WO2017204519A2

WO2017204519A2 - Vision inspection method using data balancing-based learning, and vision inspection apparatus using data balancing-based learning utilizing vision inspection method

Info

Publication number: WO2017204519A2
Application number: PCT/KR2017/005326
Authority: WO
Inventors: 오병준; 전동철; 신원종
Original assignee: (주)에이앤아이
Priority date: 2016-05-23
Filing date: 2017-05-23
Publication date: 2017-11-30
Also published as: KR101782363B1; CN109462999A; WO2017204519A3; CN109462999B

Abstract

A vision inspection method using data balancing-based learning according to the present invention comprises: a prior learning step for establishing a classification criterion, which is the boundary for locations of true and false data, by collecting, mathematizing, and schematizing, in a particular space, images of true samples that are samples of genuinely defective products and false samples that are samples of falsely defective products; a defect determination step for determining the object under inspection as genuinely defective or falsely defective by introducing, into the particular space, data to be inspected, which has been extracted as an image from an object under inspection and mathematized, and determining the area in which the data to be inspected is located in the particular space with the classification criterion as the boundary; and an additional learning step for modifying the classification criterion of the particular space by applying the data to be inspected to the prior learning step, wherein the data to be inspected is modified by means of a data balancing step so that the same number of true data and false data are included, and the defect determination step and additional learning step are repeatedly carried out.

Description

Learning-based vision inspection method through data balancing and learning-based vision inspection apparatus using data balancing

The present invention relates to a vision inspection method and a vision inspection apparatus for inspecting a defect of a polarizing layer adhered to a display panel, and more particularly, a data imbalance phenomenon in which oversampled data of a specific class is biased through balancing of learned data. The present invention relates to a vision inspection method and a vision inspection apparatus that have improved the accuracy of failure discrimination by overcoming these problems.

In general, a display panel is formed by stacking sheets or panels having respective characteristics such as a reflective layer, a light guide layer, a diffusion layer, and a polarizing layer.

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

As a method of classifying good and bad quality, it is common and highly accurate to use visual information on good and bad quality observed by naked workers.

However, relying on skilled personnel has a limitation in the amount of display panels that can be inspected per unit time with the occurrence of high costs, and the reliability of the inspection may vary depending on the skill of inspectors.

Accordingly, as a method for detecting defects, a vision inspection apparatus for photographing and inspecting a plane that is an inspection target of a display panel has been introduced and implemented. However, in the process of inspecting whether a polarization layer adhered to a display panel is defective or not, However, since the intrinsic and false causality is not clearly distinguished, the causal defect is frequently overdetected, and thus, the accuracy of detecting the intrinsic defect is poor.

In determining whether or not the polarizing layer adhered to the display panel is defective, it is possible to distinguish the caustic defects that can be classified as defective due to intrinsic defects such as lifting or stamping and damage of simple stain or protective film. Because it is difficult to carry out through.

By automating the process of detecting defects in display panels, it was possible to increase the speed of defect identification and reduce labor costs.However, due to technical limitations, the false positives that can be classified as good products are overdetected, resulting in low reliability and repeated re-examination. More steps were needed.

In addition, when a sample of accumulated data is concentrated in one class by deriving a more accurate criterion by using the data of false and intrinsic defects accumulated during the inspection process and reapplying it to the inspection system. In general, in order to reduce errors and derive more accurate criteria in applying additionally collected data to learning, it was necessary to process the collected data to solve the data imbalance between classes.

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

However, learning algorithms such as support vector machines (SVMs) assume that data rates between classes are about the same, so when data imbalances occur, the accuracy of the boundary that determines false and true defects becomes low. Could be effected.

Therefore, there is a need for a method for solving such problems.

The present invention has been made in order to solve the problems of the prior art described above by combining the discrimination technology by the recognition and learning in the vision inspection apparatus to calculate the reference data that can increase the discrimination ability of false and false false by calculating the false false The purpose of the present invention is to provide a vision inspection method and a vision inspection apparatus that reduce overdetection and improve detection accuracy of intrinsic defects.

