance as shown in FIGS. 5 and 6, and when the bit patterns are input, Output θ and v. However, here, θ takes a value obtained by subtracting 1 from each value shown in FIG. 5, that is, a value from 0 to 7, and the signal 44
The most significant 3 bits of the address of memory 432 are set as 2. The other output of the ROM 427 is a signal 441, which is the value of v (0-7) shown in FIG. That is, give v=8 in Figure 6.
Since the combination of a ₄ ′ and a ₅ ′ is actually impossible, v
The value of is always between 0 and 7. The value of the signal 441 and the value of the past feature pattern of the same point of interest are added by an adding circuit 434 and temporarily held in a latch 435. The new value of the point of interest of the retained feature pattern is then written into the memory 432. At the same time, the output of latch 435 is also input to maximum value detection circuit 436, which detects the maximum value.

Through the above processing, the directional feature pattern is stored in the memory 43.
Create within 2. The value of each point of the created feature pattern can take a value of 0 to 255, but this is compressed to a range of 0 to 15 that can be expressed with 4 bits. This data value compression (value normalization) is performed by the multiplication circuit 438 when the characteristic pattern is transferred to the memory 402. The coefficients for compression are stored in ROM43 in the form of a table.
7 is calculated and written in advance.

The characteristic pattern is transferred from the memory 432 to the memory 402.
In order to transfer the data, the signal S401 is set to "0" again, the contents of the address register 428 are sequentially changed from 0 to 511, the contents of the memory 432 are read, and the bus driver 439 is opened at the same time.

With the above processing, the directional feature pattern h〓(i,j)
This means that θ=1, 2, . . . , 8 has been created. Therefore, blurring processing is performed next.

In Figure 9, 450 is h〓(i,j)(θ=1,
2,...,8) and h〓〓(i,j)(θ=1,
2, ..., 8). This circuit blurs a multi-value pattern (4 bits deep) to obtain a multi-value pattern (4 bits deep). In the figure, 460 is the pattern g
This is a blurring processing circuit for blurring (i, j) (32×32 mesh, binary) to obtain h〓 ₀ (i, j) (8×8 mesh, multi-value). Regarding the configuration of the blurring processing circuit, a patent application filed in 1983 by the same inventor is provided.
Since it is described in No. 81871, detailed explanation will be omitted.

The final feature pattern {h〓〓(i,j)|θ=0,1,...,8} obtained as a result of the blurring process is temporarily stored in the memory 402, and then collectively stored in the main memory 1.
Transferred to 02.

The CPU 101 detects that the feature pattern has been transferred from the feature extraction unit 400, and transfers the feature pattern to the input/output control circuit IOC-A5 of the RPU 500.
03 to the memory 502, and further
Issues a command to the RPU to execute recognition processing.

Most of the memory 502 is composed of a ROM (read-only memory), and all standard patterns are written in the ROM in advance.

Since the content of the recognition process performed by the RPU 500 has been described in detail in the general description of the invention section of this specification, the role of each block in FIG. 9 will be explained.
In the figure, the IPU 530 displays one of the blur patterns h〓(i,
j) and the standard pattern H in one direction
This is a high-speed inner product calculation circuit that quickly calculates the inner product of (i, j, k, ω). Unknown pattern h〓〓(i, j)
(θ = 0, 1, 2, ..., 8) is a high-speed memory HMU
520. Therefore, to calculate the inner product of h〓〓 and H〓 for a certain θ, use HMU520
The starting address of h〓〓 in MU502 and H〓 in MU502
and the starting address of
A start command may be given to the IPU 530.

In FIG. 9, MPY510 is a high-speed multiplier for multiplying the inner product calculated by IPU530 by coefficients α〓, β〓(θ=0, 1,...,8) as in equation (26) or (29). It is a multiplication circuit.

The recognition processing program described in the general explanation section is for microinstructions in the microprocessor 501.
Hierarchical pattern matching is performed according to the program written in the ROM.

The contents of this process are described in Japanese Patent Application No. 1987-68791, so a detailed explanation will be omitted.

As a result of recognition processing using the above circuit, for the given unknown pattern characteristic pattern, the first and second character category codes, and if unrecognizable, a code indicating type 1 unrecognizable. , or a code indicating the second type of unrecognizability is obtained and transferred to the main memory 102.

When the CPU 101 receives the recognition result transferred from the RPU 500, it writes the recognition result to the floppy disk 602 in a predetermined format.

All the processing for one unknown pattern is thus completed.

As can be seen from the embodiments described above, according to the present invention, it is possible to realize a character recognition device that reads handwritten characters using the same method as printed characters without sacrificing the advantages of the pattern matching method. I can do it.

The advantages of the pattern matching method using directionality can be summarized as follows.

(1) By introducing directionality and large blurring, various deformations of handwriting within the same category can be absorbed by large blurring, and at the same time, by dividing the pattern into different planes according to direction, confusion between different categories can be reduced. Overall, it is possible to recognize fewer standard patterns for each category with higher accuracy than before.

