JPH11238135A - Method and device for image recognition - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reduce a misdecision rate while making it possible to increase weighting where a specific image part generated on the basis of difference information between master images contributes to decision making by correcting similarity according to how much the specific image is included in respective difference images. SOLUTION: An image processor 13 cuts a sample image of a character out of video data inputted by using a CCD camera 12 and inputs it. Then it is compared with one master character to obtain their similarity. Then a difference image which is present in only one of similar characters, but not present in other master characters is generated. Further, a process for removing a part of a relatively high probability that an image to be actually recognized is the other master character is performed to allow a higher-precision decision making. Then the similarity is corrected according to how much the specific image generated on the basis of the difference images is included in respective difference images.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an image recognition method and an image recognition apparatus, and more particularly, to an image recognition method and an image recognition apparatus capable of performing highly accurate recognition.

[0002]

2. Description of the Related Art Conventionally, various methods have been proposed as methods for recognizing images of characters and symbols, such as the method disclosed in Japanese Patent Application Laid-Open No. 8-77293. Also,
Japanese Patent Application Laid-Open No. 3-125287 discloses a method of removing dust patterns from read image data.

[0003]

When the conventional character recognition technology as described above is applied to the recognition of characters such as product names, symbols, and numbers printed on the surface of a semiconductor mold package, for example, misreading occurs. If this occurs, the product manufactured based on the recognition information becomes defective and a great deal of labor is required for finding and repairing. Therefore, it is necessary to reduce the misreading rate as much as possible. However, there is a problem that if complicated processing such as extraction of characteristic information is performed, processing time is required and the speed of the production line is reduced. SUMMARY OF THE INVENTION It is an object of the present invention to provide an image recognition method and an image recognition apparatus that can solve the above-described problems of the related art and can perform recognition with a low misreading rate.

[0004]

SUMMARY OF THE INVENTION According to the present invention, when determining whether a master image such as a character having an image to be recognized is similar to a master image similar to the master image, the image to be recognized is determined. The similarity is obtained by the difference from the master image, and the similarity is corrected based on how much the specific image created based on the difference information between the respective master images is included in the respective difference images. There are features.

According to the present invention, the similarity is corrected based on the degree to which the specific image created based on the difference information between the respective master images is included in each of the differential images. Makes it possible to increase the weight that contributes to the determination, and acts to reduce the erroneous determination rate.

[0006]

Embodiments of the present invention will be described below in detail with reference to the drawings. FIG. 1 is a block diagram showing a configuration example of a recognition system to which the present invention is applied. IC as an inspection object set on the work positioning jig 11
The package 10 is illuminated by the illumination device 23 with a brightness at which the character pattern on the surface is easy to read. The magnification at which the character pattern fits into the field of view is set in the optical lens of the CCD camera 12, and the character pattern is imaged on the image sensor in the camera, and the video signal is output.

[0007] The image recognition device 13 comprises a normal computer system and an image recognition program. The CPU 14 is a processing device that executes a program in the ROM 15 or the RAM 16. The A / D converter 17 is a video capture circuit that converts a video signal input from the CCD camera 12 from analog to digital and captures it.

The hard disk device interface circuit HDDI 19 is, for example, a SCSI interface circuit for exchanging signals with the hard disk device 20. The CRT interface circuit 21 outputs the image signal of the screen generated by the CPU 14 to the CRT monitor 22 and inputs operation information from a keyboard (not shown).
The I / O interface circuit 18 outputs recognition result information to an external device, for example, an RS-232C interface circuit.

The character pattern captured from the CCD camera 12 is stored in the RAM 16 in the image processing device 16 at the same size. If the memory arrangement in the RAM 16 is defined as a predetermined array, for example, X512 × Y512, pixel information forming a character pattern is stored as brightness information corresponding to the address of the corresponding memory area. FIG.
2B shows an example in which a character pattern read from the camera 12 into the memory is read out and displayed on the CRT monitor 22.

