KR102524151B1 - Labeling learning method of artificial intelligence machine for smart products inspection - Google Patents

Abstract

A labeling learning method of an artificial intelligence machine for smart vision inspection of production parts is disclosed. The labeling learning method of an artificial intelligence machine for smart vision inspection of production parts of the present invention is a labeling learning method of an artificial intelligence machine using a smart vision inspection system, comprising: (a) the artificial intelligence machine receiving image data; (b) generating, by an artificial intelligence machine, first labeling data indicating a defective area of a part, storing the first labeling data in conjunction with image data, and transmitting the first labeling data to a first labeling device; (C) when the first approval information is input to the first labeling device, the step of the first labeling device uploading the first labeling data to the artificial intelligence machine; (d) generating, by an artificial intelligence machine, second labeling data indicating a defective area, storing the second labeling data in conjunction with image data, and transmitting the second labeling data to a second labeling device; (E) when the second approval information is input to the second labeling device, the step of the second labeling device uploading the second labeling data to the artificial intelligence machine; and (f) learning, by the artificial intelligence machine, the first labeling data and the second labeling data stored in association with the image data.

Description

Labeling learning method of artificial intelligence machine for smart vision inspection of production parts {LABELING LEARNING METHOD OF ARTIFICIAL INTELLIGENCE MACHINE FOR SMART PRODUCTS INSPECTION}

The present invention relates to a labeling learning method of an artificial intelligence machine for smart vision inspection of production parts, and more particularly, to a smart vision inspection of production parts in which an artificial intelligence machine automatically recognizes physical defects of production parts. It is about a labeling learning method for artificial intelligence machines.

The machine vision-based automation market and the artificial intelligence-based vision system market are rapidly increasing. The global machine vision market is expected to grow at an average annual rate of 8.15% from $9.02 billion in 2016 to reach about $14.43 billion in 2022.

Artificial intelligence and vision-based quality inspection systems are key elements in industrial automation systems. Instead of inspecting the state of physical shape and defects, surface finish, color, etc. during the manufacturing process with the naked eye, inspection is conducted with a machine vision system consisting of cameras, lighting, and vision software to evaluate products and find defects, improve manufacturing productivity, This is because it provides excellent performance in activities such as data collection.

In particular, in the case of parts related to molds for expensive electronic products, such as power trains, brakes, and steering systems, where automobile safety and consumer demands are high, strict quality inspections are required, and vision quality inspections are essential for this.

AI vision detection technology requires labeling data for AI learning to learn AI models. The AI model learns the AI model through training data. At this time, the learning data provides data for distinguishing where the AI is located in the image for the object to be detected later and what kind of object it is.

The AI model training process determines the weight value for each node of the AI model to be optimized for object detection (evaluation) through a myriad of iterative input and validation processes of labeling data for training.

As shown in FIG. 7, the labeling task for collecting labeling data for AI learning can be typically classified into three types.

As shown in FIG. 7(a), the bounding box is a method of displaying an area of an object in a photo or video as a box. At this time, the labeling data consists of the bounding box's rectangular coordinate data (x, y, width, height) and data about the type of object. It can be processed faster than other methods and is a labeling method commonly used in various fields.

As shown in FIG. 7( b ), the image classification method is a method of classifying the meaning of an image that a person can feel when viewing an image. Labeling data is composed of classification data about the meaning of an image to be detected. For example, if the classification types of images to be detected are basketball and soccer, classification data of basketball is designated for an image corresponding to a basketball scene.

As shown in FIG. 7(c), the object area segmentation method is a method for specifying the area of an object in an image in units of pixels, and is composed of mask data for area pixels corresponding to an object and type data of the corresponding object. .

The object mask data has the same size as the target image and is composed of 0 and 1 for each pixel position. If the mask value is 1, the corresponding pixel is included in the object area, and if the mask value is 0, it means included in the background area. It is mainly used when AI detection with higher precision than other methods is required because it can accurately identify the area of an object.

In the conventional method, the task of labeling data for AI learning is focused on securing a lot of labeling data at low cost by putting in many people.

