KR102230559B1 - Method and Apparatus for Creating Labeling Model with Data Programming - Google Patents

Abstract

Disclosed is a method and apparatus for generating a labeling model based on data programming.
According to the present embodiment, a pre-processing process is performed on image information of a part, and a labeling function is prototyped based on a general framework that appropriately combines various existing image processing techniques. Provides a labeling model generation method and apparatus.

Description

Method and Apparatus for Creating Labeling Model with Data Programming}

The present embodiment relates to a method and apparatus for generating a labeling model based on data programming.

The content described below merely provides background information related to the present embodiment, and does not constitute the prior art.

In general, in order to perform supervised learning using machine learning, labeled data is required. In addition, as feature engineering has developed based on deep learning in the field of machine learning, securing a lot of labeled data has emerged as a more important task.

Existing data labeling techniques include self learning and active learning. Self-learning is one of the semi-supervised learning techniques. It is a technique that predicts the label of the remaining unlabeled data by assuming that a part of the data is labeled and using the labeled data as much as possible. This technique cannot be used if there is no label because it assumes that there is already enough labeled data. Active learning is a technique that minimizes human labeling, and does not label all data, but selectively labels only the data that is most helpful in improving model accuracy. The problem is that this technique cannot be used in new applications where such personnel are difficult to obtain, as it assumes that there are enough domain experts to do the labeling perfectly.

In recent years, data programming that can generate labels in large quantities has been proposed and is being used in more and more places. Not only domain experts but also non-experts develop a number of labeling functions (LF) that can automatically generate labels, and the opinions of those functions are collected into a single generative model. When the output values of the labeling functions pass through the generative model, the weak label of each unlabeled data is determined. Here, a weak label refers to a label that is not as accurate as a label manually generated by a human, but instead is automatically generated and has an accuracy of more than a certain threshold. Learning through data programming ends with training a discriminative model such as a convolutional neural network (CNN) using weak labels obtained through the above-described process. Data programming does not require conventional labels compared to self-learning, and does not require manual labeling compared to active learning. Implementing a labeling function in data programming can generate a large amount of weak labels. The current method has a limitation that the labeling function is mainly used only for relatively simple text data, assuming that there is a professional manpower capable of implementing the labeling function.

In this embodiment, a labeling function prototyping a labeling function based on a general framework that processes an image by performing a pre-processing process on image information of a part, and appropriately combines various existing image processing techniques. An object of the present invention is to provide a method and apparatus for generating a model.

In addition, in the present embodiment, when a labeling user interface (UI) is present, defects visible in an image are easily displayed using a labeling UI based on a crowd-sourcing technique, and patterns of the displayed defects are converted to each labeling function. An object of the present invention is to provide a method and apparatus for generating a labeling model that automatically converts.

According to an aspect of the present embodiment, there is provided an image collection unit for collecting image information of parts from an image photographing apparatus; A data preprocessor for generating processed image information by performing a preprocessing technique on the part image information; A labeling function generator that determines whether a part is defective by applying an image processing technique or a crowdsourcing technique to the processed image information, and generates at least one labeling function according to a determination result; And a model generation unit generating a labeling model for obtaining a weak label by combining the labeling function.

According to another aspect of the present embodiment, there is provided an image collection process of collecting image information of parts from an image photographing apparatus; A data pre-processing process of generating processed image information by performing a pre-processing technique on the part image information; A labeling function generation process of determining whether a part is defective by applying an image processing technique or a crowdsourcing technique to the processed image information, and generating at least one labeling function according to a determination result; And a model generation process of generating a labeling model for obtaining a weak label by combining the labeling function.

As described above, according to the present embodiment, a labeling function is prototyped based on a general framework that processes an image by performing a pre-processing process on image information of a part, and appropriately combines various existing image processing techniques. There is an effect of providing a method and an apparatus for generating a labeling model for prototyping.

In addition, according to the present embodiment, when a labeling user interface (UI) is present, defects visible in an image are easily displayed using a labeling UI based on a crowd-sourcing technique, and the displayed defect patterns are labeled for each. There is an effect of providing a method and apparatus for generating a labeling model that automatically converts to a function.