[0010] In the image processing after the capture, if the processing is performed in pixel units, the amount of calculation increases.
As shown in FIG. 4B, the image data is divided into vertical and horizontal meshes, and the value (1) of the small section is determined based on whether the average value of the brightness of the included pixel data is equal to or more than a predetermined value (0) or not (1).
Bit) is stored in a two-dimensional array having the same configuration as the mesh, and recognition processing is performed based on the data. The width of the mesh is set as wide as possible within a range in which the recognition accuracy can be sufficiently ensured.

FIG. 2 is an explanatory diagram showing the contents of a difference pattern table between master characters generated before the recognition process. In this embodiment, an example in which the images to be recognized are 26 kinds of uppercase alphabets will be described. One of the squares in the “M” row in FIG. 2, for example, the square 31 is present only in the master character in the difference data between the master character “M” and the similar master character “N”. This represents difference image data that does not exist. That is, an image having a substantially “V” shape exists only in “M” and does not exist in “N”.

In addition, an image having a substantially “＼” shape which is present only in “N” and not present in “M” is registered in the cell 30 in the “N” row. Furthermore, the square 32 in the “Q” row has Q
The image of the beard portion is registered. Since there is no image that exists only in “O” and does not exist in “Q” in the square 33 in the “O” row, nothing is registered.

Next, a method of generating the difference pattern table between master characters shown in FIG. 2 will be described. First, the degree of similarity between pre-registered master characters is determined by any known method, or a similar character is manually input. Next, among similar characters, a difference image that exists only in one master character but does not exist in the other master character is generated (logical AND with the other inverted image).

The difference image data may be registered in the table shown in FIG. 2. Further, in this difference image, a comparison is made between the probability that the image to be recognized is the other master character and that the image exists. By performing the processing excluding the target part, the determination with higher accuracy becomes possible. This processing may be performed manually, or may be performed automatically, for example, to remove the vicinity of the contact point between the image common to the two master characters and the difference image.

Next, the recognition process will be described. FIG.
5 is a flowchart showing the contents of a character recognition process to which the present invention is applied. In S1, the image processing device is CC
The character sample images (FIGS. 4B and 4E) are cut out from the video data captured by using the D camera and captured. In S2, the degree of similarity is obtained by comparing with one master character. At this time, the size, inclination, position, and the like of the captured sample image are adjusted by a well-known method so as to maximize the similarity. As a comparison method, an exclusive OR (XOR) is obtained for each pixel (small block), and the similarity is determined such that the smaller the number of pixels remaining without matching, the higher the similarity.

In S3, it is determined whether or not the similarity is equal to or more than a predetermined value. If the similarity is equal to or more than the predetermined value, the process proceeds to S7. If the similarity is less than the predetermined value, S4 is performed. Move to In S4, it is determined whether the comparison with all the master characters has been completed. If not, the process proceeds to S5. If not, the process proceeds to S6. In S5, the master character is exchanged for the next one, and the process returns to S2 to repeat the comparison process. Further, in S6, it is determined that there is no master character whose similarity is equal to or more than the predetermined value, and the sample character is determined to be unrecognizable, and the process ends. In addition, compare with all master characters,
If the maximum similarity exceeds the predetermined value, S7
It may be made to shift to.

In S7, it is determined whether or not the current master character has a similar character (registered). If there is, the process proceeds to S9, but if not, the process proceeds to S8. In S8, it is determined that the sample image is the master character currently being compared, the determination result is output to the external device, and the process ends. In S9, similar to S2, the sample image is compared with the similar character to determine the degree of similarity.
In S10, a difference image pattern registered in any of the rows and columns corresponding to the current master character in the table of FIG. 2 is read.

In step S11, the logical product of each pixel is calculated, and the difference image pattern read out in step S10 is replaced with the image of the mismatched portion between the sample image obtained in steps S2 and S9 and the master character or similar character. The degree (number of pixels or area) is determined. If two difference image data are registered in the table of FIG. 2, four types of numerical values are obtained in S11. For example, when the master character is “M” and the similar character is “N”, the mismatch data between the sample images “X” and “M” is “MxorX”, the mismatch data between “N” is “NxorX”, The difference data existing only in “M” is named “M−N”, and the difference data existing only in “N” is named “NM” (hereinafter, a character string enclosed in brackets is a logical If not, the four types of numerical values are obtained as follows.