In addition, it is to provide a vision inspection method and a vision inspection apparatus having an algorithm that is corrected so that the accuracy of the reference data for determining defects as the process is repeated by processing the data obtained through the inspection and use in learning.

In addition, it is to provide a vision inspection method and a vision inspection apparatus which improve the reliability of application of a learning algorithm by supplementing the class imbalance generated in the data collected for learning through a random sampling method.

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

In the learning-based vision inspection method through the data balancing of the present invention for achieving the above object, the first learning group is formed by including the same number of authentic samples as samples of inferior goods and false samples as samples of inferior goods in the same number. Intrinsic region and false data where true data is located by extracting a first data group having the same number of intrinsic data and false data in which the image image collected from the first learning group is formulated, and plotting the first data group in a feature space The pre-learning step of deriving the classification criteria, the boundary of the caustic region where is located, and extracting the image data from the display panel to which the polarizing layer, which is the test object, is attached, is inserted into the feature space into the feature space and characterized by Judging the area where the test data is located in the space and selling the test object as true or false The classification criteria in the feature space are corrected by applying the test data to the defective discrimination step and the pre-learning step, wherein the test data is corrected through the data balancing step so that the true data and the false data constitute the first data group with the same number. Including the step, the failure discrimination step and the further learning step are repeatedly performed.

The pre-learning step may include an image collection step of collecting a first learning image group including a true image and a pseudo image extracted from the true samples and the false samples included in the first learning group. Data is derived from each of the intrinsic and pseudo images included in the first learning image group, respectively, to form n first data points to form a first data group, and the first data group is plotted to group adjacent feature points into k groups. The pre-writing step to obtain a dictionary consisting of k average points by calculating the average of the feature points for each group and assigning the true and false data to the dictionary and plotting them in the feature space, and the true zone and the false data where the true data is located in the feature space And class classification step for determining the classification criteria which is the boundary of the caustic zone located.

Alternatively, in the dictionary creation step, the k mean clustering is performed to generate the dictionary using k mean points as the reference word.

In addition, the defect discriminating step compares the test data, which is n feature points, obtained from the test image extracted from the test object with the image information, and classifies the test data based on k average points and plots them on the feature space.

Alternatively, in the additional learning step, the classification criteria are corrected by adding the test data plotted on the feature space to the true data or the false data of the class classification step through the failure discrimination step.

In the data balancing step, when a test object is judged to be a genuine defect through a bad discrimination step, a classification standard is applied by applying one randomly selected caustic data among the causal data derived through the test data and the pre-learning step to a class classification step. If the test object is judged to be false through the bad discrimination step, the classification criteria are applied by applying one of the genuine data randomly selected from the test data and the intrinsic data derived through the pre-learning step to the class classification step. Correct it.

Alternatively, in the data balancing step, if the inspection object is judged to be incomplete through the bad discrimination step, the classification criteria are corrected by applying the average value of the caustic data derived through the test data and the pre-learning step to the class classification step. If the test object is judged to be false in the step, the classification criteria are corrected by applying the average value of the intrinsic data derived through the test data and the pre-learning step to the class classification step.

In addition, the true image, the pseudo image, and the test image are image information of a point where the modified signal of the captured image is rapidly changed while moving the inspection target surface of the inspection object according to a predetermined pattern.

In addition, the vision inspection apparatus through the data balancing of the present invention for achieving the above object is to read the transfer means for transporting the inspection object, a photographing unit for photographing the inspection surface of the inspection object, the image image obtained through the photographing unit A reading unit which selects a test image, an operation unit which maps the test data obtained through the reading unit to the feature space, and a storage unit which stores the test data obtained through the operation unit, and which stores the test data. It is determined whether the inspection object is defective by comparing the stored data and the test data, and the storage unit stores the test data and the virtual data derived through the test data.