In addition, according to the present embodiment, it is possible to reduce the cost or time required for labeling by data programming including an image processing technique or a crowdsourcing technique.

1 is a conceptual diagram schematically illustrating a labeling model generation system according to the present embodiment.
2 is a block diagram schematically showing the structure of an apparatus for generating a labeling model according to the present embodiment.
3 is a block diagram illustrating a driving process of an apparatus for generating a labeling model according to the present embodiment.
4 is a conceptual diagram schematically illustrating a labeling UI according to the present embodiment.
5 is a flowchart illustrating a method of generating a labeling model according to the present embodiment.
6 is a flowchart illustrating a method of applying an image processing technique according to the present embodiment.
7 is a flowchart illustrating a method of applying a crowdsourcing technique according to the present embodiment.

Hereinafter, the present embodiment will be described in detail with reference to the accompanying drawings. In describing the embodiments, when it is determined that a detailed description of a related known configuration or function may obscure the subject matter of the present invention, a detailed description thereof will be omitted. In addition, the accompanying drawings are for easy understanding of the embodiments disclosed in the present specification, and the technical idea disclosed in the present specification is not limited by the accompanying drawings, and all changes included in the spirit and scope of the present invention , It should be understood to include equivalents or substitutes.

In this embodiment, the term'comprising' means that elements, features, steps, or combinations thereof described in the specification exist, and the possibility of existence of one or a plurality of elements, other features, steps, or combinations thereof is excluded in advance. It should be understood as not doing. In addition, when a component is described as being "connected", "coupled" or "connected" to another component, the component may be directly connected or connected to the other component, but between each component It should be understood that other components may be “connected”, “coupled” or “connected”.

1 is a conceptual diagram schematically illustrating a labeling model generation system according to the present embodiment.

Referring to FIG. 1, the labeling model generation system according to the present embodiment includes a target part 110, a labeling model generation device 120, and a user terminal 130. Components included in the labeling model generation system are not necessarily limited thereto.

The target part 110 is photographed by an image photographing device. The target component 110 includes various components provided in various mechanical equipment in the high-tech manufacturing industry. Meanwhile, the image photographing apparatus photographs the target part 110 based on a uniform illuminance, location, and photographing distance, and generates part image information. The image capturing device refers to a device including an optical module such as a camera or lighting for acquiring part image information.

The labeling model generation device 120 collects part image information from the image photographing device. Here, the part image information means image information of the target part 110 photographed by the image photographing apparatus. The labeling model generation device 120 generally collects part image information from an image photographing device, but receives part image information by a user's manipulation or command, or loads part image information stored in a separate database. )You may.

The labeling model generation device 120 generates processed image information by processing the collected part image information. Here, the processed image information is an image obtained by applying various pre-processing techniques to part image information, and a pre-processing technique will be described later with reference to FIG. 2.

The labeling model generation device 120 generates a labeling function based on the processed image information. In more detail, the labeling model generation apparatus 120 generates a labeling function by applying either an image processing technique or a crowdsourcing technique to processed image information according to the user's selection information. Here, the user's selection information means selecting a technique that is advantageous to the user according to the user's situation.

For example, in a situation where it is difficult to secure a quality label by a domain expert, it is advantageous to select an image processing technique when an expert in the field of image processing exists, and a crowdsourcing technique when there is no expert. Select either an image processing technique or a crowdsourcing technique according to the situation of

The labeling model generation device 120 generates a generative model or a voting model by combining the generated labeling functions. The generation model or the voting model will be described later with reference to FIG. 2.

The labeling model generation device 120 generates a weak label by using a generation model or a voting model. Here, a weak label refers to a label that is not as accurate as a label manually generated by a user, but is automatically generated and has an accuracy of more than a certain threshold. Therefore, users can obtain efficiency in terms of cost and time by using a weak label. The labeling model generation device 120 transmits the generated weak label to the user terminal 130.

The user terminal 130 may train a discriminative model using a weak label. Here, the classification model is preferably learned using a convolutional neural network (CNN), but is not limited thereto.