That is, the number K1 of the pixels of the difference data "MN" existing only in "M" in the mismatch data "MxorX" between "X" and "M" is calculated by
-N "and" MxorX ", and is obtained by counting the number of" 1 ". Similarly, the number K2 of the pixels of the difference data “M−N” existing only in “M” in the mismatch data “NxorX” between “X” and “N” is “M−N” for each pixel. And "NxorX", and is obtained by counting the number of "1". Further, the difference data “N−
The pixel “M” has mismatched data “Mxo” between “X” and “M”.
The number K3 existing in “rX” is obtained by taking the logical product of “N−M” and “MxorX” for each pixel and counting the number of “1”, and the difference existing only in “N”. The number K4 in which the pixel of the data “NM” exists in the non-coincidence data “NxorX” between “X” and “N” is obtained by taking the logical product of “NM” and “NxorX” for each pixel, "1"
Is determined by counting the number of

In S12, the correction amount of each character is obtained based on the maximum four kinds of numerical values K1 to K4 obtained in S11, and the similarity is corrected. For example, K1 is "M"
Since the number of pixels of the difference data “M−N” existing only in the mismatch data “MxorX” between “X” and “M” is the number of pixels, the larger the K1 is, the more “M” is assigned to “X”. There will be no feature. Therefore, K1 works to reduce the similarity of “X” to “M”. Similarly, K3
Works to increase the similarity of “X” to “M”.
K2 works to increase the similarity of "X" to "N", and K4 works to reduce the similarity of "X" to "N". In S13, the character corresponding to the character having the highest similarity after correction is selected, the determination result is output, and the process ends.

FIG. 4 is an explanatory diagram showing an example of image information in each step of the recognition processing. In FIG. 4,
(A) and (d) are master characters "O" and [Q], respectively, and (b) and (e) are sample images to be recognized. This image is, for example, a case in which an ink stain pattern is attached to “O”, and is a case that is often seen in the process of printing a mark. (G)
This is a difference image that exists only in the master character “Q” and is generated in advance prior to the recognition process. In this embodiment, specific small sections a to e are further manually extracted from (g), h) is registered in the cell 32 of FIG. Note that a difference image existing only in [O] is not registered.

In S2 of the flowchart of FIG. 3,
The exclusive OR (XOR) of the master character data of FIG. 4A and the sample image of FIG. 4B is calculated, and the mismatch data of FIG. 4C is obtained. Then, based on the mismatch data (c), the similarity with the master character is calculated. In S9 of the flowchart of FIG. 3, an exclusive OR (XOR) of the similar character data of FIG. 4D and the sample image of FIG. 4E is obtained, and non-coincidence data of FIG. 4F is obtained. Then, the degree of similarity with the similar character is calculated based on the non-coincidence data (f).

In S10 of the flowchart of FIG. 3, the difference pattern data of FIG.
In FIG. 11, data that does not match “O”
(I) The logical product (AND) of (same as (c)) and the difference pattern (h) is obtained, and the area K3 = 0 (5) of the result (k) is obtained.
Point does not exist) is required. Similarly, the logical product ((AND)) of FIG. 4 (j) (same as (f)), which is the mismatch data with “Q”, and the differential pattern (h) is obtained, and the area K4 of the result (m) is obtained. 5 is required.

The fact that K3 is 0 (none of the five points) indicates that the sample image does not include the characteristic point of "Q".
Is positively corrected by an amount corresponding to “5”. Further, the fact that K4 is 5 also indicates that the sample image does not include the feature point of “Q”. For example, the similarity of the sample image to “Q” is an amount corresponding to “5”. Correct only to minus.

As a result of such correction, even if it is determined in the first half of the process in FIG. 3 that the sample image is most similar to "Q", the similarity is corrected by the second half of the process and the sample is corrected. The result is that the image is closer to "O" (similarity to "O" is increased). When there are a plurality of similar characters, the similar characters are taken out one by one, compared and determined, and as a result of the correction, if the similarity of the similar character is larger, the similar character is replaced with the new master character. Again, similar characters of the new master character are taken out one by one and compared and determined.

