JPS6175979A - Pattern recognizer - Google Patents

Abstract

PURPOSE:To attain pattern recognition at a high speed and with high accuracy by discriminating the peripheral picture elements of a mask and smoothing the input picture elements in case no anisotropic stroke is obtained. CONSTITUTION:The shift registers 1, 2 and 3 which sets a mask consisting of vertical and horizontal bits and scans it collates a partial pattern with a reference pattern set to the 1st ROM6 through this ROM6. When the coincidence output is detected, the outputs of two shift registers 4 and 5 which detect the peripheral picture elements of the mask are supplied to the 2nd ROM8 together with data on the peripheral picture elements of registers 1-3. In case no anisotropic stroke is formed, the correction data is supplied to an image memory 10 via a gate circuit 15. Thus it is possible to avoid the deformation of an input image owing to a smoothing process.

Description

[Detailed Description of the Invention] <Technical Field of the Invention> The present invention optically reads unknown characters and symbols (hereinafter referred to as "unknown characters"), converts them into black and white, and obtains an input image. , relates to a pattern recognition device that identifies and recognizes unknown characters by extracting features of an input image and comparing them with a standard pattern.

<Background of the Invention> As shown in FIG. 4, the conventional pattern recognition device includes a reading unit 22 that optically reads unknown characters 21 and converts the image into an image, and performs smoothing or other processing on the input image from the reading unit 22. A preprocessing unit 23 performs preprocessing, a feature extraction unit 24 extracts features of the preprocessed one-plane image, and a dictionary 2 stores the extracted features in advance.
The unknown character is matched against the standard pattern stored in 5.
It is composed of a dictionary collation unit 26 that performs ll and la.

Generally, the feature extraction unit 24 extracts features of unknown characters by focusing on black pixels of the input image and extracting features of unknown characters by focusing on white pixels (corresponding to the background part) of the input image. However, in the case of the former method, as shown in FIG. 5, the conventional method was to thin the input image G and then extract the features of the unknown character from this thin line pattern g. . However, this method requires image thinning processing, and in recent years, a method has been proposed in which the features of unknown characters are directly extracted from the input image. In this method, as shown in Fig. 6, the boundary between the black pixel and the white pixel in the input image G (indicated by the thick line in the figure) is observed, and the direction in which this boundary extends is shown in Fig. 7. , D, the paired strokes (AL I A2) (B
1. B2 ) (CI, C2) and extract the 8th
An approximate pattern E as shown in the figure is obtained.

In this method, as shown in FIG.
is scanned in a certain direction (indicated by an arrow in the figure), and during this scanning process, when a preset tracking pattern in either direction A or D is encountered, boundary tracking is started in the corresponding direction. Then, the stroke is extracted by continuing tracking until a matching tracking pattern in the same direction disappears.

FIG. 10 shows the tracking patterns for the incoming direction, all these patterns being stored in the pattern storage means along with the tracking patterns for the other directions. Therefore, in boundary tracking processing, it is necessary to match all tracking patterns in the tracking direction, and the larger the number of tracking patterns, the longer the matching process takes, which reduces pattern recognition speed.

In order to solve this problem, the inventor recently reduced the number of tracking patterns in the A and B directions to the necessary minimum and shortened the pattern matching time in the boundary tracking process, thereby speeding up the pattern recognition process. We have developed a new pattern recognition method that measures

FIG. 11 and FIG. 12 are diagrams specifically explaining this method.

FIG. 11 shows an image memory 2 in which an input image 11G is stored.
7 is enlarged, and in the figure, the tea juice between the scales corresponds to one pixel, and this image memory 27 is composed of pixels of a plurality of bits in the vertical and horizontal directions. Further, in the figure, the shaded area indicates black pixels forming the input image G, and the other areas indicate white pixels forming the background, and the boundary between the black pixel and the white pixel is indicated by a thick line in the figure.

This newly developed method sets a rectangular mask 28 composed of CI (3 pixels each vertically and horizontally) in the image memory 27 prior to boundary tracking processing of the input image G, and shifts this mask 28 line by line while tracking the input image G. This scans all rows.

In this scanning process, partial patterns of 9 bits included in the mask 28 are sequentially extracted and compared with a preset reference pattern. When both patterns match, a correction output is output and the input image G is Smooth.

