JPS61156486A - Picture processing device - Google Patents

Abstract

PURPOSE:To realize a stable pattern recognition by mounting a means which recognizes a picture image by non-linear matching a coded code hole with a standard code hole. CONSTITUTION:A memory writing circuit 101 reads any area specified beforehand in an input original through an inputting device 100 to write it in a main memory 102, and then a pre-processing circuit 103 carries out segmenting a character area and binarizing. Moreover, a primary- and secondary-feature extracting circuit 104 traces a contour and extracts contour points, as well as secondary features from code rows depending on its coding and necessity. A non-linear matching circuit 105 calculates differences between the features and respective standard pattern of input patterns. The above-mentioned each circuit is connected on a pipeline to obtain high-speed recognition. Thus it becomes possible to realize a high-speed and stable recognizing device which has a simple configuration and is suitable for various recognition objects.

Description

DETAILED DESCRIPTION OF THE INVENTION [Technical Field] The present invention relates to an image processing device capable of recognizing handwritten characters, printed characters, and all other patterns.

[Prior art]

There have been various proposals for pattern recognition in the past.0 For example, when performing pattern recognition using contour information of a pattern, the contour line direction at each contour point is determined by 1 to 8 shown in Figure 1.
The code was obtained as a quantized code according to the direction code.

However, with this method, if there is a pattern as shown in Figure 2, for example, the direction code obtained when tracing the contour clockwise from 51 in the figure as the starting point is as shown in Figure 3. The direction code, which originally changes continuously, becomes discontinuous in the code, as shown in parts A and B in Figure 3 (corresponding to a and b in Figure 2), which causes inconvenience during recognition. arise. .

This means that changes in the starting point due to rotation of the original pattern, etc., make subsequent recognition difficult, which has the disadvantage of increasing processing time.

〔the purpose〕

In view of the above points, the present invention eliminates the above-mentioned drawbacks, captures the outline of a pattern as continuous data, and provides an image processing device that is capable of stable pattern recognition unaffected by pattern positional deviation, changes in rotation size, etc. It is about providing.

〔Example〕

Hereinafter, the present invention will be described in detail with reference to the drawings. Images in the present application indicate characters, marks, symbols, pictures, etc.

FIGS. 4(a) and 4(b) are diagrams showing an example of an image processing apparatus to which the present invention can be applied. 1 is league, 2 is OCR
, league, or printer, 3 is a printer, and 4 is a computer with a display, which can issue control commands for each device and display desired data.

5 is a keyboard.

Further, in (b), IO is a main body of a hand-held image processing apparatus, 11 is a light source, and 12 is an original to be read. Note 1
3 is an optical system that guides the light 1itll and the light from the pattern on the original 12 to a photoelectric conversion element (for example, planar C0D) 14; 15 is an optical system that processes output data from the photoelectric conversion element 14 and/or
Or it is a processing unit that performs output. Note that it is not necessary to limit the structure to this kind. Alternatively, it may be one of the systems equipped with a disk device, microfilm, workstation, etc.

Next, the processing section 15 will be explained in detail. 5th-
FIG. 1 is a block diagram of the processing section 15. A preprocessing circuit 16 performs binarization, noise removal, etc. on the input pattern. Next, the CPU 17 and memory 18 extract features of the pattern. Regarding feature extraction, the flow is shown within the broken line in Fig. 5-2. That is, the contour information of the pattern is encoded by performing primary feature extraction. Depending on the pattern or extraction method, the code is used as is. Further, features other than the code string may be extracted by secondary feature extraction, and the features obtained by the above may be checked against a dictionary in the memory 18 or the like, or similarities and correlations may be calculated. , the pattern may be determined.

Furthermore, the above feature extraction may be repeated, or other feature extraction may be performed and then a pattern may be determined.

Next, feature extraction according to the present invention will be explained in more detail. In primary feature extraction, a predetermined number of pixels forming the outline of a pattern are first extracted.・
Next, the directional differences in the contour tangential directions are determined based on the extracted points and are sequentially coded.