Learning-based vision inspection method through data balancing and learning-based vision inspection apparatus through data balancing of the present invention for solving the above problems by acquiring and formulating the data of the intrinsic and false false of the display panel as image information Based on the learning-based algorithm that derives the boundary by drawing the boundary in the feature space and discriminates true and false defects, it improves the accuracy of defect discrimination based on the characteristics of true and false defects and overdetects false negatives. A reducing effect can be obtained.

In addition, by accumulating the data of intrinsic and false inferiority acquired through repeated failure determination step, the boundary between intrinsic and false inferiority is continuously corrected, and the data to be further trained are not concentrated in any one area. By maintaining the balance of the data, it is possible to prevent the increase of the error that may be caused by the data imbalance or the performance degradation of the learning algorithm.

Therefore, it is possible to utilize the boundary with improved accuracy through the learning algorithm for vision inspection.

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

1 is a flowchart illustrating a learning-based vision inspection method through data balancing according to an embodiment of the present invention.

2 is a flowchart illustrating data for determining classification criteria in a learning-based vision inspection method through data balancing according to an embodiment of the present invention.

3 is an image showing a true image of a true sample in a learning-based vision inspection method through data balancing according to an embodiment of the present invention.

4 is an image showing a pseudo image of a pseudo sample in a learning-based vision inspection method through data balancing according to an embodiment of the present invention.

5 is a schematic diagram illustrating a process of deriving a dictionary from a learning-based vision inspection method through data balancing according to an embodiment of the present invention.

6 is a schematic diagram illustrating a process of assigning a first image group to a dictionary in a learning-based vision inspection method through data balancing according to an embodiment of the present invention.

7 is a schematic diagram showing a pre-learning step and a bad discrimination step in a learning-based vision inspection method through data balancing according to an embodiment of the present invention.

8 is a graph illustrating a three-dimensional example of a state in which an intrinsic region and a pseudo region are derived in a feature space in a learning-based vision inspection method through data balancing according to an embodiment of the present invention.

9 is a schematic diagram illustrating a process of performing data balancing in a two-dimensional graph in the further learning step of the learning-based vision inspection method through data balancing according to an embodiment of the present invention.

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 Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. In the description of this embodiment, the same name and the same reference numerals are used for the same configuration and additional description thereof will be omitted.

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

1 is a flowchart illustrating a learning-based vision inspection method through data balancing according to an embodiment of the present invention, and FIG. 2 is a classification criterion in the learning-based vision inspection method through data balancing according to an embodiment of the present invention. A flow chart showing the data used to determine

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

1 to 2, a first learning group including an intrinsic sample 100 that is a sample of an inferior goods and a pseudosample 200 that is a sample of an inferior goods in the same number is provided.

The first data group having the same number of the intrinsic data 120 and the false data 220 in which the image image collected from the first learning group is modified is extracted.

The first data group is plotted on the feature space 400 to derive the classification criteria 410 which is a boundary between the intrinsic region 130 where the intrinsic data 120 is located and the caustic region 230 where the caustic data 220 is located. The pre-learning step (S100) is performed.

The polarization layer 304, which is the inspection object 300, is extracted as an image image from the adhered display panel 302 and substituted with the modified test data 320 on the feature space 400.

The failure discrimination step (S200) of determining the area of the test data 320 on the feature space 400 based on the classification criterion 410 and determining the test object 300 as intrinsic or false caustic is performed. .

The classification criteria 410 on the feature space 400 is modified by applying the test data 320 to the pre-learning step S100, but the test data 320 is the same number as the true data 120 and the false data 220. Further learning step (S300) is corrected through the data balancing step (S310) to form a first data group, the failure determination step (S200) and the additional learning step (S300) is repeatedly performed.

In the following, each step described above is described in detail.

The first learning group includes the same number of intrinsic samples 100 as samples of inferior goods and pseudo samples 200 as samples of inferior goods as A.