The user terminal 130 refers to an electronic device capable of receiving various web page data via a network according to a user's key manipulation. The user terminal 130 includes a tablet PC, a laptop, a personal computer, a smart phone, a personal digital assistant (PDA), and a mobile communication terminal. Mobile Communication Terminal) and the like. The user terminal 130 may include a web browser for accessing the labeling model generating device 120 via a network, a memory for storing a program, and a microprocessor for executing and controlling a program.

2 is a block diagram showing the structure of an apparatus for generating a labeling model according to the present embodiment.

Referring to FIG. 2, the labeling model generation apparatus 120 according to the present embodiment includes an image collection unit 210, a data preprocessor 220, a labeling function generation unit 230, and a model generation unit 240. . Components of the labeling model generation apparatus 120 are according to an embodiment, and are not essential components to reproduce the present embodiment, and some components may be added, changed, or deleted.

The image collection unit 210 collects part image information from the image photographing device. More specifically, the image collection unit 210, where the part image information is image information obtained by photographing various target parts 110, from a separately provided database (not shown) or an external cloud server (not shown). It may be loaded or may be input by user's input information. The image collection unit 210 may collect part image information and store it as image data, and the labeling model generation device 120 may use the image data to determine whether the target part 110 is defective.

The data preprocessor 220 generates processed image information by performing a preprocessing technique on the part image information. Here, the processed image information means image information from which noise on the part image information is removed so that it can be easily used for data programming. The data preprocessor 220 generates processed image information using a light reflection correction method, an ROI setting method, or a contrast enhancement method.

In the case where the difference in pixel values over the entire area of the part image information is wider than the difference in pixel values due to defects of the target part 110 in the part image information, all part images The light reflection is corrected by subtracting the average pixel value for the information.

The region of interest setting method refers to a method of defining an area occupied by the target part 110 in the part image information as a region of interest (RoI), and filling an area outside the region of interest with a constant value. Since the ROI setting method assumes that the position of the target component 110 is uniform on the component image information, the ROI (hard-coded RoI) is used without detecting the ROI using a special algorithm.

The contrast enhancement method uses min-max normalization to enhance the contrast when the contrast between the defective part and the non-defect part is small and it is difficult to identify a small defect area with the naked eye. do.

The labeling function generator 230 determines whether a part is defective by applying an image processing technique or a crowdsourcing technique to the processed image information, and generates at least one labeling function according to the determination result. Here, the labeling function (LF) refers to a function that automatically checks whether or not there is a defect by using processed image information as an input value and whether or not a surface defect of the target part 110 is an output value.

In more detail, since it is advantageous for the user to secure a label, it is advantageous for the user to select an image processing technique when an expert in the field of image processing exists, and a crowdsourcing technique when an expert does not exist. Selects an image processing technique or a crowdsourcing technique based on the selection information selected according to the user's situation. The labeling function generation unit 230 includes an image processing unit 232 that applies an image processing technique to processed image information and a crowd sourcing unit 234 that applies a crowdsourcing technique.

The labeling function generator 230 according to the present embodiment may select an image processing technique or a crowdsourcing technique according to the presence or absence of a labeling user interface (UI). In addition, the labeling function generator 230 does not necessarily select one of an image processing technique or a crowdsourcing technique, but may select both techniques to generate a labeling function.

The image processing unit 232 classifies a component defect candidate group by filtering processed image information based on an image processing technique, and generates a labeling function using only the final defect image information that satisfies a certain criterion among the component defect candidate groups. The image processing unit 232 will be described later with reference to FIG. 3.

The crowd sourcing unit 234 displays a suspected part defect area on the processed image information based on user input information using a labeling UI, and generates defect pattern information based on the suspected part defect area, and defines it as a labeling function. When the user specifies a region in the processed image information that is suspected of having a defect in the target part 110, the crowd sourcing unit 234 specifies the region as a suspected part defect region. The crowd sourcing unit 234 generates defect pattern information based on the line displayed in the suspected part defect area and defines it as a labeling function. The labeling UI will be described later with reference to FIG. 5.

The crowd sourcing unit 234 may generate defect pattern information using a plurality of suspected part defect areas, and define an area corresponding to the defect pattern information in the processed image information as a labeling function. For example, the crowd sourcing unit 234 generates a plurality of defect pattern information based on a part defect suspicious area generated by a plurality of users, and uses the generated defect pattern information to input parts on the processed image information in the next sequence. You can create a defect suspicious area.

