KR101782363B1 - Vision inspection method based on learning data - Google Patents

Abstract

A learning-based vision inspection method through data balancing of the present invention includes: a pre-learning step of collecting and formulating a video image from an intrinsic sample, which is a sample of an intrinsic defective product, and a false sample, which is a sample of a false defective product, displaying the video image on a feature space, and acquiring a classification criterion which is a boundary of an intrinsic area where the intrinsic data is positioned and a false area where the false data is positioned; a failure discrimination step of inserting test data extracted and formulated as the video image from a test object into a feature space and discriminating an area in which the test data is located on the feature space with the classification criterion as a boundary so as to discriminate the test object as an intrinsic defect or a false defect; and an additional learning step in which the test data is applied to the pre-learning step so as to correct the classification criterion in the feature space, wherein the test data is corrected through a data balancing step such that the intrinsic data and the false data compose a first data group with the same number. The failure discrimination step and the additional learning step are repeatedly performed.

Description

[0001] VISION INSPECTION METHOD BASED ON LEARNING DATA [0002]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a vision inspection method and a vision inspection apparatus for inspecting defects of a polarizing layer adhered to a display panel and more particularly, And more particularly, to a vision inspection method and a vision inspection apparatus which improve the accuracy of a defect determination.

In general, a display panel is formed by laminating a sheet or panel having characteristics of a reflective layer, a light guiding layer, a diffusion layer, a polarizing layer, or the like.

In the production process of such a display panel, a process of sorting good products and defective products is essentially performed by checking the adhesion state of a sheet or a panel.

As a method of classifying good products and defective products, it is common and highly accurate to compare visual information of good products and defective products which are observed through naked eyes by skilled workers.

However, relying on skilled personnel has a disadvantage in that the amount of display panel that can be checked per unit of time is limited with the occurrence of high cost, and the reliability of the inspection may vary depending on the skill level of inspection personnel.

As a method for detecting defects, a vision inspection apparatus for photographing and inspecting a plane to be inspected of a display panel has been introduced and performed. However, in the process of inspecting whether or not a polarizing layer adhered to a display panel is defective , The distinction between the intrinsic defect and the false defect is not clear, so that the false defect is frequently detected, and thus, the accuracy of detection of the defect is low.

This is because, in judging whether or not the polarizing layer adhered to the display panel is bad, it is possible to distinguish the false defects that can be classified as defective due to intrinsic defects such as lifting or dropping, simple stains, This is because it is difficult to carry out through.

Although it was possible to increase the fault detection speed and reduce the labor cost by automating the process of detecting the defective display panel, it was found that due to the technical limit, the false defect that can be classified as good product was detected, Furthermore, a number of processes were required.

When a sample of accumulated data is biased to a class in a process of deriving a more precise criterion by utilizing data of pseudo-impairment and imperfection accumulated by carrying out the inspection process and reapplying it to the inspection system In addition, in order to reduce the error and to derive more precise criteria in applying the collected data to the learning, there is a need to solve the data imbalance among the classes by processing the collected data.

If the number of data belonging to one class is more or less than the number of data belonging to another class, it can be seen that the data is unbalanced.

However, learning algorithms such as SVM (support vector machine) are based on the assumption that the data ratios between classes are almost similar, so that if the data imbalance occurs, the accuracy of the boundary determining the pseudorandom and intrinsic error decreases Could be caused.

Therefore, a method for solving such problems is required.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned problems of the prior art, and it is an object of the present invention to provide a vision inspection apparatus which can obtain reference data capable of enhancing discrimination between true defectiveness and false defectiveness, And to provide a vision inspection method and a vision inspection apparatus in which the detection is reduced and the detection accuracy of the true defect is improved.

It is another object of the present invention to provide a vision inspection method and a vision inspection apparatus having an algorithm that corrects data obtained through inspection to be used for learning so that accuracy of reference data for identifying defects increases as the process is repeated.

