KR102168724B1 - Method And Apparatus for Discriminating Normal and Abnormal by using Vision Inspection - Google Patents

Abstract

Disclosed are a method for determining abnormality using image inspection and an apparatus thereof. The present embodiment may comprise: a defect suspicious determination unit which receives an original image from a photographing unit, generates a tile image by dividing the original image by a preset tile unit according to a size of the original image, and determines a suspicious defect area by primarily detecting whether a defect in the tile image is detected using machine vision; and a defect determination unit which selects an adjacent area based on a position of the suspicious defect area, sets the position of the suspicious defect area and the adjacent area as a defect determination area, and determines whether or not the defect is defective using only the defect determination area using an artificial intelligence model. According to the present invention, computation is fast due to low complexity of image analysis.

Description

Method and Apparatus for Discriminating Normal and Abnormal by using Vision Inspection}

An embodiment of the present invention relates to a method and apparatus for determining an abnormality using image inspection.

The contents described below merely provide background information related to the present embodiment and do not constitute the prior art.

In general, before food, semiconductor, and electronic devices are shipped, it is essential to check the condition of products by inspecting defects. Conventionally, a generally practiced inspection method takes a lot of time to inspect a large amount of products by manually performing an inspection by a plurality of personnel in a workplace using inspection equipment, and there is a limit to the accuracy of determining whether or not it is normal.

To solve the above problem, there is a method of inspecting the product by adsorbing the product using a circular plate or drum and then rotating it. However, this method can also be carried out faster than manual inspection, but it is possible to inspect a large amount of product. It takes a lot of time and the accuracy is poor.

In this embodiment, an original image photographed of an object to be inspected is divided into a preset tile unit, and a primary inspection is performed to detect a defective region using machine vision, A method and apparatus for detecting defects by setting the location of the suspected defect area and the adjacent area as a defect location candidate, and performing a secondary inspection to detect a defect determination area using a previously learned artificial intelligence model. It has a purpose to provide.

According to an aspect of the present embodiment, the vision inspection apparatus receives an original image from a photographing unit, generates a tile image by dividing the original image into a preset tile unit according to the size of the original image, and machine vision Using (Machine Vision), a defect suspicious determination unit that determines a suspicious area by first detecting whether or not a defect in the tile image is used, and an adjacent area is selected based on the location of the suspicious defect area, and The location and the adjacent area may be set as a defect determination area, and a defect determination determination unit may be configured to determine whether or not a defect is determined for only the defect determination area using an artificial intelligence model.

Meanwhile, the defect determination determining unit may recognize a contour of a suspected defect object within the suspected defect region, and determine the suspected defect object as a defect determined object or a defect unconfirmed object based on the contour of the suspected defect object.

In addition, the defect determination determination unit determines the suspected defect object as a defect determined object or a defect unconfirmed object based on a probability value (non-defect probability) and a probability value (defect probability), and the probability value (non-defect probability) is the It is a possibility that it is not a defect, and a probability value (defect probability) is a possibility that the suspected defect object is a defect, and the defect determination determination unit may generate defect type information and defect characteristic information for the defect determination object.

In addition, the vision inspection apparatus further includes a product defect determination unit that determines whether or not a product is defective based on information on a defect determination object from the defect determination unit, defect type information on the defect determination object, and defect characteristic information. I can.

According to another aspect of the present embodiment, in the vision inspection method, the vision inspection device receives the original image from the photographing unit, and the suspected defect determination unit divides the original image into a preset tile unit according to the size of the original image. A tile image is created, and the step of determining a suspected defect area by first detecting whether a defect in the tile image using Machine Vision is used, and a defect determination unit adjacent to the location of the suspected defect area Selecting an area, setting the location of the suspected defect area and the adjacent area as a defect determination area, and determining whether or not a defect is a defect only for the defect determination area using an artificial intelligence model.

In addition, the defect determination determining unit may recognize a contour of a suspected defect object within the suspected defect region, and determine the suspected defect object as a defect determination object or a defect unconfirmed object based on the contour of the suspected defect object.