The crowdsourcing unit 234 compares the defect pattern information with the suspected part defect area on the next processed image information using the correlation coefficient, squared difference, and cosine similarity to determine the degree of similarity. Can be measured. The crowd sourcing unit 234 may determine that there is a defect in the corresponding part when the measured similarity is greater than or equal to a specific threshold.

The crowd sourcing unit 234 selects representative defect image information including representative defect pattern information among the processed image information displaying the suspected part defect region, and automatically generates the labeling function using the representative defect image information. do. For example, the expert selects representative defect pattern information determined to be capable of representing the representativeness of the corresponding defect among the previously generated defect pattern information. The crowd sourcing unit 234 specifies processed image information in which representative defect pattern information exists among processed image information as representative defect image information, and generates a labeling function based on the representative defect image information. Accordingly, since the crowd sourcing unit 234 generates a labeling function based on a representative defect pattern, the accuracy of the labeling model generation device 120 may be improved. For example, the crowd sourcing unit 234 may obtain high-accuracy labels in a short time by selecting k labeling functions defined by the defect pattern information in the order of specific criteria (precision, recall, etc.).

The model generator 240 generates a labeling model for obtaining a weak label by combining a labeling function. Here, the labeling model includes a generative model or a voting model.

The model generation unit 240 trains a probabilistic model by combining a labeling function based on a development dataset according to a user's selection to generate a generated model. Since the performance of the generated model is determined not by each labeling function, but by a combination of labeling functions, it is desirable to secure a large amount of development data sets to detect combinations of labeling functions that complement each other.

The model generation unit 240 generates a voting model by obtaining a voting value for whether a component is defective based on the result value of each of the labeling functions. The model generator 240 obtains a vote value of a defect if it is determined that any one of a plurality of labeling functions for the processed image information is defective.

The model generator 240 may determine the type of model to be used according to the number of labeling functions. For example, it is preferable to use a generation model when the number of labeling functions is greater than or equal to a specific threshold value, and a voting model when the number of labeling functions is less than or equal to a specific threshold value. The voting model has a recall rate of 93.54%, a precision of 83.03%, and an accuracy of 88.88% when there are 7 labeling functions, and a generative model has a recall of 91.75%, a precision of 81.84%, and an accuracy of 87.57% when the labeling function is 30.

When the labeling model generation device 120 generates a labeling function, the generated labeling function may be used to determine other defects, and may additionally implement information on other defects to the labeling function. For example, when the labeling function means a defect pattern related to a scratch, the labeling model generation apparatus 120 may additionally implement a labeling function so that other types of defects other than the scratch cannot be detected.

The labeling model generation device 120 may generate a labeling model by mounting a labeling model generation program. The labeling model generation device 120 generates and provides a labeling model to a user by driving a labeling model generation program according to a user's manipulation or command.

To describe the labeling model generation program in more detail, the labeling model generation program may be a program installed in the computer when the labeling model generation device 120 is a computer. The labeling model generation program may be stored in a computer-readable recording medium and utilized through the user terminal and the labeling model generation device 120 to provide a labeling model generation service. Here, the computer-readable recording medium includes all types of recording devices that store data that can be read by a computer system. That is, as a computer-readable recording medium, a magnetic storage medium (e.g., ROM, floppy disk, hard disk, etc.), optical reading medium (e.g., CD-ROM, DVD, etc.), and carrier wave (e.g., Internet It includes a storage medium such as (transmitted through). In addition, a computer-readable recording medium may be distributed to computer servers and systems connected through a network, and stored as code that can be read by the labeling model generating apparatus 120 and executed.

On the other hand, to describe the form in which the labeling model generation program is mounted in the labeling model generation device 120 according to the present embodiment, the labeling model generation device 120 mounts the labeling model generation program in an embedded form, or It may be mounted in an embedded form in an OS (Operating System) mounted in the labeling model generating device 120, or installed in the OS in the labeling model generating device 120 by a user's manipulation or command.