Although the embodiments of the present invention have been disclosed above, the present invention may have the following modifications. In the embodiment, the example in which the difference pattern is generated by the difference with the similar character is disclosed.However, instead of the comparison with the similar character, the part that should not be present in the master character is input manually or the like. If there is a partial pattern, by correcting in the direction of decreasing the degree of similarity, it is possible to reliably detect a "must not" pattern. In the embodiment, the example in which the alphabet printed on the IC package is recognized is disclosed. However, the present invention is applicable to the recognition of any captured or input character, symbol, or graphic image.

[0027]

As described above, in the present invention,
Since the degree of similarity is corrected based on how much the specific image created based on the difference information between the master images is included in each difference image, the weight of the specific image portion contributing to the determination is increased. It is possible to reduce the erroneous determination rate. Further, the processing is relatively simple and the processing can be performed at high speed. Therefore, there is an effect that the yield is improved when used for product inspection or the like.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a configuration example of a recognition system to which the present invention is applied.

FIG. 2 is an explanatory diagram showing the contents of a difference pattern table between master characters.

FIG. 3 is a flowchart showing the contents of a character recognition process to which the present invention is applied.

FIG. 4 is an explanatory diagram showing an example of image information in each step of a recognition process.

[Explanation of symbols]

10: inspection object (IC package), 11: work positioning jig, 12: CCD camera, 13: image recognition device,
14 CPU, 15 ROM, 16 RAM, 17 A
/ D converter, 18 ... I / O interface circuit, 19
... HDD interface circuit, 20 ... Hard disk drive, 21 ... CRT interface circuit, 22 ... CR
T monitor device, 23 ... lighting device

Claims (3)

[Claims] When determining whether an image to be recognized is similar to a first master image or a second master image, the similarity is determined by a difference between the image to be recognized and each master image. The similarity is corrected based on the degree to which the specific image created based on the difference information between the master images is included in each of the difference images, and the determination is made based on the corrected similarity. An image recognition method characterized by performing. (1) First difference data, which is a difference between the image data to be recognized and the first master image data, is obtained, and the image data to be recognized and the first master data are determined from the first difference data. Calculating a first similarity with the image data; (2) calculating second data which is a difference between the image data to be recognized and the second master image data;
Calculating a second similarity between the image data to be recognized and the second master image data from the difference data. (3) First specific image data that is at least a part of image data that exists in the first master image data but does not exist in the second master image data exists in the first difference data. Calculating a first correction amount corresponding to the amount to be corrected. (4) Not present in the first master image data,
Obtaining a second correction amount corresponding to the amount of the second specific image data, which is at least a part of the image data existing in the master image data, existing in the first difference data. (5) First specific image data that is at least a part of image data that is present in the first master image data but is not present in the second master image data exists in the second difference data. Calculating a third correction amount corresponding to the amount to be corrected. (6) Not present in the first master image data,
Obtaining a fourth correction amount corresponding to the amount of the second specific image data, which is at least a part of the image data existing in the master image data, in the second difference data. (7) A step of correcting the first similarity in a direction in which the similarity increases based on the first correction amount, and correcting the first similarity in a direction in which the similarity decreases based on the second correction amount. . (8) A step of correcting the second similarity in a direction in which the similarity increases based on the third correction amount, and correcting the second similarity in a direction in which the similarity decreases based on the fourth correction amount. . (9) According to the magnitude of the corrected first and second similarities,
Determining whether the image data to be recognized is similar to the first master image data or the second master image data. An image recognition method comprising: 3. An image recognizing apparatus for determining which master image an image to be recognized is similar to, and a similarity calculating means for calculating respective similarities based on a difference between the image to be recognized and each master image. A similarity correction unit that corrects the respective similarities according to how many specific images created based on the difference information between the respective master images are included in the respective difference images; An image recognition device, comprising: a determination unit for performing a determination.