The present invention is to provide a vision inspection method and a vision inspection apparatus which improve the reliability of application of a learning algorithm by complementing the imbalance between classes generated in data collected for learning through a random sampling method.

The problems of the present invention are not limited to the above-mentioned problems, and other problems not mentioned can be clearly understood by those skilled in the art from the following description.

In order to accomplish the above object, a learning-based vision inspection method using data balancing according to the present invention is characterized in that a first learning group is formed by including the intrinsic specimens, which are samples of genuine defective products, and the false specimens, Extracting a first data group having the same number of intrinsic data and pseudo-data as the video image collected from the first learning group and extracting the first data group on the feature space and extracting the intrinsic region and the false data A learning step of deriving a classification criterion that is a boundary of a pseudo region in which a polarizing layer is attached, and a step of assigning, on a feature space, The area where the test data is located in the space is judged, and the test object is judged to be an intrinsic defect or a false defect. The test data is applied to the defect discrimination step and the pre-learning step to correct the classification criterion in the feature space, and the test data is further learned through the data balancing step so that the intrinsic data and the pseudo- And a failure determination step and a further learning step are repeatedly performed.

The pre-learning step includes an image collecting step of collecting a first learning image group composed of a genuine image and a false image extracted from the respective intrinsic specimens included in the first learning group and the image information from the specimen, The data is derived from n respective feature points from each of the intrinsic images and the false images included in the first learning image group to form a first data group, a first data group is schematized to group adjacent feature points into k groups, A dictionary creation step of obtaining a feature point average for each group to create a dictionary composed of k average points and a step of creating a dictionary on the feature space by substituting the intrinsic data and the pseudo data in advance and extracting the intrinsic region and the pseudo- And a classifying step of determining a classifying criterion which is a boundary of the located caustic area.

Alternatively, the dictionary creation step performs K average clustering to create the dictionary with k average points as reference words.

In the defect determination step, the test data, which is n feature points obtained from the image of the subject extracted from the image information, is compared with the dictionary, and the test data is classified based on k average points and displayed on the feature space.

Alternatively, the additional learning step corrects the classification criterion by adding the test data schematized on the characteristic space through the defect discrimination step to the intrinsic data or the caustic data of the class classification step.

In the data balancing step, if the test object is judged to be an intrinsic defect through the defect determination step, the test data and the one of the pseudo data previously selected through the pre-learning step are applied to the class classification step, And if the inspected object is judged to be a false defect through the defective discrimination step, one of the genuine data randomly selected from the test data and the intrinsic data derived through the pre-learning step is applied to the class classification step, .

Alternatively, in the data balancing step, the average value of the test data and the pseudo data derived through the pre-learning step when the object to be inspected is determined to be defective through the defect determination step is applied to the class classification step to correct the classification criterion, The average value of the test data and the intrinsic data derived through the pre-learning step is applied to the class classification step to correct the classification criterion.

The intrinsic image, the false image, and the image to be inspected are image information of a point where the inspected surface of the inspected object is moved according to a certain pattern and the modulated signal of the photographed image is abruptly changed.

In order to accomplish the above object, the present invention provides a vision inspection apparatus through data balancing, comprising: a transfer unit to which an object to be inspected is transferred; a photographing unit to photograph an inspection surface of the object to be inspected; And a storage unit for storing the test data obtained through the operation unit. The operation unit may include a storage unit for storing the test data obtained through the operation unit, The stored data is compared with the test data to determine whether the test object is defective. The storage unit stores the test data and the virtual data derived from the test data together.

In order to solve the above problems, a learning-based vision inspection method using data balancing and a learning-based vision inspection apparatus using data balancing according to the present invention acquires data on intrinsic defectiveness and false defectiveness of a display panel as image information, Based on a learning-based algorithm that identifies intrinsic and intrinsic defects by deriving the boundaries by plotting them on the feature space, it improves the accuracy of the fault discrimination based on the characteristics of intrinsic and intrinsic defects, Can be obtained.