The first learning image group includes the same number of the true image 110 derived from the true sample 100 and the pseudo image 210 derived from the pseudo sample 200.

The first data group includes the same number of intrinsic data 120 that modifies the intrinsic image 110 and caustic data 220 that modifies the pseudo image 210.

Therefore, in one embodiment of the present invention, the true sample 100 and the true image 110 are the same A, and the pseudo sample 200 and the pseudo image 210 are the same A.

The feature space 400 is a multidimensional space having a plurality of variables and refers to a space where specific data is plotted according to an embodiment of the present invention.

In the pre-learning step (S100), the image collecting step (S110), which acquires the true image 110 and the false image 210, which are image images of points suspected to be defective from the true sample 100 and the false sample 200, is preliminary. A creation step S120 and a class classification step S130 are included.

3 is an image showing a true image of a true sample in the learning-based vision inspection method through data balancing according to an embodiment of the present invention, Figure 4 is a learning-based through data balancing according to an embodiment of the present invention In the vision inspection method, pseudo images of pseudo samples are shown.

3 to 4, in the image collecting step S110, an image is collected through the photographing unit 520 moving and moving the measurement target planes of the true sample 100 and the false sample 200 in a predetermined pattern. In the derivation of a result that satisfies a predetermined condition in an image of a specific position, the image is captured at the position to acquire the true image 110 or the false image 210.

As an embodiment of the present invention, when a change in brightness of an image captured by the photographing unit 520 rapidly increases or decreases, the spot may be photographed as shown in FIGS. 3 to 4 to obtain an image image. have.

In the pre-creation step S120, the intrinsic data 120 and the pseudo data 220 derived from n feature points are derived from the intrinsic image 110 and the pseudo image 210, respectively.

A × n pieces of intrinsic data 120 are obtained, and A × n pieces of pseudo data 220 are obtained.

5 is a schematic diagram showing a process of deriving a dictionary from a learning-based vision inspection method through data balancing according to an embodiment of the present invention, and FIG. 6 is a learning-based through data balancing according to an embodiment of the present invention. It is a schematic diagram showing the process of assigning the first group of learning images in advance in the vision inspection method.

5 to 6, the intrinsic data 120 and the pseudo data 220 derived from the intrinsic image 110 and the pseudo image 210 are colored in the intrinsic image 110 and the pseudo image 210. Features such as brightness, saturation, shape, curvature, and the like may be extracted and collected at n points.

The intrinsic data 120 and the false data 220 are converted into a vector amount having characteristics such as hue, brightness, saturation, shape, curvature, etc. as variables, and multidimensional space corresponding to the variables constituting the vector amount. Is plotted on the phase.

FIG. 5 briefly illustrates a process in which n feature points are derived as vector quantities from the intrinsic image 110 and the pseudo image 210, and are illustrated in a multi-dimensional space.

This is exemplary and the authentic data 120 and the false data 220 may be extracted based on various factors according to an embodiment to which the present invention is applied.

As described above, A × n true data 120 from A true images 110 and A × n false data 220 from A pseudo images 210 are collected in a vector amount, and the true data 120 And pseudo data 220 are plotted on a multidimensional space as described above, grouping adjacent intrinsic data 120 and pseudo data 220 into k groups, respectively, and intrinsic data belonging to each group ( 120 and the average points of the caustic data 220 are obtained to calculate k average points.

At this time, each group is defined as a word and k average points are defined as a dictionary to create a dictionary having k words.

The pre-creation step S120 may be performed using a k mean clustering algorithm which clusters the given data into k clusters.

The class classification step S130 substitutes the authentic data 120 into the dictionary created through the dictionary creation step S120 to histogram the authentic data 120 belonging to each of the k words, as shown in FIG. 6. The histogram step is performed first.