The diagrams on the left side of FIGS. 12(1) to (8) show the reference pattern, and the diagrams on the right side of the same diagrams show the smoothed corrected image. The standard pattern is the 12th number 111 + 2)
+51 +6), the black pixel 29 that protrudes by 1 bit (hereinafter referred to as the "protrusion") or the black pixel 29 in FIG.
As shown in 3+ 14117+ +81, it has a white pixel 30 (hereinafter referred to as a "concavity") that is recessed by 1 bit, and when the partial pattern in the mask 28 matches any of the reference patterns, the convex part of the partial pattern The black pixels forming the recess are corrected to white pixels, and the white pixels forming the recess are corrected to black pixels.

In Figure 12, the X marks indicate bits from which black pixels have been removed in the input image, and the O marks indicate bits to which black pixels have been added.As a result, in the corrected image, the boundaries between white pixels and black pixels are straight and smooth. be converted into

Thus, Fig. 12 shows the tracking patterns () corresponding to ) to (8).
Figure 10! 21 t3+ +71 (81 corresponds to this) can be reduced, and the time for pattern matching in boundary tracking processing can be shortened.

However, it has been found that if the smoothing process is unconditionally performed, the corrected input image may be deformed and its characteristics may change. For example, in FIG.
When the mask 28 is positioned at the portion indicated by 1, a partial pattern having concave portions to be smoothed appears within the mask 28. However, this part 31 is an important part that makes up the D direction stroke in the input image G, and if this is corrected using the above method, the D direction stroke will not be extracted. There is an inconvenience that

<Object of the Invention> The present invention provides a novel method that smoothes an input image to reduce the number of tracking patterns in the A and B directions to the necessary minimum, and prevents deformation of the input image due to the smoothing process. An object of the present invention is to provide a pattern recognition device, thereby increasing the speed and accuracy of pattern recognition processing.

<Structure and Effects of the Invention> In order to achieve the above object, the present invention scans an input image with a mask as preprocessing for boundary tracking, and scans an input image with a mask so that a partial pattern in the mask matches a predetermined reference pattern having uneven parts. When doing so, a matching output is sent out, and when there is this matching output, peripheral pixels of the mask are determined, and when forming a stroke in the C and D directions between this and the partial pattern, the input image is without performing smoothing processing,
On the other hand, when the C and D direction strokes are not configured, a correction output is sent out to flatten and ossify the input image.

According to the present invention, among the tracking patterns in the A and B directions, it is possible to reduce the α uneven tracking patterns as shown in FIG. It is possible to prevent deformation of the input image without disturbing the constituent pixels of the input image. As a result, the number of traced patterns can be reduced, pattern matching time in boundary tracking can be shortened, pattern recognition processing can be made faster, and productivity can be improved.

<Explanation of Examples> FIG. 1 shows an example of a device according to the invention.

In the illustrated example, the mask 28 has a configuration of 3 bits each in the vertical and horizontal directions.
Three shift registers 1 and 2 for setting and scanning
．． 3 and two shift registers 4.5 added to detect the pixels on the four sides of the mask 28, a reference pattern is set, and the partial pattern of the mask 28 circle is used to match the reference pattern. The first ROM (Ready
0nly Memory) 5, an address register 7 that accesses the address of the ROM 6 according to the output of the shift register 1.2.3, and a correction output for smoothing the input image according to the determination result of peripheral pixels. an address register 9 that accesses the address of the second ROM 13 according to the outputs of the first ROM 5 and the soft register 1.3.4.5, and an image memory that stores the correction output, etc. NOT circuit 11, AND circuit 12, 13, and OR circuit 1 that supply to 10
4, and a gate circuit 15 including 4.

Each of the five shift registers 1, 2, 3, 4.5 has a number of bits equivalent to one row of the image memory 10, and each of the shift registers 1 to 5 has a total of 5 bits in the stroke number memory 10.
Image data for a row is set. From the three-bar soft registers 1, 2, and 3 associated with the mask 28, the output of the first three bits is sent simultaneously to the address register.
The contents of the address register 7 are sequentially updated in response to the simultaneous shift operation of each shift register. The total 9-bit data content in the address register 7 is determined by the mask 28.
This corresponds to the content of the partial pattern in the ROM 6, and also corresponds to the address of the first ROM 6 accessed by the address register 7. In this way, the addresses of the first ROM 6 corresponding to the data contents of the reference pattern are as shown in FIG.
) is stored, and when the partial pattern in the mask 28 matches the reference pattern shown in FIG. 12 (1) to (8), the format shown in FIG. 2 is stored from the corresponding address in the ROM 5. data is output.