Further, by accumulating the data of the intrinsic defect and the false defect obtained through the repeated defect discrimination step and further learning, the boundaries of the intrinsic defect and the false defect are continuously corrected, and the data to be further learned is not biased into any one area So that it is possible to prevent the increase of the error or the performance degradation of the learning algorithm caused by the unbalance of the data.

Thus, learning algorithms can be used to enhance the accuracy of vision boundaries.

The effects of the present invention are not limited to the effects mentioned above, and other effects not mentioned can be clearly understood by those skilled in the art from the description of the claims.

FIG. 1 is a flowchart illustrating a learning-based vision inspection method using data balancing according to an exemplary embodiment of the present invention.
FIG. 2 is a flowchart illustrating data defining a classification criterion in a learning-based vision inspection method using data balancing according to an exemplary embodiment of the present invention.
FIG. 3 is an image showing an intrinsic image of an intrinsic sample in a learning-based vision inspection method using data balancing according to an embodiment of the present invention.
4 is an image showing a false image of a false sample in a learning-based vision inspection method using data balancing according to an embodiment of the present invention.
FIG. 5 is a schematic diagram illustrating a process of deriving a dictionary in a learning-based vision inspection method using data balancing according to an embodiment of the present invention.
FIG. 6 is a schematic view showing a process of assigning a first image group in advance in a learning-based vision inspection method using data balancing according to an embodiment of the present invention.
FIG. 7 is a schematic view showing a learning step and a failure determination step in a learning-based vision inspection method using data balancing according to an embodiment of the present invention.
FIG. 8 is a graph illustrating a three-dimensional representation of a state in which an intrinsic region and a false region are derived on a feature space in a learning-based vision inspection method using data balancing according to an exemplary embodiment of the present invention.
FIG. 9 is a schematic diagram illustrating a process of performing data balancing in a learning step of a learning-based vision inspection method using data balancing according to an exemplary embodiment of the present invention.
FIG. 10 is a state diagram illustrating a learning-based vision inspection apparatus through data balancing according to an embodiment of the present invention.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, preferred embodiments of the present invention will be described with reference to the accompanying drawings. In describing the present embodiment, the same designations and the same reference numerals are used for the same components, and further description thereof will be omitted.

The learning-based vision inspection method through data balancing according to the present invention can be implemented as follows.

FIG. 1 is a flowchart illustrating a method of inspecting a vision-based learning through data balancing according to an embodiment of the present invention. FIG. 2 is a flowchart illustrating a method of inspecting a vision-based vision by data balancing according to an exemplary embodiment of the present invention. As shown in FIG.

The learning-based vision inspection method using data balancing according to an embodiment of the present invention is a learning-based vision inspection method for inspecting defects of the polarizing layer 304 adhered to the display panel 302.

Referring to Figs. 1 and 2, a first learning group formed by including the intrinsic specimen 100, which is a sample of intrinsic defects, and the tentative specimen 200, which is a sample of a false defects, are provided in the same number.

The first data group including the same number of the intrinsic data 120 and the false data 220 extracted from the first learning group is extracted.

The first data group is displayed on the feature space 400 to derive the classification reference 410 that is the boundary between the intrinsic region 130 where the intrinsic data 120 is located and the false region 230 where the false data 220 is located Learning step S100 is performed.

The inspected data 320 extracted from the display panel 302 to which the polarizing layer 304 as the inspection object 300 is adhered and formulated is substituted into the feature space 400.

The defect determination step S200 of determining the region where the test data 320 is located on the feature space 400 with the classification reference 410 as a boundary is performed to determine the test object 300 as an intrinsic defect or a false defect .

The test data 320 is applied to the pretest learning step S100 to correct the classification criterion 410 on the feature space 400 so that the test data 320 can be obtained by using the same number of the intrinsic data 120 and the false data 220 (S300) that is corrected through a data balancing step (S310) so as to form a first data group, and the failure determination step (S200) and the additional learning step (S300) are repeatedly performed.