3 is a block diagram illustrating a driving process of an apparatus for generating a labeling model according to the present embodiment.

Referring to FIG. 3, the labeling model generation apparatus 120 according to the present embodiment acquires part image information from the part photographing apparatus using the image collection unit 210. The labeling model generation device 120 generates processed image information using the data preprocessor 220. The labeling model generation device 120 applies an image processing technique or a crowdsourcing technique to the processed image information according to a user's selection.

The image processing unit 232 using the image processing technique includes a filtering unit 232_1, a candidate group classifying unit 232_2, and a defect determining unit 232_3. Components included in the image processing unit 232 are not limited thereto, and some components may be added, deleted, or changed.

The filtering unit 232_1 extracts feature information for checking whether a part is defective in the processed image information based on an edge detector. Here, the feature information refers to a pixel value changed in response to a defect such as a scratch on the surface of the target part 110 in the processed image information and line information detected in the processed image information based on this. In more detail, the filtering unit 232_1 is an edge detector, and image processing algorithms such as Wavelet Transform, Canny Edge Detection, Adaptive Threshold, Morphology Transformation, etc. You can use

The candidate group classifying unit 232_2 classifies the component defect candidate group based on the feature information. The candidate group classifier 232_2 may classify a component defect candidate group by applying a line segment detection, a Hough Line Transform, or the like to the extracted feature information, but is not limited thereto.

The defect determination unit 232_3 finally determines whether or not the target component 110 is defective based on a threshold in each image of the component defect candidate group. More specifically, the defect determination unit 232_3 defines pixel information corresponding to the broken area of the line included in the feature information as a parameter, and compares the length of the line and the threshold value to Determine whether there is a defect. Here, the parameter includes information on how many pixels are allowed to be cut off in the middle in defining the continuity of the line corresponding to the scratch defect. That is, the defect determination unit 232_3 may measure the length of the line based on the parameter, and may finally determine the presence or absence of a defect in the target part 110 by comparing the length of the line with a specific threshold value.

The labeling model generation device 120 generates at least one labeling function using a crowdsourcing technique. The labeling model generation device 120 generates a generation model or a voting model by combining the generated labeling function. The labeling model generation device 120 acquires a weak label by using the generation model or the voting model, and the user finally generates a classification model based on the weak label.

4 is a conceptual diagram schematically illustrating a labeling UI of an apparatus for generating a labeling model according to the present embodiment.

The labeling model generation device 120 may display a suspected part defect area in the processed image information as well as a general user (hereinafter referred to as a non-expert) who is not a domain expert. Accordingly, the labeling UI according to the present embodiment can easily display a part defect suspicious area even by a non-expert, thereby enabling a sufficient number of users to intervene, and thus a large amount of labeling functions can be obtained.

Referring to FIG. 4, the labeling UI includes an image display section and a button. The image display section is a place in which the suspected part defect area is displayed in the processed image information, and can be displayed in various ways by touch input or coordinate input. A suspected part defect area is displayed, defect pattern information is generated based on this, and a labeling function corresponding to the defect pattern information is generated.

5 is a flowchart illustrating a method of generating a labeling model according to the present embodiment.

5, the labeling model generation apparatus 120 according to the present embodiment collects part image information from the image photographing apparatus (S602). In step S602, the labeling model generating device 120 may directly receive part image information from the user.

The labeling model generation device 120 generates processed image information by performing a pre-processing technique on the part image information (S604). In step S604, the labeling model generating apparatus 120 uses a light reflection correction method, an ROI setting method, or a contrast enhancement method as a preprocessing technique, but is not limited thereto.

The labeling model generation device 120 checks whether the labeling UI exists (S606). When there is no labeling UI, the labeling model generation device 120 generates a labeling function by applying an image processing technique to the processed image information (S608). A method of applying the image processing technique to the processed image information in step S608 will be described later with reference to FIG. 7.

When the labeling UI exists, the labeling model generation device 120 generates a labeling function by applying a crowdsourcing technique to the processed image information (S610). In step S610, the labeling model generation apparatus 120 may generate a labeling function using the image processing technique if the user selects an image processing technique even when the labeling UI exists.