In addition, the defect determination determination unit determines the suspected defect object as a defect determined object or a defect unconfirmed object based on a probability value (non-defect probability) and a probability value (defect probability), and the probability value (non-defect probability) is the It is a possibility that it is not a defect, and a probability value (defect probability) is a possibility that the suspected defect object is a defect, and the defect determination determination unit may generate defect type information and defect characteristic information for the defect determination object.

In addition, the product defect determination unit may further include determining whether the product is defective based on information on the defect determination object, defect type information on the defect determination object, and defect characteristic information.

As described above, according to the present embodiment, the original image photographing the object to be inspected is divided into a preset tile unit, and a primary inspection is performed to detect a suspected defect region using machine vision. It has the effect of detecting a defect by performing a second inspection on the defect judgment area by setting a defect judgment area for the suspected defect object and an adjacent area, and using a previously learned artificial intelligence model.

According to the present exemplary embodiment, since a high-resolution image is divided into predetermined tile units to detect a defect area, and then an abnormality is determined, the complexity of image analysis is low, so that a fast operation is possible.

1A and 1B are diagrams illustrating an abnormality determination system using image inspection according to the present embodiment.
2 is a schematic diagram of a vision inspection apparatus according to the present embodiment.
3 is a diagram illustrating a method of dividing an original image into a preset tile unit according to the present embodiment.
4 is a view showing a vision inspection apparatus according to the present embodiment.
5 is a diagram illustrating an example of determining a defect type according to the present embodiment.
6 is a conceptual diagram illustrating a method of generating an image of a suspected defect region according to the present embodiment.
7 is a conceptual diagram illustrating a method of generating an artificial intelligence learning model of a defect determination unit according to the present embodiment.
8 is a conceptual diagram showing the operation of the product defect determination unit according to the present embodiment.

Hereinafter, this embodiment will be described in detail with reference to the accompanying drawings.

1A and 1B are diagrams illustrating an abnormality determination system using image inspection according to the present embodiment.

An abnormality determination system using image inspection according to the present embodiment includes a photographing unit 110 and a vision inspection device 120. Components included in the abnormality determination system using image inspection are not necessarily limited thereto.

The photographing unit 110 refers to a photographing module that photographs an image or an image, including a camera module. The photographing unit 110 is installed at a specific location to generate a photographed image photographing a preset location.

The photographing unit 110 includes a camera, a lens, and a light. The camera and lens included in the photographing unit 110 are preferably high-resolution cameras and lenses for machine vision, but are not limited thereto. When the photographing unit 110 receives a signal to which the object to be inspected on the conveyor belt enters from a separate sensor, it emits light and then generates a photographed image of the object to be inspected using a camera and a lens.

Preferably, the photographing unit 110 generates a photographed image by photographing an object to be inspected on a conveyor belt, as shown in FIGS. 1A and 1B. The photographing unit 110 resizes a photographed image of an object to be inspected in high resolution, or transmits it to the vision inspection apparatus 120 in a state of an original image without deteriorating the resolution. The photographing unit 110 communicates with the vision inspection apparatus 120 by wire or wirelessly, and transmits an original image photographing an object to be inspected to the vision inspection apparatus 120.

The vision inspection apparatus 120 performs primary image analysis using machine vision, and determines whether there is a defect by performing secondary image analysis using artificial intelligence. The primary image analysis used in the vision inspection apparatus 120 may be an image analysis that does not use separate machine learning. The secondary image analysis may be an analysis of whether there is a defect based on a learning model based on machine learning.

For example, the vision inspection device 120 not only determines the presence of foreign substances based on the photographed images of food and beverage products based on the first image analysis and the second image analysis, but also determines whether there is an abnormality based on a photographed image of a semiconductor or display. Can be determined. When the vision inspection device 120 is applied to a food product manufacturing process, it is possible to detect a case in which foreign substances (eg, iron pieces, plastic pieces) that cannot be consumed by humans are included in the food product manufacturing process.