In particular, the crowd-sourcing AI learning data labeling platform, which is widely used recently, performs primary labeling on training data through general non-expert labelers of many people in low-wage regions or countries, and labeling semi-experts It is a method of securing secondary high-quality labeling data through primary labeling data inspection.

Crowd-sourced data labeling for AI learning is only applicable when labeling is possible by common sense knowledge of common people because many non-experts perform the labeling task, and it is difficult to apply this method if specialized knowledge about the labeling target is required. .

However, the AI learning labeling data required to perform vision quality inspection for defect detection of manufactured parts can only be judged by very limited people, such as quality personnel at the manufacturing plant where the corresponding part is actually produced, and such quality experts can be judged at the manufacturing plant. It is difficult to apply the existing data labeling method for AI learning as it is difficult to obtain from outside.

Since manufacturing parts have different defect type characteristics depending on the shape, material, and process even if they are the same part, uniform defect data labeling for parts cannot be performed, and defect data labeling is performed separately for each manufacturing process. Should be.

Therefore, the existing data labeling work for AI learning is to separately photograph and collect defective parts in factories or companies that produce each manufactured part, and the on-site quality manager in the manufacturing process sends each filmed video at an additional time outside of the production schedule. Let labeling of bad data be performed in

This method requires costs for additional work by field quality personnel, and since field quality personnel directly perform additional work outside of the production process, there is a problem that labeling data for AI learning takes a long time. This acts as a burden on companies or factories that want to introduce a vision quality inspection system, and becomes a factor that hesitates to introduce an actual vision quality inspection system.

An object of the present invention is to provide a labeling learning method of an artificial intelligence machine for smart vision inspection of production parts made to quickly build labeling data for AI learning even for production parts requiring specialized knowledge for labeling work.

The above object is, according to the present invention, a labeling learning method for an artificial intelligence machine using a smart vision inspection system, the smart vision inspection system comprising: a camera for generating image data of parts in a production line; a first labeling device receiving the image data; An artificial intelligence machine for transmitting and receiving data with the first labeling device; and a second labeling device for transmitting and receiving data to and from the artificial intelligence machine, (a) receiving the image data by the artificial intelligence machine; (b) generating, by the artificial intelligence machine, first labeling data indicating a defective area of the part, storing the first labeling data in conjunction with the image data, and transmitting the image data to the first labeling device; (c) uploading, by the first labeling device, the first labeling data to the artificial intelligence machine when first approval information is input to the first labeling device; (d) generating, by the artificial intelligence machine, second labeling data indicating the defective area, storing the second labeling data in association with the image data, and transmitting the second labeling data to the second labeling device; (e) uploading, by the second labeling device, the second labeling data to the artificial intelligence machine when second approval information is input to the second labeling device; And (f) the artificial intelligence machine for learning the first labeling data and the second labeling data stored in association with the image data. It is achieved by the labeling learning method.

The first labeling data may include a symbol designating the defective area; parts classification data designating the type of the parts; and defect classification data designating the type of the defect area.

The second labeling data may include mask data designating the defective area inside the coordinate data in units of pixels.

In the step (c), when first corrected labeling data is input instead of the first approval information to the first labeling device, the first labeling data is replaced with the first corrected labeling data, and then the first labeling device may be made to upload the first modified labeling data to the artificial intelligence machine.

In the step (e), when second corrected labeling data is input to the second labeling device instead of the second approval information, the second labeling data is replaced with the second corrected labeling data, and then the second labeling device may be made to upload the second modified labeling data to the artificial intelligence machine.

The first corrected labeling data may include a symbol designating the defective area; parts classification data designating the type of the parts; and defect classification data designating the type of the defect area.

The second corrected labeling data may include mask data designating the defective area in units of pixels.

If the artificial intelligence machine does not generate the first labeling data in the step (b), after the first corrected labeling data is stored in association with the image data in the first labeling device, the first labeling device performs the artificial intelligence The first corrected labeling data may be uploaded to an intelligent machine, and the step (d) may be initiated.

If the artificial intelligence machine does not generate the second labeling data in step (d), after the second corrected labeling data is stored in association with the image data in the second labeling device, the second labeling device performs the artificial intelligence The second corrected labeling data may be uploaded to an intelligent machine, and the step (f) may be initiated.