Each of the above steps will be described in detail below.

The first learning group includes the same number as the intrinsic specimen 100, which is a sample of the intrinsic defective product, and the simulated specimen 200, which is a sample of the false defective product.

The first learning image group includes the same number of the intrinsic image 110 derived from the intrinsic specimen 100 and the intrinsic image 210 derived from the pseudo-specimen 200. [

The first data group includes the same number of the intrinsic data 120 in which the intrinsic image 110 is formulated and the false data 220 in which the virgin image 210 is formulated.

Therefore, it is assumed in the embodiment of the present invention that the intrinsic specimen 100 and the intrinsic image 110 are the same A, and the false specimen 200 and the false image 210 are also the same A.

The feature space 400 is a multi-dimensional space having a plurality of variables, and refers to a space in which specific data according to an embodiment of the present invention is depicted.

The pre-learning step S100 includes an image collecting step S110 for obtaining an intrinsic image 110 and a false image 210, which are image images of suspected bad points from the intrinsic specimen 100 and the false specimen 200, Creating step S120 and class classification step S130.

FIG. 3 is an image showing an intrinsic image of a genuine sample in a learning-based vision inspection method using data balancing according to an exemplary embodiment of the present invention. FIG. This is an image that shows the false image of the pseudo-sample in the vision inspection method.

3 to 4, in the image collection step S110, an image is collected through the photographing unit 520 that photographs the moving target planes of the intrinsic specimen 100 and the tentative specimen 200 in a predetermined pattern And acquiring the intrinsic image 110 or the false image 210 by imaging the image of the specific position at the time of deriving a result satisfying predetermined conditions.

In an embodiment of the present invention, when a change in brightness of an image picked up through the photographing unit 520 rapidly increases or decreases, the point may be photographed as shown in FIGS. have.

The preprocessing step S120 derives the intrinsic data 120 and the pseudo-data 220 derived from the intrinsic image 110 and the voiced image 210, respectively, into n feature points.

The intrinsic data 120 is obtained A x n, and the mantissa data 220 is also obtained A x n.

FIG. 5 is a schematic diagram illustrating a process of deriving a dictionary in a learning-based vision inspection method using data balancing according to an exemplary embodiment of the present invention. FIG. FIG. 3 is a schematic diagram showing a process of assigning a first learning image group in advance in the vision inspection method. FIG.

5 through 6, the intrinsic data 120 and the tentative data 220 derived from the intrinsic image 110 and the tentative image 210 are stored in the intrinsic image 110 and the tentative image 210, Such as brightness, saturation, shape, curvature, etc., can be extracted from n points.

The intrinsic data 120 and the tentative data 220 are converted into a vector quantity having characteristics such as color, lightness, saturation, shape, curvature, and the like as variables, and a multidimensional space Lt; / RTI >

FIG. 5 schematically shows a process in which n feature points are derived as vector quantities from the intrinsic image 110 and the tentative image 210 and are expressed in a multidimensional space.

This is an illustrative example, and the intrinsic data 120 and the false data 220 can be extracted based on various factors according to an embodiment to which the present invention is applied.

As described above, A x n number of false data 220 are collected from A number of intrinsic images 110, A x n number of intrinsic data 120, A number of false images 210, and the intrinsic data 120 ) And the false data 220 are schematized on the multidimensional space as described above, and are grouped into k groups by grouping the adjacent intrinsic data 120 and the false data 220, and the intrinsic data 120 belonging to each group 120 and the false data 220 are calculated to calculate k average points.

At this time, each group is defined as a word, and a k-word dictionary is created by defining k average points as a dictionary.

The preprocessing step (S120) may be performed using K average clustering, which is an algorithm for grouping the given data into k clusters.