(2) When extracting the directionality, the directionality is extracted by spatial differentiation without line thinning, so the pattern matching method that introduces the conventional directionality, Patent Application No. 1983
-It is more accurate than No. 100377, and can even read low-quality characters (patterns that are blurred, etc.).

Furthermore, since the directionality is extracted by spatial differentiation, the character pattern does not need to be a binary pattern as in the past, and can be easily extended to a system that recognizes multivalued patterns as they are.
Conventionally, the result of photoelectric conversion (multi-value pattern) is
Because the process was converted into binary data to reduce the amount of processing, a large amount of information was lost, and low-quality characters, such as those written with a ballpoint pen, became blurred or cut off, which caused problems. Therefore, a recognition method that extracts directionality as is without performing binarization is effective.

(3) Since printed and handwritten character patterns can be recognized using the same method, hardware costs are lower than in the past.

(4) Since standard pattern creation can be done almost automatically without requiring any human intervention, development costs are lower than in the past.

(5) Conventional structural analysis methods are vulnerable to changes in topology such as cut-off, contact, blurring of characters, and protrusion of end points at some three-junction points, and standard patterns must be prepared for each deformation. Naturally, unpredictable phenomena such as "cutting" have been a problem. However, according to the present invention, even with these topological changes,
If the character shape as a whole resembles the character shape of that category, it can be recognized correctly. Therefore,
Compared to structural analysis methods, it can be said that this recognition method is more resistant to visual deformation as described above.

[Brief explanation of drawings]

Figure 1 is a process flowchart outlining the method of the present invention, Figure 2 is a diagram showing the principle of normalization of position and size, and Figure 3 is quantization of the direction obtained by spatial differentiation of a two-dimensional pattern. Figure 4 shows the directions obtained by spatial differentiation and the method for obtaining their strengths. Figures 5 and 6 show the vectors obtained by spatial differentiation. Figure 7 shows the value _of the point spreading function for the blurring process. Figure ₈ shows the value of the point spread function for the blurring process. The feature pattern h〓(θ=0,1,...,
In an example of 8), a is the number “0”, b is the number “1”,
c shows an example of the number "2", FIG. 9 is a block diagram of an embodiment of the method of the present invention, FIG. 10 is a detailed block diagram of the position and size normalization circuit 270,
FIG. 11 is a waveform diagram of the control signal for the operation of writing a pattern into the memory 281 of the size regular circuit;
The figure is a waveform diagram of the control signal when normalizing the size using the same size normalization circuit, Figure 13 is a block diagram of the circuit that performs directional feature extraction, and Figure 14 is the directional extraction. FIG. 15 is a waveform diagram of a control signal when the circuit performs directionality extraction, and is a diagram showing an example of an input pattern and a standard pattern, and an example of the norm in each direction.

Claims (1)

[Claims] 1. In a pattern recognition method using a pattern matching method, after cutting out unknown characters one by one from a two-dimensional character pattern obtained by photoelectric conversion, the size and position of each unknown character are normalized, Local inclination θ of the contour part of the normalized character pattern
(value θ is an integer from 1 to N) and the strength v of the contour (value v
is an integer from 1 to M), and then extracts N two-dimensional directional feature patterns h from the direction θ and intensity v.
(θ), {θ=1,2,...,N}, and further N
After creating N blurred directional patterns (θ) by performing convolution processing using a two-dimensional point spread function on each of the directional feature patterns, A standard blurring direction pattern H (θ; k, ω), which is stored in advance as a standard pattern with {θ=1, 2, ...,
N, k = 1, 2, ..., K} (k is a standard pattern number, ω is a function of k indicating a character category) and the blurring direction pattern (θ) is calculated, and character recognition is performed. A pattern recognition method that is characterized by performing the following. 2. In the pattern recognition method described in claim 1, pattern matching between the blurred directional pattern and the standard blurred directional pattern for unknown characters is determined by the similarity between the directional feature patterns in each corresponding direction θ. A pattern recognition method characterized in that each degree of similarity is multiplied by a predetermined weight and added to determine the degree of similarity of the entire character. 3 In the pattern recognition method described in claim 2, the weight amount in each predetermined direction is the square root of the sum of squares of the values of each point of the standard pattern in the predetermined direction, and the directionality in which the blurring process has been performed. A pattern recognition method characterized in that a value corresponding to the maximum value among the square roots of the sum of squares of the values of each point of a feature pattern is set. 4. In the pattern recognition method described in claim 2, the weight amount in each predetermined direction is the square root of the sum of squares of the values of each point of the standard pattern in the predetermined direction, and the directionality in which the blurring process has been performed. A pattern recognition method characterized in that the method corresponds to the minimum value among the square roots of the sum of squares of the values of each point of a feature pattern. 5. In the pattern recognition method described in claim 2, the weight amount for each predetermined direction is the square root of the sum of squares of the values of each point of the standard pattern in the predetermined direction, and the directionality in which the blurring process has been performed. A pattern recognition method characterized in that the value corresponds to the difference between the maximum value and the minimum value among the square roots of the sum of squares of the values of each point of a characteristic pattern.