A flowchart of contour point extraction is shown in FIG.

That is, first, the contour tracing starting point is determined by a method such as scanning the image by rask-scanning and first determining the pixels that correspond to the pattern, and the position is stored (STEPI). Next, find the contour point adjacent to the right side, for example, on the contour (STEP 3),
Find a distance statement from the starting point (STEP 4). The distance statement from the starting point is accumulated according to the following equations (1) and (1)', for example.

A poem in which adjacent contour points are located on the top, bottom, left, and right = Mi +1 (1) A poem in which adjacent contour points are located in a diagonal direction = Sentence +1.4 (1) From the above, the length L of one circumference of the contour can be found up to 5TEP6. Next, based on this value, a certain number N of contour points are extracted. First, the starting point is selected as the extraction point (ST
EP8), find adjacent contour points and a long contour sentence from their starting point (STEP II, 12). Here, the method of finding the sentence is to extract contour points as it progresses, and to judge the following conditions (ST
EP 13).

Here, N is a predetermined number of contour points to be extracted. If the formula (2) holds true, return to 5TEP8 and extract the contour points, and while it does not hold, 5TEPII
, 12 are repeated.

Then, the process ends when N contour points are extracted (STE
PIO).

If the above method is applied to the pattern shown in FIG. 7, for example, the pixels (S1 to 520) indicated by diagonal lines in the figure are extracted. However, N=20, the contour tracing starting point was Sl, and the contour was traced clockwise.

Specifically, in the pattern shown in FIG. 7, the total length of the pattern is L=98.6 and N=20, so the side of one pixel is 1. Calculate with the diagonal as 1.4. The numerical value shown within the pixel in FIG. 7 is the distance. That is, from Sl to 3
When proceeding to two pixels (indicated by diagonal lines), the distance becomes 5.4, and in step 13, since 5.4 > 4.93, the shaded pixel is set as the extraction point S2. In the same way, as shown in Fig. 7, up to the next extraction point, it progresses in the order of 6.8, 7.8, 9.2, and at the next diagonally shaded part, the value becomes 10.6. In step 13 of FIG. 6, 10.6 > 4.93 x 2, and this pixel is extracted as an extraction point (S3). Similarly, extraction points 51 to S2 shown in FIG.
0 is extracted.

Next, the direction difference in the contour tangent direction is sequentially determined based on the contour extraction points and coded. For example, when the contour extraction points are lined up as shown in Figure 8, the direction difference in the contour tangent direction is α in the figure. ,
For example, as shown in FIG. 9, it is quantized in eight directions and coded. Here, the first contour tangent direction (5iSit1 in FIG. 8) is always made to match the code 0 direction. In this way, by encoding the direction difference in the contour tangent direction,
It is possible to obtain a code string that changes periodically no matter how many times it goes around the contour without causing any unreasonable discontinuity points. - As an example, FIG. 10 shows a code string obtained using all the contour points of the pattern in FIG. 2. As is clear from comparing it with FIG. It is suitable for identification processing, and is stable against changes in the starting point of the code.

FIG. 11 shows the result of performing the processing detailed above on the pattern shown in FIG. 7. The code string shown in parentheses in the figure expresses the contour shape well without being affected by the fine irregularities of the pattern, and its length is also normalized.

Next, the determination section first performs non-linear matching between the code string obtained from the next feature extraction and the standard code string stored in memory in advance to determine the similarity between the input cover pattern and the category to which the standard code string belongs. Calculate the distance D (the more similar the value, the smaller the value). The calculation formula is shown below. The code string of a certain input pattern is air
a2 -------, aN When the standard code string is cl, c2 -------, CN, calculate the following decomposition formula. However, d(Ci, aj)=lci-ajl+f(lci -aj
l) (4) Here, d(ci, aj) is the distance between the standard code ci and the input code aj, and f(lci-a
N) corresponds to a penalty for the positional shift of the corresponding code on the code string, and is determined, for example, as shown in FIG.