7 is a schematic diagram showing a pre-learning step and a failure discrimination step in a learning-based vision inspection method through data balancing according to an embodiment of the present invention, and FIG. 8 is a diagram illustrating learning through data balancing according to an embodiment of the present invention. It is a graph showing three-dimensional examples of the state in which the intrinsic region and the pseudo region are derived from the feature space in the vision inspection method.

Based on the histogram obtained through the above histogramization step, the word and the intrinsic data 120 belonging to each word are shown in the feature space 400 as a variable, and as shown in FIG. 8, the feature space ( The intrinsic region 130 is derived on 400.

In addition, the histogram step of histogramting the pseudo data 220 belonging to each of the k words by assigning the pseudo data 220 to the dictionary created through the dictionary creation step S120 is performed as shown in FIG. 6. do.

Based on the histogram obtained through the caustic data 220, the word and the causal data 130 belonging to each word are shown as a variable on the feature space 400, and as shown in FIG. The caustic region 230 is derived on 400.

The classification criterion 410 may be expressed as a line, a plane, or a hyper plane that is a boundary between the intrinsic region 130 and the caustic region 230 on the feature space 400, and the intrinsic region 130 and the caustic region. The 230 may be obtained using a support vector machine (SVM) technique that obtains a boundary of each other in the feature space 400 so as to maximize each other.

Defective determination step (S200) is consistent with the preset conditions by continuously photographing the surface on which the polarizing layer 304 is bonded to the display panel 302, which is the inspection object 300, along a predetermined path through the photographing unit 520. A test image collection step (S210) of acquiring a test image 310, which captures an image of a location indicating a value, and the intrinsic image 110 and the false image 210 described above with respect to the test image 310. And converting into a test image 310 by using the same method as the method of converting to the caustic data 220, and inserting the test image 310 into a dictionary to be shown on the feature space 400, and classifying step S130. Determining whether it belongs to the intrinsic region 130 or the caustic region 230 on the basis of the classification criteria 410 obtained through the (), the failure discrimination step according to the classification criteria to determine whether it is an intrinsic defective or caustic defective according to the belonging region ( S220).

In the additional learning step S300, a step of correcting the dictionary may be added by reprocessing k average points by adding the test image 310 acquired through the failure discriminating step S200 in advance.

Alternatively, the classification standard correction step (S320) of adding the test data 320 obtained through the failure determination step (S200) to the feature space and correcting the classification criteria 410 by reflecting the added test data 320 is Can be performed.

The step of correcting such a dictionary, the classification standard correction step (S320) is an effect of increasing the learning sample by adding new information obtained by performing a bad discrimination to the bad discrimination criteria previously learned through the pre-learning step (S100). Can be obtained.

By increasing the learning sample, it is possible to increase the accuracy of performing the defect discrimination, and the boundary between the intrinsic defect and the caustic defect can be more clearly defined in the feature space 400, thereby reducing the phenomenon of caustic defect overdetection.

In addition, the additional learning step S300 may include a data balancing step S310.

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

Referring to FIG. 9, in the data balancing step S310, when the inspection object 300 acquired through the defective discrimination step S200 is an intrinsic defect, the feature space learned in the classification criterion correction step S320 may be used. The data added to 400) serves as the intrinsic data 120, and if the test object 300 is false, the data added to the feature space 400 learned in the classification standard correction step S320 is false. It serves as data 220.

An imbalance problem of data that may occur between classes classified as true data 120 or false data 220

It may be expressed as an equation, and may be expressed as a binary classification problem for finding a hyperplane on the feature space 400 that becomes an input space with respect to the input vector x.

The intrinsic data 120 or the false data 220 are classified into two classes of '+1' or '-1', and the class classification problem may be regarded as finding w to obtain excellent generalization performance.

In this case, the case where the ratio of the number of data belonging to the intrinsic data 120 and the causal data 220 is significantly different may be regarded as an imbalanced data problem.