The vision inspection apparatus 120 mounts an artificial intelligence model for secondary image analysis in software, and determines whether or not a defect (eg, contains foreign substances) in the received image by using the artificial intelligence model. The vision inspection apparatus 120 ships the product when a defect is not detected in the image (eg, foreign material is not present), and moves to a removal line when a defect is detected (eg, foreign material is present) in the image.

For example, when the vision inspection device 120 is applied to the manufacturing process of a manufactured product, the vision inspection device 120 may determine whether there is an abnormality based on the original image photographed of the manufactured product. In other words, it is possible to determine whether the binding of the manufactured product exists in a normal position or whether an oil seal exists at a preset position inside the cylinder. The vision inspection apparatus 120 according to the present embodiment may acquire an image photographed of an object to be inspected, analyze the image, and automatically determine whether the object to be inspected is defective.

2 is a schematic diagram of a vision inspection apparatus according to the present embodiment.

The vision inspection apparatus 120 according to the present embodiment includes a defect suspect determination unit 210, a defect determination determination unit 220, and a product defect determination unit 230. Components included in the vision inspection apparatus 120 are not necessarily limited thereto.

Each component included in the vision inspection device 120 is connected to a communication path connecting software modules or hardware modules inside the device, so that they can operate organically with each other. These components communicate using one or more communication buses or signal lines.

Each component of the vision inspection apparatus 120 shown in FIG. 2 refers to a unit that processes at least one function or operation, and may be implemented as a software module, a hardware module, or a combination of software and hardware.

The suspected defect determination unit 210 receives an original image from the photographing unit 110. The defect suspicious determination unit 210 checks the size of the original image and generates a tile image by dividing the original image into a preset tile unit according to the size of the original image. The suspected defect determining unit 210 may detect whether or not a defect is in the tile image using machine vision to generate an image of a suspected defect area including a suspected defect object.

The defect determination determining unit 220 determines a defect determination region including an adjacent region based on the position of the suspected defect object on the image of the suspected defect region. The defect determination determining unit 220 determines whether a defect is a defect only object in the defect determination region using a previously learned artificial intelligence model, and determines a defect determined object and a defect unconfirmed object.

The product defect determination unit 230 may finally determine whether to combine the product based on the defect determination object determined by the defect determination unit 220. The final defect of a product can be determined depending on the type of defect and its sensitivity to defects.

The defect suspicious determination unit 210 according to this embodiment primarily performs defect suspicious determination based on machine vision without separate machine learning, and the defect determination determination unit 220 determines suspicious defects based on machine learning. It may be determined whether or not a defect for a suspected defect object determined to be a defect existing in the area map that is primarily determined by the unit 210. Through this method, it is possible to detect defects more quickly and accurately by applying artificial intelligence machine vision with a large amount of computation in the second order only to the defect determination area including the suspicious object and only for the defect determination area including the suspicious object. Judgment may be possible.

More specifically, the defect suspicious determination unit 210 may perform a rough primary analysis on the entire input image by applying a machine vision technology to which a separate machine learning is not applied. Candidate areas that are suspected to be defective are detected, and the suspicious area including a suspected defect object having a low probability may be determined by adjusting the sensitivity.

The suspected defect determination unit 210 analyzes an edge, intensity distribution, etc. by applying a machine vision technology that does not use machine learning, and determines the suspected defect object in an area different from the distribution of the normal image. It can be designated as a suspected defect area to include.

The suspected defect determination unit 210 detects a plurality of image patches of size W2 x H2 centered on the location of the pixel suspected to be a defect (the area of the suspected defect) through the primary image analysis, and machine learning is applied artificial intelligence. It may be transmitted to the defect determination unit 220, which is a secondary image analysis module.

The defect determination determining unit 220 may perform a final determination as to whether or not a defect is present and a type of defect based on the image of the suspected defect region including the suspected coupling region received from the suspected defect determining unit 210. In more detail, the defect determination determining unit may perform determination of a defect using classification, object detection, and segmentation techniques.