Figure 2 (here, the 0th to 2nd bits (here, the 1st
Data specifying each of the reference patterns in Figures 2 (1) to (8) is set. For example, the reference pattern in Figure 12 (1) is rooOJ, and the reference pattern in Figure 12 (2) is r.
The reference pattern shown in FIG. 12 (8), hereinafter referred to as oolJ, has the same data structure as rllllJ. In addition, the third bit (here, data indicating whether the central pixel in the mask 28 is black or white (“1” for a black pixel, “0” for a white pixel) is set, and the fourth bit The bit contains data indicating whether or not the partial pattern in the mask 28 matches the basic Buddha pattern shown in FIG. Set.

Of the configuration data (above), the data of the 0th to 2nd bits is sent to the address register 9 that accesses the second ROM 8.
The third and fourth bits of data are output to the gate circuits 15, respectively, and the address register 9 contains, in addition to the above data, data P1 to P1 of predetermined bits in the shift registers 1.3 and 4.5. P12 has been input. These data indicate whether a predetermined pixel around the mask 28 is a black or white pixel; for example, data P
5. P6 is data indicating the pixel state at the position shown in FIG. 3 (13).

The figure on the left side of Figure 3 1) to (8) shows the reference pattern, and the figure on the right side of the same figure shows the peripheral pixel positions to be checked for each reference pattern. P5 of peripheral pixel data (for example, FIG. 3, 13). P6
) are white pixels, it is determined that the partial pattern within the mask 28 does not constitute a stroke in the C-D direction with its surrounding surface pixels; on the other hand, when any of them is a black pixel, the mask 28 It is determined that the partial pattern within 28 constitutes a C- and D-direction streak with its surrounding pixels.

Thus, the output contents of the ROM 5 and the data 9 P1~
P L2 (7) Contents tz vs. 5t le 2nd R,O
Data specifying black and white for the central one-sided element in the mask 28 is set in the MB address. For example, if the partial pattern within the mask 28 matches the third r3+3+ reference pattern and the peripheral pixel data P5. When any of P6 is a white pixel, the corresponding address of ROM8 is
Correction data for converting the center pixel (corresponding to the recess) in the mask 28 into a black pixel is stored. On the other hand, peripheral pixel data P5. When any of P6 is a black pixel, the corresponding address of ROM 8 is filled with mask 28.
The data that maintains the center pixel within the pixel as a white pixel is stored.

These data in the ROM 9 are output to the image memory 10 via the AND circuit 13 of the gate circuit 15 and the OR circuit 14. The logic of the AND circuit 13 is established only when the data of the fourth bit of the output data of the first ROM 5 is "1", that is, when the partial pattern in the mask 28 matches any reference pattern. The logic of the other AND circuit 12 is established when the data of the fourth bit is rOJ, that is, when the partial pattern and the reference pattern do not match, and in this case, the center pixel data in the mask 28 (3 The data of the cursor hits) is supplied to the image memory 10 as is.

By the above operation, the mask 28 forming a convex portion or a concave portion
The central pixel in the image is subjected to black-and-white conversion hR under predetermined conditions, thereby flattening and ossifying the input pixel.

[Brief explanation of the drawing]

FIG. 1 is a circuit block diagram of an example of a sediment storage system according to the present invention, and FIG.
The figure is an explanatory diagram showing the format of data stored in the first RO^4, Fig. 3 is an explanatory diagram showing the relationship between the reference pattern and the surrounding pixels of the mask, and Fig. 4 is an explanatory diagram showing the overall configuration of the pattern recognition device. Block diagram; FIG. 5 is an explanatory diagram showing thinning processing of an input image in the conventional method; FIG. 6 is an explanatory diagram showing boundary tracking processing; FIG. 7 is an explanatory diagram showing the tracking direction;
FIG. 9 is an explanatory diagram showing an approximate pattern of an unknown character obtained by the conventional method, FIG. 9 is an explanatory diagram showing a method of scanning an input image in the conventional method, and FIG. 10 is an explanatory diagram showing a tracking pattern.
FIG. 11 is an enlarged view showing the input image stored in the image memory, and FIG. 12 is an explanatory view showing the reference pattern and the corrected image pattern. 7... Image memory 1.2.3, 4.5
...Shift register 7.9...Address register
6.8...ROM10... image (Elephant Memory Patent Applicant: Tateishi Electric Co., Ltd.) (11) (9) (/1) ノも千／ノ り 9 Takashi 72 fig.

Claims (1)

[Claims] means for storing an input image; means for scanning the input image with a mask and sequentially extracting partial patterns within the mask; and a reference pattern having uneven portions is set, and a matching output is obtained by matching the partial pattern with the reference pattern. A pattern recognition device comprising means for transmitting, and means for discriminating peripheral pixels of the mask upon receiving a matching output, and transmitting a correction output for smoothing the input image according to the determination result.