When such an unbalanced data problem occurs, data is biased into one class of the intrinsic data 120 or the false data 220, which is a classification standard correction step of correcting the defective discrimination criteria previously learned through the pre-learning step (S100). S320) may result in lowering the reliability of the failure discrimination criteria.

Therefore, when the test data 320 obtained by either the true data 120 or the false data 220 is biased, the classification criteria 410 is the true region (as shown in FIG. 9A). 130 or the caustic region 230 may be biased toward either side.

In particular, when the defect inspection of the polarization layer 304 of the display panel 302 is performed, the detection frequency of the intrinsic defect is significantly higher than that of the caustic defect, and the bias of such data may act as a deterrent to improving the accuracy of the defect discrimination. Can be.

Therefore, when the test data 320 obtained in the failure discrimination step S200 belongs to the intrinsic region 130 on the feature space 400, randomly selected among the caustic data 220 belonging to the caustic region 230 to correspond to this. One caustic data 220 is added to the feature space 400 together with the acquired test data 320 to correct the classification criteria 410.

In detail, when the inspection of the inspection object 300 is performed through the defective determination step S200, a total of four intrinsic defect determinations are derived, the caustic data 220 previously derived together with the obtained four inspection data 320. Is added to the feature space 400 to perform the process of resetting the classification criteria 410.

If the false negative determination is derived by performing the inspection of the inspection object 300 through the failure determination step (S200), the derived intrinsic data 120 is added to the feature space 400 together with the inspection data 320. To reset the classification criteria (410).

As another embodiment, if the true defect determination is derived by performing the inspection of the inspection object 300 through the failure determination step (S200), the average value of the acquired caustic data 220 together with the test data 320 is characterized. The classification criteria 410 are reset in addition to the space, and when deriving the false negative determination, the average value of the acquired true data 120 is added together with the test data 320 in the feature space to classify the classification criteria 410. Can be reset.

In this embodiment, since the data balancing is performed by obtaining the average value of the data obtained by the simple operation only, without performing the process of selecting data at random, it is possible to simplify the calculation process and improve the processing speed.

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

As shown in FIG. 10, the learning-based vision inspection apparatus based on data balancing according to an embodiment of the present invention may include an inspection surface of a transport means 510 and an inspection object 300 to which the inspection object 300 is transferred. The reading unit 520 and the reading unit 530 which select a test image 310 by reading an image image acquired through the photographing unit 520 and the photographing unit 520 that move continuously along a predetermined path. A storage unit for storing the test data 320 obtained through the operation unit 540 and the operation unit 540 to plot the test data 320 obtained by modifying the test image 310 obtained through the feature space 400 ( 550, and the operation unit 540 determines whether the inspection object 300 is defective in comparison with the data previously stored in the storage unit 550 and the test data 310, and the storage unit 550 includes the test data ( 320 and virtual data derived through the data 320 may be stored together.

In detail, the reading unit 530 performs the image collecting step S110 and the test image collecting step S210 described above, and captures an image obtained by continuously photographing the test object 300 through the photographing unit 520. The image is obtained by analyzing the position corresponding to the preset condition.

The calculation unit 540 is a pre-creation step (S120), class classification step (S130), failure determination step (S220) according to the classification criteria is performed.

In the storage unit 550, an additional learning step S300 including a data balancing step S310 and a classification criterion correction step S320 is performed.

As described above, a preferred embodiment according to the present invention has been described, and the fact that the present invention can be embodied in other specific forms in addition to the above-described embodiments without departing from the spirit or scope thereof has ordinary skill in the art. It is obvious to them. Therefore, the above-described embodiments should be regarded as illustrative rather than restrictive, and thus, the present invention is not limited to the above description and may be modified within the scope of the appended claims and their equivalents.