According to the present invention, by uploading image data in which the first labeling data or the first corrected labeling data is stored and then uploading the image data in which the second labeling data or the second corrected labeling data is stored to the artificial intelligence machine, a specialized labeling operation is performed. Even for production parts that require specific knowledge, it is possible to provide an artificial intelligence machine labeling learning method for smart vision inspection of production parts that can quickly build labeling data for AI learning.

1 is a flowchart of a labeling learning method of an artificial intelligence machine for smart vision inspection of production parts according to an embodiment of the present invention.
Figure 2 is a schematic diagram showing a smart vision inspection system used in the labeling learning method of the artificial intelligence machine for smart vision inspection of production parts of FIG.
Figure 3 is a flow chart showing in detail the step (b) of the labeling learning method of the artificial intelligence machine for smart vision inspection of the production part of FIG.
Figure 4 is a flow chart showing in detail the step (d) of the labeling learning method of the artificial intelligence machine for smart vision inspection of the production part of FIG.
FIG. 5(a) is a diagram showing image data of a defectively processed part of FIG. 2 .
FIG. 5( b ) is a diagram showing image data output on the screen of the first labeling device 12 of FIG. 2 .
FIG. 5(c) is a diagram showing image data output on the screen of the second labeling device 13 of FIG. 2 .
6 is a diagram showing various image data.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. However, in describing the present invention, descriptions of already known functions or configurations will be omitted to clarify the gist of the present invention.

The labeling learning method of an artificial intelligence machine for smart vision inspection of production parts of the present invention is made to quickly build labeling data for AI learning even for production parts that require specialized knowledge for labeling work.

1 is a flowchart of a labeling learning method (S100) of an artificial intelligence machine 14 for smart vision inspection of a production part 1 according to an embodiment of the present invention.

2 is a schematic diagram showing the smart vision inspection system 10 used in the labeling learning method (S100) of the artificial intelligence machine 14 for smart vision inspection of the production part 1 of FIG.

3 is a flowchart showing in detail step (b) (S120) of the labeling learning method (S100) of the artificial intelligence machine 14 for smart vision inspection of the production part 1 of FIG. 1.

4 is a flowchart showing in detail step (S140) of the labeling learning method (S100) of the artificial intelligence machine 14 for smart vision inspection of the production part 1 of FIG. 1.

Figure 5 (a) is a diagram showing the image data (1A) of the defective part (1) of FIG. FIG. 5( b ) is a diagram showing image data 1A output on the screen of the first labeling device 12 of FIG. 2 . FIG. 5(c) is a diagram showing image data 1A output on the screen of the second labeling device 13 of FIG. 2 .

6 is a diagram showing various image data 1A.

As shown in FIG. 1, the labeling learning method (S100) of the artificial intelligence machine 14 for smart vision inspection of the production part 1 according to an embodiment of the present invention is (a) steps (S110), ( b) step S120, (c) step S130, (d) step S140, (e) step S150 and (f) step S160.

As shown in FIG. 2, the labeling learning method (S100) of the artificial intelligence machine 14 for smart vision inspection of the production part 1 according to an embodiment of the present invention uses the smart vision inspection system 10. .

The smart vision inspection system 10 includes a camera 11, a first labeling device 12, an artificial intelligence machine 14, and a second labeling device 13.

The camera 11 generates image data 1A of the part 1 in the production line. The camera 11 is installed on one side of the production line. One or more cameras 11 may be provided. The camera 11 can generate image data 1A of the part 1 by automatic or manual operation.

The first labeling device 12 may be a tablet or personal computer provided in the production line. The first labeling device 12 receives image data 1A from the camera 11 . The first labeling device 12 and the camera 11 may communicate by wire or wirelessly.

The first labeling device 12 may output the image data 1A received from the camera 11 to the screen. In the first labeling device 12, an application program capable of specifying the type and location of the part 1 in the image data 1A and the type and location of the defective area 1B may be installed.

The professional worker P1 operates the first labeling device 12 . The professional worker P1 may be a professional person capable of determining the type of the part 1 and the defective area by looking at the image data 1A of the part 1.