The final inspection result information of the defect determination unit 220 may classify a suspected defect object as a normal (or non-defective object) probability value (probability of nondefectiveness), and a suspected defect object as a defect (defect determination object). It may include information on a probability value (defect probability). The defect determination determining unit 220 may determine a suspected defect object as a defect unconfirmed object or a defect determined object based on the non-defect probability and the defect probability. In addition, the final inspection result information may further include defect type information and defect characteristic information corresponding to the defect determination object when the suspected defect object is classified as a defect determination object. The defect characteristic information may include information on characteristics of a defect such as a defect location, a defect length, and a defect area.

The product defect determination unit 230 may determine whether to determine whether to determine a product having a defect determination object as a product (defective) or a product (non-defective) based on defect type information and defect characteristic information of the defect determination object. Defects of products may be judged differently depending on the type of defect and characteristics of defects for each product.

3 is a diagram illustrating a method of dividing an original image into a preset tile unit according to the present embodiment.

The vision inspection apparatus 120 according to the present embodiment checks the size of the photographed original image (high-resolution image). When the captured original image is, for example, a high-resolution image of 8,000×8,000 (pixel), the vision inspection apparatus 120 divides the captured original image into a preset tile unit according to the size of the captured original image (8,000×8,000).

Since the vision inspection device 120 determines whether there is a defect in the image in units of a preset tile, when determining whether there is an abnormality in the input high-resolution original image (8,000 × 8,000), the amount of calculation is reduced, so that fast calculation is possible. .

The defect suspicious determination unit 210 of the vision inspection apparatus 120 may determine whether a defect in the image is detected in units of tiles based on slicing tiling and determine the image of the suspected defect area.

The defect determination unit 220 of the vision inspection apparatus 120 may perform object detection to determine an object determined to be a defect on an image of a suspected defect area in units of tiles. The object detection may detect a suspected defect object and display a separate first contour in an area including the suspected defect object. For example, the first contour may have a rectangular shape. The first contour including the suspected defect object may be a defect determination area.

Thereafter, instance segmentation is performed on the defect determination region to determine the boundary line of the object as the second contour, and detailed information on the defect may be extracted. An object judged to be an actual defect among suspected defect objects may be determined as a defect determination object. Conversely, an object that is judged not to be an actual defect among suspected defect objects may be determined as a defect unconfirmed object.

The vision inspection apparatus 120 may learn about an artificial intelligence model by receiving feedback on information about a defect-determined object and/or a defect-undefined object among the suspected defect objects.

Specifically, separate machine learning is performed on the defect-determined object and/or the defect-unconfirmed object of the vision inspection apparatus 120, and the learning result of the defect-determined object and/or the defect-unconfirmed object is applied to the suspected defect determination unit 210. I can. The accuracy of extraction of the suspected defect object by the suspected defect determining unit 210 may be increased based on a learning result of the object not determined to be defective and/or the object determined to be defective.

4 is a view showing a vision inspection apparatus according to the present embodiment.

In FIG. 4, a method of learning an artificial intelligence model of a defect determination unit for analyzing a secondary image in a vision inspection apparatus is disclosed.

The vision inspection apparatus 120 performs primary image analysis for extracting a defect candidate image patch DB using machine vision technology. The primary image analysis may be performed by the above-described defect suspicious determination unit 210.

The vision inspection apparatus 120 includes an image processor and software for machine vision for primary image analysis. The vision inspection apparatus 120 emits light from the photographing unit 110 to an object to be inspected on a conveyor belt, and then acquires a photographed image of the object to be inspected using a camera and a lens. The vision inspection apparatus 120 performs image processing and analysis on the captured image received from the photographing unit 110 using an image processor and software.

The vision inspection apparatus 120 extracts an image patch for a defect candidate image using machine vision. The vision inspection apparatus 120 adds a label for learning to the defect candidate image patch DB. The label contains information such as defect type, defect location, and defect shape. Information about the defect candidate image patch and the label for the defect candidate image patch is input, and an artificial neural network may be trained based on such supervised learning.

5 is a diagram illustrating an example of determining a defect type according to the present embodiment.

The vision inspection device 120 receives an original image from the photographing unit 110. The vision inspection apparatus 120 checks the size of the original image and generates a tile image by dividing the original image into a preset tile unit according to the size of the original image.