Finally, the distance @D between the input turn and the category is found as D=g(N,N) (5).

For example, if the input code based on the input pattern shown in Figure 11 and a certain standard code are correlated as shown in Figure 13 (this correspondence is expressed as (3), then calculate equation (4) to the end). The distance d (ci, aD) between each code is as shown in FIG. 14, and the sum thereof is equal to D.

Next, referring to the drawings, regarding the calculation and matching of D,
Explain in detail.

As shown in FIG. 11, the input pattern code string is a
j=1. l, 1, 0, 0. -3.0,1,3,0,0°3.0,0,0. -2. -2.0, 4, 1. It is clear that The code string of the standard pattern for one character “2°” is ci=1.1.O,O,-2,0,0,3,
0,0°2,0,0,0. -1. -1.0, 3, 1.1
Then, the method for calculating D becomes the flow shown in FIG. 15.

That is, in equation (3), g(1,1)=d(c,a) is cl=1. a1=
1, g (1, 1) = O g (1, 2) = d (c4, az) is c
1=l, a2=2, so g(1,2)=d(cl,a
z) + min (g (1, 1)) = 0.1 Similarly, g (2, 2) = d (c2.az) + min (g
(1,2) g(2,1) , g(1,1)) = I
c2-a21+f (I c2-az l) +mi
n (0, 1, 0, 1, 0) =O.

Note that f (lci-ajl) is i and j shown in Figure 12.
(corresponding sampling points). Below, when t+J is performed from 1 to 20, g (i, j) at 1=j=20 as shown in the upper right of Figure 15.
The value D = 8.1 is calculated.

On the other hand, for the same pattern shown in FIG. 7, the standard pattern is changed to the standard pattern of "°3" (ci=1.1.-2.
1, 1, 1, 1, 3, 0. -1. -1. -1.0°2.
0. -1. -1.0, 3.1), the result is shown in FIG. In other words, when using a standard pattern of 3'', g (i, j) at 1=j=20
The value of is calculated as 15.2. As is clear from the above, for the pattern shown in Figure 7, the standard pattern °2'' is 8.1 < 15 compared to the standard pattern ゛3pa.
．． Since the distance is 2, it can be determined that the distance is short, and the pattern shown in FIG. 7 has a high probability of being determined to be 2.

Also, what is the correspondence between each extracted point of standard pattern "2" and each extracted point of the pattern shown in FIG.

As shown by the thick line arrow in Figure 15, the correspondence is as shown in Figure 17.As is clear from Figure 0, there is no unreasonable correspondence at each position. Recognize.

Also, if necessary, change the extraction point selection conditions shown in Figure 6.
The accuracy may be further improved by extracting the shifted extraction points multiple times. Hereinafter, for the character pattern "3", the above-mentioned extraction points are shifted and various distances of 1!10 (
A method that allows for more accurate recognition by calculating Equation 5) will be explained.

The contour point extraction method in that case is shown in FIG.

Note that the details are the same as those in FIG. 6 and will therefore be omitted. Here, 8 shown in step 8 and step 16 is an offset amount from the first contour extraction point (a value indicating how much it deviates from the first contour extraction point), and is a value of 0≦δkl. . (If δ = 0, it matches the first contour extraction point.) For example, when there is an input cover turn as shown in Figure 19-1, the first contour extraction point and the direction found from it. The code becomes as shown in Figure 19-2.

On the other hand, when the contour extraction is redone with δ=touch and the code string is obtained, the result is as shown in FIG. 19-3. In other words,
As you can see by comparing the position of the first extraction point with Figure 19-2, the unevenness of the pattern outline is faithfully reflected in the code (the location indicated by "→" in Figure 19-3). Therefore, the probability of correct recognition increases significantly through subsequent processing. In FIG. 2, only two are detected, but in FIG. 19-3, three are detected faithfully to the shape of the pattern, making it much more likely that accurate recognition will be performed. (Refer to the ↑ marks above the code strings in Figures 15-2 and 15-3.) Note that the value of δ is not limited to stations; for example, 115,
215.