The second labeling device 13 may be a tablet or personal computer provided remotely from the production line. The second labeling device 13 transmits and receives data with the artificial intelligence machine 14 . The second labeling device 13 and the artificial intelligence machine 14 may communicate by wire or wirelessly.

The second labeling device 13 may output the data transmitted by the artificial intelligence machine 14 to the screen. In the second labeling device 13, an application capable of performing area division in pixel units for the defective area 1B in the image data 1A may be installed.

The external worker (P2) operates the second labeling device (13). The external worker P2 may be a non-professional person who cannot determine the type of the part 1 and the defective area 1B by looking at the image data 1A of the part 1.

The artificial intelligence machine 14 can perform machine learning such as deep learning for recognizing and learning image data 1A, determining and learning types of parts 1 and defective areas 1B. can

Deep learning technology, a type of machine learning, can learn by going down to a multi-level deep level based on data. Deep learning can represent a set of machine learning algorithms that extract core data from a plurality of data as the level increases. Since deep learning technology is a widely known technology, a detailed description thereof will be omitted.

The artificial intelligence machine 14 automatically detects (localizes) the defective area 1B of the image data 1A, an artificial intelligence model for detecting the location of the defect, and segments the detected defective area 1B into pixels. and an artificial intelligence model for segmentation of defective areas for designation.

The artificial intelligence machine 14 transmits, receives, and stores data with the first labeling device 12 and the second labeling device 13 . The artificial intelligence machine 14 may include a cloud platform function. That is, data can be stored in the artificial intelligence machine 14 connected to the Internet, and data can be freely transmitted and received with the first labeling device 12 and the second labeling device 13 connected to the Internet.

As shown in FIG. 3, (a) step S110 is a step in which the artificial intelligence machine 14 receives image data 1A from the camera 11. Fig. 5(a) shows image data 1A photographed by the camera 11.

When the first labeling device 12 receives the image data 1A from the camera 11 (S110), the first labeling device 12 transmits the image data 1A to the artificial intelligence machine 14.

When step (a) (S110) is completed, step (b) (S120) is performed. (b) In step S120, the artificial intelligence machine 14 first determines whether first labeling data is generated (S121).

As shown in FIG. 5(b), the first labeling data includes a symbol D1-1 designating a defective area, parts classification data designating the type of the part 1, and designating the type of the defective area. It includes bad classification data.

Machine learning involves algorithms that learn based on empirical data, make predictions, and improve their own performance. The artificial intelligence machine 14 may improve the generation frequency and accuracy of the first labeling data as learning of the part 1 and the defective area 1B by machine learning progresses.

When the first labeling data is generated, the artificial intelligence machine 14 stores the first labeling data in conjunction with the image data 1A. The first labeling data is stored as data in a different form associated with the image data 1A. For example, the first labeling data may include coordinate data of a specific region of the image data 1A.

Then, the first labeling data is transmitted to the first labeling device 12 . At this time, the first labeling device 12 outputs the received first labeling data and the input window on the screen. The first labeling data is output on the screen of the first labeling device 12 in an overlapping manner with the image data 1A.

The professional worker P1 can check whether the first labeling data is correct through the image data 1A and the first labeling data output on the screen. If it is determined that the first labeling data is correct, the professional worker (P1) inputs the first approval information to the input window (S122).

(c) Step (S130) is a step in which the first labeling device 12 uploads the first labeling data to the artificial intelligence machine 14.

When the first approval information is input to the first labeling device 12 (S122), the first labeling device 12 uploads the first labeling data to the artificial intelligence machine 14 (S130). The first labeling data may be uploaded and stored in the cloud platform together with meta data such as generation time and part (1) ID.

As shown in FIG. 3, when the professional worker P1 determines that the first labeling data is incorrect, the first labeling device 12 may input the first corrected labeling data instead of the first approval information (S123). .

The first correction labeling data includes symbol data D3-1 designating the defective area 1B, parts classification data designating the type of the part 1, and defect classification data D3 designating the type of the defective area 1B. -3) may be included.

The professional worker P1 may designate the defective area 1B in the image data 1A output on the screen through various first labeling devices 12 including a touch screen of a tablet or a mouse of a PC. The symbol data D3-1 designating the defective area 1B may be a bounding rect or a screen drawing image.