Claims

A learning-based vision inspection method for inspecting defects of a polarizing layer adhered to a display panel,

A first learning group is formed including the same number of intrinsic samples as the samples of inferior goods and the false samples as samples of the inferior goods, and the intrinsic data and the caustic data in which the image images collected from the first learning groups are modified are provided in the same number. Extracting a first data group, and drawing the first data group on a feature space to derive a classification criterion that is a boundary between an intrinsic region in which the intrinsic data is located and a causal region in which the pseudo data is located;

The image data is extracted from the display panel to which the polarization layer, which is an object to be inspected, is attached to the feature space, and is modified into the feature space, and the area where the test data is located in the feature space is determined based on the classification criteria. A defect discrimination step of judging the inspection object as true or false; And

The classification data on the feature space is modified by applying the test data to the pre-learning step, wherein the test data is corrected through a data balancing step so that the true data and the false data constitute the first data group with the same number. A further learning step;

The learning-based vision inspection method through data balancing in which the failure discrimination step and the additional learning step are repeatedly performed.
The method of claim 1,

The pre-learning step,

An image collection step of collecting a first learning image group consisting of an intrinsic image and a pseudo image extracted as the image information from each of the intrinsic specimens and the pseudo specimens included in the first learning group;

The intrinsic data and the pseudo data are derived from each of the intrinsic image and the pseudo image included in the first learning image group as n feature points, respectively, to form the first data group and to diagram the first data group. A dictionary generation step of grouping adjacent feature points into k groups, obtaining a mean of feature points for each group, and creating a dictionary including k mean points; And

The class classification is performed by mapping the intrinsic data and the causal data into the dictionary to determine a classification criterion that is a boundary between the intrinsic region in which the intrinsic data is located and the intrinsic region in which the caustic data is located. step;

Learning-based vision inspection method through data balancing comprising a.
The method of claim 2,

The dictionary creation step,

A learning-based vision inspection method through data balancing for performing the k-means clustering to create the dictionary with k mean points as the reference word.
The method of claim 2,

The failure determination step,

Learning through data balancing which compares the test data which are n feature points obtained from the test image extracted from the test object as image information from the dictionary and classifies the test data on the basis of the k average points and plots them on the feature space. Based vision inspection method.
The method of claim 4, wherein

The further learning step,

Learning-based vision inspection method through data balancing for correcting the classification criteria by adding the test data plotted on the feature space through the bad discrimination step to the true data or the false data of the class classification step.
The method of claim 5,

The data balancing step,

If it is determined that the inspection object is an intrinsic defect through the failure discrimination step, the classification is performed by applying one randomly selected caustic data of the test data and the causal data previously derived through the pre-learning step to the class classification step. If the test object is judged to be false false through the failure determination step, the classifying step of randomly selected one of the authentic data from the test data and the intrinsic data derived through the pre-learning step Learning-based vision inspection method through data balancing to apply the correction to the classification criteria.
The method of claim 5,

The data balancing step,

If it is determined that the inspection object is an intrinsic defectiveness through the failure determination step, the classification criteria are corrected by applying the average value of the caustic data derived through the test data and the pre-learning step to the class classification step, and If the inspection object is determined to be false through a bad discrimination step, data balancing is performed by applying an average value of the intrinsic data derived through the test data and the pre-learning step to the class classification step to correct the classification criteria. Learning-based vision testing through learning.
The method of claim 4, wherein

The true image, the pseudo image and the test image,

A learning-based vision inspection method using data balancing, which is image information of a point at which a modified signal of a photographed image is rapidly changed while moving the inspection target surface of the inspection object according to a predetermined pattern.
Transport means for transporting the inspection object;

A photographing unit for photographing an inspection surface of the inspection object;

A reading unit which selects an image to be examined by reading an image image obtained through the photographing unit;

An operation unit which plots the test data obtained by modifying the test image obtained through the reading unit on a feature space; And

And a storage unit for storing the test data plotted through the operation unit.

The operation unit compares the data previously stored in the storage unit with the illustrated test data and determines whether the inspection object is defective, and the storage unit stores data together with the virtual data derived through the test data and the test data. Learning-based vision inspection device through balancing.