The class classification step S130 substitutes the intrinsic data 120 into the dictionary created through the pre-creation step S120 to histogram the intrinsic data 120 belonging to each of the k words as shown in Fig. 6 The histogramming step is performed first.

FIG. 7 is a schematic diagram illustrating a learning step and a failure determination step in a learning-based vision inspection method using data balancing according to an embodiment of the present invention. FIG. 8 is a flowchart illustrating a method of learning through data balancing according to an exemplary embodiment of the present invention. 3 is a graph illustrating a state in which an intrinsic region and a pseudo-region are derived on a feature space in a vision-based vision inspection method.

8, the intrinsic data 120 belonging to a word and each word are displayed on the feature space 400 as a variable on the basis of the histogram obtained through the histogramming step described above, The intrinsic region 130 is derived.

The histogramming step of substituting the pseudo-data 220 into the dictionary created through the preprocessing step S120 to histogramize the pseudo-data 220 belonging to each of the k words as shown in Fig. 6 is performed do.

The words and the pseudo data 130 belonging to each word are shown on the feature space 400 as variables and the feature space 400 as shown in Fig. 8, based on the histogram obtained through the pseudo- 400). ≪ / RTI >

The classification reference 410 may be expressed as a line, a plane, or a hyperplane which is a boundary between the intrinsic region 130 and the pseudo region 230 on the feature space 400, and the intrinsic region 130 and the pseudo- A support vector machine (SVM) technique may be used to determine the boundary between the feature space 400 and the feature space 400 to maximize the space between them.

In the defect determination step S200, the surface on which the polarizing layer 304 is adhered is continuously photographed on the display panel 302, which is the inspection object 300, along a predetermined path through the photographing unit 520, (S210) of acquiring an image 310 of an image of a position indicating the position of the subject image 310 and the intrinsic image 110 and the false image 210 of the image 310, And the visible image 310 are previously substituted and displayed on the feature space 400 using the same method as the method of transforming the image into the visible image 310 and the false data 220, Judging whether the genuine region 130 or the false region 230 belongs to the classification reference 410 obtained through the classification step 410 and determining whether the genuine region 130 or the false region 230 is genuine or false S220).

In the additional learning step S300, a step of correcting the dictionary through the process of adding the image 310 to be inspected obtained in the failure determination step S200 in advance and re-reproducing k average points may be added.

Alternatively, a classification criterion correction step S320 for adding the test data 320 obtained through the failure judgment step S200 on the feature space and correcting the classification criterion 410 by reflecting the added test data 320 is performed .

In the step of correcting the dictionary, the classification criterion correction step (S320) may be performed by performing defect determination on the learned defect criterion learned through the pre-learning step (SlOO), adding new information to be acquired, Can be obtained.

It is possible to increase the accuracy of performing the failure discrimination by increasing the number of learning samples and to reduce the phenomenon in which the false defect is detected by making the boundary between the true defect and the false defect more apparent in the feature space 400. [

The additional learning step S300 may include a data balancing step S310.

FIG. 9 is a schematic view illustrating a process of performing data balancing in an additional learning step using test data in a learning-based vision inspection method using data balancing according to an exemplary embodiment of the present invention.

Referring to FIG. 9, the data balancing step S310 is a step of balancing the data of the characteristic space (step S320), which has been learned in the classification criterion correction step S320 when the inspection object 300 acquired through the defect determination step S200 is an intrinsic defect The data added to the characteristic space 400 learned in the classification criterion correction step S320 when the object 300 to be inspected is a false defect, And serves as data 220.

와 같은 수식으로 표현될 수 있으며, 입력벡터 x에 대해 입력공간(input space)이 되는 특징공간(400)상의 초평면(hyperplane)을 찾는 이진 분류 문제로 표현될 수 있다.The problem of data asymmetry that can occur between classes that are classified as intrinsic data 120 or fuzzy data 220

And can be expressed as a binary classification problem for finding a hyperplane on the feature space 400 as an input space for the input vector x.