Alternatively, the professional worker P1 directly marks the defective position of the part 1 using a pen or a sticker S2 for indicating the defective part (see FIG. 6), photographs the part 1, and performs first labeling. It is also possible to store image data 1A in device 12. S1 means that the defective position is marked with a pen.

And the professional worker (P1) can designate the type of the part (1) by selecting one of the types of parts (1) displayed on the screen of the first labeling device (12).

In addition, the professional operator (P1) can designate the type of the corresponding defective area (1B) by selecting one or multiple types of the defective area (1B) displayed on the screen of the first labeling device (12).

When the first corrected labeling data is input to the first labeling device 12 instead of the first approval information (S123), the first labeling data is replaced with the first corrected labeling data. That is, the first corrected labeling data is stored in association with the image data 1A. Then, the first labeling device 12 uploads the first corrected labeling data 1A to the artificial intelligence machine 14 (S130).

As described above, in step (b) (S120), the artificial intelligence machine 14 first determines whether the first labeling data is generated (S121). If the artificial intelligence machine 14 does not generate the first labeling data, the first labeling device 12 outputs the image data 1A and an input window on the screen. In addition, the professional worker P1 may input the first corrected labeling data to the first labeling device 12 (S123).

After the first corrected labeling data is linked to and stored in the image data 1A in the first labeling device 12, the first labeling device 12 uploads the first corrected labeling data to the artificial intelligence machine 14 (S130) And, step (d) (S140) is initiated.

When step (c) (S130) is completed, step (d) (S140) is performed. (d) In step S140, the artificial intelligence machine 14 first determines whether the second labeling data is generated (S141).

As shown in FIG. 5(c), the second labeling data includes mask data D2-1 designating the defective area 1B inside the coordinate data D1-1 in units of pixels, and the defective area 1B It includes defect classification data for recognizing the type of.

Mask data refers to data having a value representing the presence or absence of a defective region 1B for each pixel position having the same size as the picture of the component 1.

If the value of the mask data D2-1 is 0, it means that the pixel at the corresponding position in the picture of the component 1 is not included in the defective area 1B.

If the value of the mask data D2-1 is 1 (or not 0), it means that the pixel at the corresponding position in the image data 1A of the part 1 is included in the defective area 1B. In addition to this, the mask data D2-1 may include data capable of distinguishing pixels included in the defective region 1B in various ways.

Machine learning involves algorithms that learn based on empirical data, make predictions, and improve their own performance. As learning of the part 1 and the defective area 1B by machine learning progresses, the generation frequency and accuracy of the second labeling data may be improved.

When the second labeling data is generated, the artificial intelligence machine 14 stores the second labeling data in conjunction with the image data 1A. The second labeling data is stored as data in a different form associated with the image data 1A. For example, the second labeling data may include coordinate data of a specific region of the image data 1A.

Then, the image data 1A is transmitted to the second labeling device 13 . At this time, the second labeling device 13 outputs the received second labeling data and the input window on the screen. The second labeling data is overlapped with the image data 1A and output on the screen of the second labeling device 13 .

The external worker P2 can check whether the second labeling data is correct through the image data 1A and the second labeling data output on the screen. If it is determined that the second labeling data is correct, the external worker P2 inputs the second approval information into the input window (S142).

Step (e) (S150) is a step in which the second labeling device 13 uploads the second labeling data to the artificial intelligence machine 14.

When the second approval information is input to the second labeling device 13 (S142), the second labeling device 13 uploads the second labeling data to the artificial intelligence machine 14 (S150). The second labeling data may be uploaded and stored in the cloud platform together with meta data such as generation time and part (1) ID.

As shown in FIG. 4, when the second labeling data is determined to be incorrect, the external worker P2 may input second corrected labeling data instead of the second approval information to the second labeling device 13 (S143). .

The second corrected labeling data may include mask data D4-1 designating the defective area 1B in units of pixels.

As shown in FIG. 5(c), the external worker P2 has a defective area of the image data 1A output on the screen through various second labeling devices 13 including a touch screen of a tablet or a mouse of a PC. (1B) can be specified in units of pixels.