The vision inspection apparatus 120 determines a defect suspicious area in which a defect in the tile image is detected using machine vision. The vision inspection apparatus 120 recognizes the outline of the instance by performing instance segmentation for distinguishing objects (or instances) within the tile image.

The vision inspection apparatus 120 detects a defect based on the contour of each instance and an artificial intelligence learning model using machine vision. The vision inspection apparatus 120 determines defect type information based on the contours of each instance. As described above, the vision inspection device 120 performs instance segmentation on the defect determination area to determine the defect type based on the boundary line of the object and the artificial intelligence learning model, as described above, as pinhole, stripe, bubble, white point, fold, ignorance contamination, and ignorance. It can be determined by any one of minor wrinkles and taping marks.

6 is a conceptual diagram illustrating a method of generating an image of a suspected defect region according to the present embodiment.

In FIG. 6, a method for dividing an original image in units of tiles and determining an image of a suspected defect region in which a defect exists in the tile image is disclosed. Referring to FIG. 6, a tile size for determining an image of a suspected defect region may be adjusted based on a determination result of an existing defect and a defect extraction sensitivity 640. The existing defect determination result may include defect location information 610, defect size information 620, and defect type information 630.

According to the present embodiment, the size of the divided tile may be adjusted to be relatively larger as the position where defects do not occur frequently on the original image. Conversely, the size of the tile to be divided may be adjusted to be relatively smaller as the location where defects frequently occur on the original image.

In addition, the size of the tile may be adjusted to be relatively larger as the size of the defect is relatively large based on the size of the defect occurring in a specific location. Conversely, as the defect size is relatively small, the tile size can be adjusted to be relatively small.

Also, as the type of defect occurring in a specific location has a relatively small influence on the quality of the product, the size of the tile may be adjusted to be relatively larger. Conversely, as the type of defect has a relatively large influence on the quality of the product, the size of the tile can be adjusted to be relatively smaller.

In this way, the speed of defect determination and accuracy of defect determination may be further increased by adjusting the tile size in consideration of the area requiring more determination and the size of the defect.

In addition, according to the present embodiment, as the defect extraction sensitivity 640 is set relatively high, the tile size is set to be relatively small. Conversely, as the defect extraction sensitivity 640 is set relatively low, the tile size is relatively large. Can be set.

In addition, in the present embodiment, as the reliability of the defect determination 650 for the type of defect is lower, the size of the tile for the corresponding portion may be set to be relatively small. As a result of learning about the artificial intelligence model, the defect determination reliability of the artificial intelligence model may be different for each defect type, and the smaller the defect determination reliability is, the smaller the tile size may be for more accurate determination.

The adjustment of the tile size for the original image as described above may be performed by continuously reflecting the determination result for the defect determination object as described above.

7 is a conceptual diagram illustrating a method of generating an artificial intelligence learning model of a defect determination unit according to the present embodiment.

In FIG. 7, a method for learning about a defect based on the classification of the defect determination unit is disclosed. Referring to FIG. 7, the defect determination determining unit may perform learning based on at least one classification model.

The first learning model 710 may be a model trained based on a defect image for a product and a normal image for a product, which are actual training data. The second learning model 720 may not be an actual defect determination product, but may be a model that is trained based on a defect image generated in a product similar to the product to be determined for defects.

A defect image of a product and a normal image of a product, which are actual training data, may be expressed in terms of first training data 715. An image of a defect generated in a product similar to a product to be determined that is not an actual defect determination product may be expressed in terms of second learning data 725.

The third learning model 730 may be a model in which training is performed based on the third training data 735 obtained by mixing the first training data 715 and the second training data 725. The mixing ratio of the third training data 735 may be determined based on the similarity between the product to be determined for defects and the product used when the second training data 725 is generated. The third learning data 735 may be generated by relatively increasing the mixing ratio of the second learning data 725 as the similarity between the product to be determined to be defective and the product used when the second training data 725 is generated is higher.

In the initial product inspection step, a determination of a defect may be performed based on a determination result of the first learning model 710, the second learning model 720, and the third learning model 730.