The intrinsic data 120 or the pseudo data 220 are classified into two classes of '+1' and '-1', and the class classification problem can be regarded as searching for a w that provides excellent generalization performance.

At this time, the case where the ratio of the number of data belonging to the intrinsic data 120 and the number of data belonging to the false data 220 is significantly different can be regarded as an imbalanced data problem.

In the case of such an unbalance data problem, data is concentrated in one of the intrinsic data 120 or the false data 220, which is a classification reference correction step (step S100) for correcting the learned failure determination reference S320) may result in lowering the reliability of the failure detection standard.

Therefore, when the test data 320 obtained in either case of the intrinsic data 120 or the false data 220 is biased, as shown in FIG. 9A, 130 or the pseudo region 230 may be shifted to either side.

Particularly, when the defect inspection of the polarizing layer 304 of the display panel 302 is performed, the detection frequency of the intrinsic defect is remarkably higher than the false defect, and the bias of such data acts as an impediment to improve the accuracy of the defect discrimination .

Therefore, when the test data 320 obtained in the defect determination step S200 belongs to the intrinsic region 130 in the feature space 400, the pseudo data 220 included in the pseudo region 230 is randomly selected A single verifiable data 220 is added on the feature space 400 together with the obtained test data 320 to correct the classification criterion 410. [

More specifically, if a total of four true defect judgments are obtained by performing inspection of the object 300 through the defect discrimination step S200, the four false data 320 together with the obtained false data 220, Is added to the feature space 400 to reset the classification reference 410.

When the false determination is obtained by performing verification of the inspection object 300 through the defect determination step S200, the extracted derived intrinsic data 120 is added to the feature space 400 together with the test data 320 And the classification criteria 410 are reset.

As another embodiment of the present invention, if the intrinsic defect determination is derived by performing the inspection of the object 300 through the defect determination step S200, the average value of the previously obtained pseudo data 220 is used as the feature data 320 together with the test data 320 And adds the average value of the acquired intrinsic data 120 to the classification space 410 together with the test data 320 on the feature space when the false determination is derived. Can be reset.

In this embodiment, since data balancing is performed by obtaining an average value of data obtained by simple operations only without randomly selecting data, an operation process can be simplified and the processing speed can be improved.

The learning-based vision inspection apparatus through data balancing according to the present invention can be implemented as follows.

FIG. 10 is a state diagram illustrating a learning-based vision inspection apparatus through data balancing according to an embodiment of the present invention.

10, the learning-based vision inspection apparatus using data balancing according to an exemplary embodiment of the present invention includes a conveying unit 510 for conveying an object to be inspected 300, an inspection surface of the object to be inspected 300 A reading unit 530 and a reading unit 530 for reading the image image obtained through the photographing unit 520 and selecting the image 310 to be inspected, An operation unit 540 for visualizing the test image 320 obtained by modifying the image 310 obtained through the operation unit 540 on the feature space 400 and a storage unit 540 for storing the test data 320 obtained through the operation unit 540 The arithmetic unit 540 determines whether the inspection object 300 is defective by comparing the data stored in the storage unit 550 with the inspection data 310. The storage unit 550 stores the inspection data 320 and the test data 320 may be stored together.

More specifically, the reading unit 530 performs the image capturing step S110 and the image capturing step S210 described above, and sequentially captures the image obtained by photographing the object 300 through the image capturing unit 520 And acquires a video image at a position that meets predetermined conditions.

The operation unit 540 performs pre-creation step S120, class classification step S130, and defect classification step S220 according to the classification criteria.

The storage unit 550 performs an additional learning step S300 including a data balancing step S310 and a classification criterion correction step S320.

It will be apparent to those skilled in the art that the present invention can be embodied in other specific forms without departing from the spirit or scope of the invention as defined in the appended claims. It is obvious to them. Therefore, the above-described embodiments are to be considered as illustrative rather than restrictive, and the present invention is not limited to the above description, but may be modified within the scope of the appended claims and equivalents thereof.