When the second corrected labeling data is input to the second labeling device 13 instead of the second approval information (S143), the second labeling data is replaced with the second corrected labeling data. That is, the second corrected labeling data is stored in association with the image data 1A. Then, the second labeling device 13 uploads the second corrected labeling data to the artificial intelligence machine 14 (S150).

As described above, in step (d) (S140), the artificial intelligence machine 14 first determines whether the second labeling data is generated (S141). If the artificial intelligence machine 14 does not generate the second labeling data, the second labeling device 13 outputs the image data 1A and the input window on the screen. And the external worker (P2) can input the second correction labeling data to the second labeling device (13) (S143).

After the second corrected labeling data is stored in association with the image data 1A in the second labeling device 13, the second labeling device 13 uploads the second corrected labeling data to the artificial intelligence machine 14 (S150) And, step (f) (S160) is initiated.

As shown in FIG. 4, (f) the artificial intelligence machine 14 learns the first labeling data and the second labeling data stored in association with the image data 1A.

As the number of times the first labeling device 12 uploads the first labeling data to the artificial intelligence machine 14 (S130) increases, the learning data used for learning the AI model for detecting defective locations increases, thereby detecting defective locations. The defect location detection performance of the AI model for AI is improved.

Therefore, as the number of times the first labeling device 12 uploads the first labeling data to the artificial intelligence machine 14 (S130) increases, the frequency of manual work in which the professional operator P1 directly designates the defective location (S123) It can be reduced, so step (b) (S120) can proceed quickly.

In addition, as the number of times the second labeling device 13 uploads the second labeling data to the artificial intelligence machine 14 (S150) increases, the learning data used for learning the AI model for dividing the defective area 1B increases. Thus, the performance of dividing the bad region 1B of the AI model for dividing the bad region 1B is improved.

Therefore, as the number of times the second labeling device 13 uploads the second labeling data to the artificial intelligence machine 14 (S150) increases, the external worker P2 sets the defective area 1B pixel by pixel to the mask data D4- 1) can be reduced in frequency of manual work designating (S143), so step (d) (S140) can proceed quickly.

According to the present invention, the quality inspector in the production line inputs and processes only essential elements such as defect location and defect type (S123), and performs pixel unit segmentation on the defect area 1B necessary for learning AI modeling By processing step S143 in two ways, there is an advantage in that labeling (S123) for defective data of the manufactured part 1, which can only be performed by a professional person such as a field quality manager in the manufacturing field, can be quickly performed.

In addition, by supporting marking on the surface of parts (1) using a pen such as an oil pen, which is a method of marking the location of defects of parts (1) actually used at manufacturing sites, and marking stickers (S2) for marking the location of defects, it is possible to help those who are not familiar with IT technology. There is an advantage in that the quality manager at the manufacturing site can quickly perform labeling (S123).

In addition, by performing the AI data labeling work for determining pass/fail of the part (1) in conjunction with the production process stage of the part (1), the time and cost of securing labeling data for learning the AI model required when introducing the vision quality inspection system for determining pass/fail of the part (1) , and has the advantage of being able to introduce the vision quality system faster than before.

In addition, according to the present invention, by uploading the first labeling data or the first corrected labeling data to the artificial intelligence machine 14 and then uploading the second labeling data or the second corrected labeling data, specialized knowledge in labeling work It is possible to provide a labeling learning method (S100) of an artificial intelligence machine (14) for smart vision inspection of a production part (1) made to quickly build labeling data for AI learning even if it is a required production part (1). .

In the foregoing, although specific embodiments of the present invention have been described and shown, the present invention is not limited to the described embodiments, and it is common knowledge in the art that various modifications and variations can be made without departing from the spirit and scope of the present invention. It is self-evident to those who have Therefore, such modifications or variations should not be individually understood from the technical spirit or viewpoint of the present invention, and modified embodiments should fall within the scope of the claims of the present invention.