315, 415, etc., the contour extraction may be repeated while changing it little by little, allowing for more accurate recognition.

Furthermore, the determination results obtained through the series of processes described above may be outputted as recognition results as they are, or may be stored as candidate categories and then proceed to another recognition process.

Next, we will explain the case where the distance D (Equation 5) is calculated using the code string of the pattern shown in Figures 19-2 and 19-3 and the standard code of the character "°3゛°." Even more accurate recognition is possible by simply changing the
0 and 21 are standard code strings of °゛3''9 ct=1,
1. -2.1,1゜1.1,3,0. -1. -1. -1
．． 0, 2, 0. -1. −1.0,3゜l and °“3 pa input turn code string aj=l, 0゜1.0,0,0,1
,1,0,4,0. -2. -1. -1.0,0゜-1,
-2, 0, 3, when the distance and extraction point are shifted °
“3° Input Cover Turn Code Row” J=L1+L L
nO, 1, 1, l, 1, 3, 0. -1. -1. -1.2
．． -1. It is an explanatory diagram when calculating the distance from -1°-1, 3, 1.

The method for determining the distance has been explained in FIGS. 15 and 16, so the details will be omitted. In the end, the distance changes from 17.5 to 8, making it possible to recognize the pattern more accurately. Also, by looking at the ↑ marks in FIGS. 20 and 21, it can be seen that each pixel is extracted as shown in (a) and Cb) of FIG. 22, respectively.

Note that it is not necessary to use this distance as an aid for pattern determination; similarity or any other determination method may be used. In other words, the above basically used a method called dynamic programming (D, P), but by allowing positional shifts and multiple correspondence in the code,
For example, it realizes nonlinear matching that can absorb deformations such as handwritten characters. Furthermore, in the present invention, by considering positional deviation on the code as an evaluation amount (penalty) as shown in FIG. 12, it is possible to eliminate inappropriate correspondence (so-called overmatching) while allowing deformation. , enabling accurate and fast recognition.

Then, for each standard code, a distance representing pattern similarity is calculated, and an incoming cover turn is determined according to, for example, the following conditions. In other words, when the smallest distance for a certain input cover turn is Dl, and the next smallest distance is D2, if both the following conditions ■ and ■ are satisfied, then the input cover turn belongs to the category corresponding to Dl. Otherwise, it is determined that the corresponding category does not exist.

■ Di < Kl ■ D2 - DI > K2 or the categories corresponding to Dl and D2 are equal, however, Kl and 2 are constants determined in advance depending on the degree of deformation of the recognition target and the degree of confusion between categories. be.

Furthermore, the above determination result may be output as is as a recognition result, or may be stored as a candidate category and proceed to another recognition process.

In the above embodiment, the contour point extraction method is as shown in FIGS. 6 and 18.
The method is not limited to the one shown in the figure, but any method that extracts a certain number of pixels that represent the characteristics of the pattern shape without being affected by small irregularities or noise in the contour shape of the pattern can be used. It can be applied without affecting the processing at all. Also shown in Figure 9 with code 0 7
I + 4 lh Yuμ 矧 烏 4 h,
1ヂ yen 1--μ Ll Xiang Ko, Ku(
For example, as shown in FIG. 23) or roughening is effective.

In addition, a penalty (k) for multiple correspondences is added, such as a treasure ship where k > 1, or a penalty for the positional shift of the code shown in Figure 12. By changing the constraint conditions for deformation,
Matching suitable for various recognition targets can be performed.

It goes without saying that the methods described above may be repeated or may be combined.

Next, an example of the hardware configuration of the method described above is shown in FIG.
(digital copy reader in the figure) and writes it to the main memory 102, then the preprocessing circuit 1
03, the character area is cut out and binarized. Furthermore, the primary and secondary feature extraction circuit 104 performs contour tracing and contour point extraction, and also encodes the contour and extracts secondary features from the code string as necessary. And nonlinear matching circuit 1
According to 05, each standard pattern (code) of the input cover turn
A high recognition speed can be obtained by connecting each circuit whose degree of difference is 0 or more from 0 to 1 on a pipeline.