Thereafter, the accuracy of each defect type of the first learning model 710, the second learning model 720, and the third learning model 730 may be determined in a product inspection step in which the inspection of the first critical product number is completed. For example, for defect 1, the first learning model 710 and the third learning model 730 may exhibit an accuracy of a threshold or higher, and the learning model having the highest accuracy may be the third learning model 730. have. With respect to defect 2, the first learning model 710 and the second learning model 720 may exhibit an accuracy greater than or equal to a threshold value, and the learning model having the highest accuracy may be the second learning model 720.

In an initial product inspection step in which the first critical number of products has been inspected, a model having a high accuracy is actually extracted, and feedback for additional product inspection results is input for only the corresponding model to perform additional learning.

Thereafter, in the product inspection step in which the inspection of the second critical product number is completed, only the learning model having only the highest accuracy may be selected to perform more accurate determination for each defect type.

8 is a conceptual diagram showing the operation of the product defect determination unit according to the present embodiment.

In FIG. 8, a product defect determination operation based on a defect determination object of the product defect determination unit 230 is started. Referring to FIG. 8, as described above, the product defect determination unit 230 determines a product having a defect determination object as a product (defective) or a product (non-defective) based on defect type information and defect characteristic information of the defect determination object. You can decide whether to judge.

The defect type may be set to a defect type (defect characteristic not determined) 800 and a defect type (defect characteristic determination) 820. A product with a defect corresponding to the defect type (defect characteristic not determined) 800 may be determined as a product (defect) regardless of the size of the defect.

A product having a defect corresponding to the defect type (defect characteristic determination) 820 may be determined as a product (defective) or a product (non-defective) in consideration of defect characteristic information.

For each of the at least one defect type (defect characteristic determination), a defect determination threshold value for each of the at least one defect characteristic information may be set. For example, a threshold value for each defect characteristic information (defect location, defect length, defect area) can be set for defect type 1 (defect characteristic determination), and lower defect characteristics for defect type 2 (defect characteristic determination) A threshold value can be set for each of the information (defect location, defect length, defect area). This threshold may be adjusted according to the first sensitivity setting 850 of product defect determination.

The product defect determination unit 230 may determine whether the defect type (defect characteristic determination) exceeds a defect determination threshold, and determine the defect type (defect characteristic determination) present in the product.

The product defect determination unit 230 sets a weight for each defect type (defect characteristic determination) present in the product according to the secondary sensitivity setting 870 of product defect determination, and determines whether the product is finally determined as a product (defect). You can decide whether or not.

The above description is merely illustrative of the technical idea of the present embodiment, and those of ordinary skill in the technical field to which the present embodiment belongs will be able to make various modifications and variations without departing from the essential characteristics of the present embodiment. Accordingly, the present exemplary embodiments are not intended to limit the technical idea of the present exemplary embodiment, but are illustrative, and the scope of the technical idea of the present exemplary embodiment is not limited by these exemplary embodiments. The scope of protection of this embodiment should be interpreted by the following claims, and all technical ideas within the scope equivalent thereto should be construed as being included in the scope of the present embodiment.

110: photographing unit 120: vision inspection device
210: defect suspicious determination unit 220: defect determination determination unit
230: product defect determination unit

Claims (8)