S100: Labeling learning method
S110: Step (a)
S120: Step (b)
S130: Step (c)
S140: Step (d)
S150: (e) step
S160: Step (f)
10: vision inspection system
11 : Camera
12: first labeling device
13: second labeling device
14: AI machine
1: Parts
1A: image data
1B: bad area
S2: Sticker
D1-1: coordinate data
D2-1: mask data
D3-1: symbol data
D3-3: Defect Classification Data
D4-1: mask data
P1: Field workers
P2: External worker

Claims (9)

As a labeling learning method for an artificial intelligence machine using a smart vision inspection system,
The smart vision inspection system,
A camera that generates image data of parts in a production line;
A first labeling device provided in the production line to receive the image data and having an application program installed therein which allows a professional worker to designate the type of the part and the type and location of the defective area of the part from the image data;
An artificial intelligence model for detecting a defective location for transmitting and receiving data with the first labeling device and automatically detecting (localization) the defective area of the image data, and segmenting and designating the detected defective area in pixel units An artificial intelligence machine including an artificial intelligence model for segmentation of defective regions to perform; and
A second labeling device provided with an application program installed remotely from the production line to transmit and receive data to and from the artificial intelligence machine, and allowing an external worker to divide the defective area in pixel units in the image data; ,
(a) receiving the image data by the artificial intelligence machine;
(b) generating, by the artificial intelligence machine, first labeling data indicating a defective area of the part, storing the first labeling data in conjunction with the image data, and transmitting the first labeling data to the first labeling device;
(c) uploading, by the first labeling device, the first labeling data to the artificial intelligence machine when the first approval information by the professional operator is input to the first labeling device;
(d) generating, by the artificial intelligence machine, second labeling data indicating the defective area, storing the second labeling data in association with the image data, and transmitting the second labeling data to the second labeling device;
(e) uploading, by the second labeling device, the second labeling data to the artificial intelligence machine when the second approval information by the external worker is input to the second labeling device; and
(f) learning, by the artificial intelligence machine, the first labeling data and the second labeling data stored in association with the image data;
In step (e),
When second corrected labeling data by the external worker is input to the second labeling device instead of the second approval information, after the second labeling data is replaced with the second corrected labeling data, the second labeling device A labeling learning method of an artificial intelligence machine for smart vision inspection of production parts, characterized in that for uploading the second corrected labeling data to the artificial intelligence machine. According to claim 1,
The first labeling data,
a symbol designating the defective area; and
parts classification data designating the type of the parts; and
A labeling learning method of an artificial intelligence machine for smart vision inspection of production parts, characterized in that it includes defect classification data specifying the type of the defect area. According to claim 2,
The first labeling data includes coordinate data of a specific region of the image data,
The second labeling data,
A labeling learning method of an artificial intelligence machine for smart vision inspection of production parts, characterized in that it includes mask data designating the defective area inside the coordinate data in pixel units. According to claim 1,
In step (c),
When first corrected labeling data is input instead of the first approval information to the first labeling device, after the first labeling data is replaced with the first corrected labeling data, the first labeling device sends the artificial intelligence machine to the first corrected labeling data. A labeling learning method of an artificial intelligence machine for smart vision inspection of production parts, characterized in that for uploading the first corrected labeling data. delete According to claim 4,
The first modified labeling data,
a symbol designating the defective area; and
parts classification data designating the type of the parts; and
A labeling learning method of an artificial intelligence machine for smart vision inspection of production parts, characterized in that it includes defect classification data specifying the type of the defect area. According to claim 1,
The second modified labeling data,
A labeling learning method of an artificial intelligence machine for smart vision inspection of production parts, characterized in that it includes mask data designating the defective area in pixel units. According to claim 1,
If the artificial intelligence machine does not generate the first labeling data in the step (b), after the first corrected labeling data is stored in association with the image data in the first labeling device, the first labeling device performs the artificial intelligence A labeling learning method of an artificial intelligence machine for smart vision inspection of production parts, characterized in that the first corrected labeling data is uploaded to the intelligent machine, and the step (d) is initiated. According to claim 8,
If the artificial intelligence machine does not generate the second labeling data in step (d), after the second corrected labeling data is stored in association with the image data in the second labeling device, the second labeling device performs the artificial intelligence A labeling learning method of an artificial intelligence machine for smart vision inspection of production parts, characterized in that the second corrected labeling data is uploaded to the intelligent machine, and the step (f) is initiated.

Images