In addition, the control C0P and U control each of the above-mentioned circuits, and further output the final recognition effect to the external device 106 based on the dissimilarity value obtained from the nonlinear matching circuit 105. , for the secondary feature extraction circuit 104,
Instructs to re-encode.

〔effect〕

As described in detail above, according to the present invention, a certain number of contour points of a pattern are extracted, and based on this, the direction difference in the contour tangent direction is determined and coded, and the above code is compared with a standard code stored in memory in advance. By performing non-linear matching to calculate the distance representing the similarity between the input cover pattern and the category to which the standard code belongs, taking into account the degree of pattern deformation, and recognizing the input cover pattern, various recognition methods can be achieved. It has become possible to realize a recognition device that is suitable for the target, fast, and stable with a simple configuration.

[Brief explanation of the drawing]

FIG. 1 is a diagram showing a conventional direction code, FIG. 2 is a diagram showing an example of a pattern, FIG. 3 is a diagram showing the relationship between contour points and direction codes, and FIG. 4 is an image to which the present invention can be applied. FIG. 1 is a diagram showing an example of a processing system. FIG. 5-1 is a block diagram of the processing section. Figure 5-2 is a diagram showing the flow of pattern feature extraction;
The figure shows the contour point extraction method in the primary feature extraction section.
FIG. 7 is a diagram showing an example of a pattern and its contour extraction points, FIG. 8 is a diagram showing the relationship between the contour extraction points and the difference in the direction of the contour tangent, and FIG. A diagram showing the direction code, FIG. 1O is a diagram showing the relationship between contour points and the direction code, FIG. 11 is a diagram showing the direction code according to the present invention,
FIG. 12 is a diagram showing a penalty for positional deviation of codes used when calculating pattern distances by performing nonlinear matching, and FIG. 13 is a diagram showing an example of code correspondence.
Figure 14 is a diagram showing the distance between codes for the correspondence shown in Figure 13, and Figure 15 is a calculation of the distance between the code string of the input pattern shown in Figure 11 and the code string of the standard pattern of character 2-pa. An explanatory diagram of the method, Fig. 16 is an explanatory diagram of the method for calculating the distance #D between the code string of the input pattern shown in Fig. 11 and the code string of the standard pattern of character “°3゛”, and Fig. 17 is an explanatory diagram of the standard pattern Figure 18 showing the correspondence between 2 and the input cover pattern.
The figure is a flowchart for selecting extraction points by shifting them.
Figure 19-1 is a diagram showing an example of an input cover pattern, Figure 19-1
Figure 2 shows an example of encoding the input pattern shown in Figure 19-1, and Figure 19-3 shows the encoding shown in Figure 19-2. Figure 20 is a diagram showing how to calculate the distance between the code string shown in Figure 19-2 and the standard pattern code string of the character "3". Figure 21 is a diagram showing a method for calculating the distance between the code string shown in Figure 19-3 and the code string of the standard pattern of the character "3", and Figure 22 is a diagram showing the method for calculating the distance between the code string shown in Figure 19-3 and the code string of the standard pattern of the character "3".
1, FIG. 23 is a diagram showing subdivided direction codes, and FIG. 24 is a system control block diagram of the present invention. 106 is a control CPU 105 is a nonlinear matching circuit 104 is a feature extraction circuit

Claims (1)

[Claims] Detection means for tracing contour information of an image pattern in a predetermined direction to detect contour points; Extraction means for selecting a predetermined number of contour points from among the contour points detected by the detection means; Based on the extraction points. Sequentially find the directional difference in the contour tangent direction,
An image processing apparatus comprising: a coding means; and a recognition means for performing non-linear matching between the code string encoded by the coding means and a standard code string to recognize the image pattern.