When an original image is received from the photographing unit and a tile image is generated by dividing the original image into a preset tile unit according to the size of the original image, defect location information and defects included in the existing defect determination result Size information and defect type information are extracted, and based on the defect location information, the size of the tile to be divided is adjusted to be larger than a preset threshold size, as the location where defects do not occur frequently, based on the defect location information. If the defect size is larger than the preset threshold size, the tile size is adjusted to be larger than the preset threshold size, and the smaller the effect on the product quality based on the defect type information, the tile size is made larger than the preset threshold size. The tile size is adjusted to be larger than the preset threshold size as the preset defect extraction sensitivity is set lower, and the whole of the original image in the tile image is adjusted using Machine Vision without machine learning. A defect suspicious determination unit for determining a suspicious defect area including a suspicious defect object in an area different from the distribution of the normal image;
Selecting an adjacent area based on the location of the suspicious defect area, setting the location of the suspicious defect area and the adjacent area as a defect determination area, and learning based on the defect image that occurred in the actual product and the normal image of the product. A first training model that is trained with first training data that is actual training data, a second training model that is trained with second training data that is trained on the basis of a defect image that occurs in a product subject to defect determination and a similar product, the first training data, and Objects having defects having an accuracy of greater than or equal to a threshold value using at least two or more learning models among the third learning models learned by mixing the second training data with the third training data A defect determination unit configured to determine a defect determination object; And
If the defect characteristics of the defect-determined object is determined as information on the type of defect that has not been determined, it is determined as a defect of the product, regardless of the size of the defect. Product defect determination unit that determines whether a product is defective by applying a weight to the characteristics exceeding a preset threshold among the defect location, defect length, and defect area.
Vision inspection device comprising a. The method of claim 1,
The defect determination determining unit,
Vision inspection, characterized in that the contour of the suspected defect object is recognized within the suspected defect area, and the suspected defect object is determined as the defect determined object or the defect unconfirmed object based on the contour of the suspected defect object Device. The method of claim 2,
The defect determination determining unit determines the suspected defect object as the defect determined object or the defect unconfirmed object based on a probability value (non-defect probability) and a probability value (defect probability),
The probability value (probability of non-defectiveness) is the probability that the suspected defective object is not a defect,
The probability value (defect probability) is a probability that the suspected defect object is a defect. delete Receiving an original image from a photographing unit in a suspected defect determination unit, and generating a tile image by dividing the original image into a preset tile unit according to the size of the original image;
Extracting defect location information, defect size information, and defect type information included in an existing defect determination result when the suspected defect determination unit generates a tile image;
Based on the defect location information, the defect suspicious determination unit adjusts the size of the divided tile to be larger than a preset threshold size, as the location where defects do not occur frequently, based on the defect location information, and the defect size is determined based on the defect size information. If it is larger than the set threshold size, the size of the tile is adjusted larger than the preset threshold size, and the smaller the effect on product quality based on the defect type information, the larger the tile size is adjusted than the preset threshold size. Adjusting the size of the tile larger than the preset threshold size as the set defect extraction sensitivity is set lower;
The defect suspicious determination unit determines a defect suspicious area including a suspicious defect object in an area different from the distribution of the normal image within the tile image for the whole of the original image using Machine Vision to which machine learning is not applied. The process of doing;
Selecting an adjacent area based on the position of the suspected defect area by a defect determination unit and setting the position of the suspected defect area and the adjacent area as a defect determining area;
A first learning model that learns from the first learning data that is actual learning data learned based on the image of a defect that occurred in the actual product by the defect determination unit, the normal image of the product, and the image of a defect that occurs in a product similar to the product subject to defect determination. At least two or more of a second learning model trained with the second training data learned based on and a third training model trained with the third training data mixed with the first training data and the second training data Determining an object having a defect having an accuracy equal to or greater than a threshold value as a defect determination object using only the defect determination region;
Determining a defect of a product irrespective of the size of the defect when the defect determination unit determines the defect characteristics of the defect determination object as information about the type of defect that has not been determined; And
When the product defect determination unit determines the defect type information for the defect-determined object, a weight is applied to a property exceeding a preset threshold among the defect location, defect length, and defect area included in the defect property information. The process of determining whether there is a defect
Vision inspection method comprising a. The method of claim 5,
The defect determination determining unit recognizes a contour of a suspected defect object within the suspected defect region, and determines the suspected defect object as the defect determined object or the defect unconfirmed object based on the contour of the suspected defect object. Vision inspection method characterized by. The method of claim 6,
The defect determination determining unit determines the suspected defect object as the defect determined object or the defect unconfirmed object based on a probability value (non-defect probability) and a probability value (defect probability),
The probability value (probability of non-defectiveness) is the probability that the suspected defective object is not a defect,
The probability value (defect probability) is a probability that the suspected defect object is a defect.
delete

Images