100: intrinsic specimen 110: intrinsic image
120: intrinsic data 130: intrinsic region
200: False sample 210: False image
220: falsifying data 230:
300: object to be inspected 302: display panel
304: polarization layer 310: image to be imaged
320: Test data
400: Feature Space 410: Classification Criteria
510: transfer means 520: photographing unit
530: Reading section 540:
550:

Claims (9)

A learning-based vision inspection method for inspecting defects of a polarizing layer adhered to a display panel,
A first learning group is formed by including the same number of false samples that are samples of intrinsic defective products and a false sample that is a sample of false defects and the same number of the intrinsic data and the false data in which the video image collected from the first learning group is formulated Learning step of extracting a first data group that is a boundary between the first data group and the first data group and deriving a classification criterion that is a boundary between an intrinsic area where the intrinsic data is located and a pseudo area where the pseudo data is located;
The method comprising the steps of: inputting test data obtained by extracting and formulating a video image from a display panel to which a polarizing layer as an object to be inspected is adhered on the feature space, determining an area where the test data is located on the feature space, A defect judgment step of judging that the inspection object is an intrinsic defect or a false defect; And
Wherein the test data is applied to the pre-learning step to correct a classification criterion in the feature space, wherein the test data is corrected through a data balancing step such that the intrinsic data and the pseudo data constitute the first data group An additional learning step,
Wherein the pre-
An image collecting step of collecting a first learning image group composed of a genuine image and a false image extracted from each of the intrinsic specimens included in the first learning group and image information from the pseudo specimen;
Wherein the intrinsic data and the pseudo data are derived from n feature points respectively from the respective intrinsic images and the pseudo images included in the first learning image group to form the first data group and the first data group A dictionary creation step of grouping adjacent feature points into k groups and obtaining a feature point average for each group to create a dictionary composed of k average points; And
A classifying unit for classifying the intrinsic data and the pseudo data into the dictionary and drawing it on the feature space and determining a classification reference that is a boundary between the intrinsic region where the intrinsic data is located and the pseudo- Comprising:
Wherein the defective discrimination step and the additional learning step are repeatedly performed.
delete The method according to claim 1,
The pre-
K-means clustering to create the dictionary with k average points as a reference word.
The method according to claim 1,
In the failure determination step,
Learning by data balancing in which the test data, which is n feature points obtained from the image to be inspected extracted from the image to be inspected, is compared with the dictionary and the test data is classified based on the k average points and the test data is displayed on the feature space Based vision inspection method.
5. The method of claim 4,
Wherein the additional learning step comprises:
And adding the test data schematized on the feature space to the intrinsic data or the false data of the class classification step through the defect determination step to correct the classification criterion.
6. The method of claim 5,
The data balancing step includes:
If the inspected object is judged to be an intrinsic defect through the fault discrimination step, the test data and the one of the pseudo data derived in advance through the pre-learning step are randomly selected to be applied to the classification step, And a step of classifying at least one of the genuine data randomly selected from the test data and the intrinsic data pre-learned through the pre-learning step, when the inspected object is judged as a false defect through the defect determination step, Based vision inspection method using data balancing that corrects the classification standard by applying the method of claim 1. [
6. The method of claim 5,
The data balancing step includes:
Correcting the classification criterion by applying an average value of the test data and the pseudo data derived through the pre-learning step to the class classification step when the inspected object is determined to be a genuine defect through the defect determination step, Data balancing for correcting the classification criterion by applying an average value of the test data and the intrinsic data previously extracted through the pre-learning step to the class classification step when the inspection object is judged to be a false defect through the defect determination step Learning - based vision inspection method.
5. The method of claim 4,
Wherein the intrinsic image, the false image,
Based vision inspection through data balancing, which is image information of a point where a surface of an object to be inspected of the object to be inspected moves according to a certain pattern and a modulated signal of the imaged image is abruptly changed.
delete

Images