The labeling model generation device 120 generates a labeling model by combining the generated labeling function (S612). In step S612, the labeling model includes a generation model or a voting model.

The labeling model generating device 120 generates a weak label based on the labeling model (S614). The weak label is finally used to generate a classification model (S616).

6 is a flowchart illustrating a method of applying an image processing technique according to the present embodiment.

Referring to FIG. 6, the labeling model generating apparatus 120 extracts feature information from processed image information based on an edge detector (S702). In step S702, the feature information refers to information including a pixel value changed in response to a defect such as a scratch in the processed image information and a line detected in the processed image information based on the changed pixel value.

The labeling model generation device 120 classifies a component defect candidate group using the feature information (S704). The labeling model generation device 120 checks whether there is a broken area of the line included in the feature information (S706). The labeling model generation device 120 defines pixel information corresponding to the broken area of the line as a parameter (S708). In step S708, the parameter means information for defining the continuity of the line corresponding to the scratch defect.

The labeling model generation device 120 measures the length of the line based on the parameter, and compares the length of the line with a specific threshold (S710).

The labeling model generation apparatus 120 determines whether or not there is a component defect in each image of the component defect candidate group according to the comparison result (S712). In step S712, the labeling model generating apparatus 120 generates a labeling function based on the defect image information determined to have a defect.

7 is a flowchart illustrating a method of applying a crowdsourcing technique according to the present embodiment.

Referring to FIG. 7, the labeling model generation apparatus 120 according to the present embodiment displays a suspected part defect region on processed image information based on user input information using a labeling UI (S802). In step S802, the user's input information includes drag input using a mouse and coordinate input using a keyboard.

The labeling model generation device 120 generates defect pattern information based on a plurality of suspected part defect areas (S804). In step S804, the suspected part defect area refers to a specific area indicated by the user as a suspected defect area, and the defect pattern information refers to a defect pattern existing in the suspected part defect area, eg, a defect pattern such as a scratch.

The labeling model generation device 120 checks whether the user has selected representative defect pattern information (S806). When the user selects representative defect pattern information, the labeling model generation device 120 automatically generates a labeling function using representative defect image information including representative defect pattern information (S808).

When the user does not select representative defect pattern information, the labeling model generation apparatus 120 defines a region corresponding to the defect pattern information as a labeling function (S810).

In FIGS. 5 to 7, steps S602 to S616, S702 to S712, and S802 to S810 are described as sequentially executing, but are not limited thereto. In other words, since it is possible to change and execute the steps illustrated in FIGS. 5 to 7 or to execute one or more steps in parallel, FIGS. 5 to 7 are not limited to a time series order.

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 embodiments are not intended to limit the technical idea of the present embodiment, but to explain the technical idea, and the scope of the technical idea of the present embodiment is not limited by these 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: target part 120: labeling model generation device
130: user terminal 210: image collection unit
220: data preprocessing unit 230: labeling function generation unit
232: image processing unit 232_1: filtering unit
232_2: Candidate classification unit 232_3: Defect judgment unit
234: crowd sourcing unit 240: model generation unit

Claims (11)

An image collection unit for collecting image information of parts from the image photographing apparatus;
A data preprocessor for generating processed image information by performing a preprocessing technique on the part image information;
A labeling function generator that determines whether a part is defective by applying an image processing technique or a crowdsourcing technique to the processed image information, and generates at least one labeling function according to a determination result; And
Including a model generation unit for generating a labeling model (Labeling Model) for obtaining a weak label (Weak Label) by combining the labeling function,
The labeling function generation unit,
An image processing unit for classifying a component defect candidate group by filtering the processed image information based on the image processing technique, and generating the labeling function using only defect image information satisfying a predetermined criterion among the component defect candidate groups,
The image processing unit,
Based on an edge detector, feature information for checking whether a part is defective in the processed image information is extracted, the part defect candidate group is classified using the feature information, and a line included in the feature information Defining pixel information corresponding to the broken area of (Line) as a parameter, and determining the length of the line and whether the part is defective based on the parameter for each image of the part defect candidate group Comparing threshold values for determining whether the part is defective, generating defective image information according to the determination result, and generating the labeling function using the defective image information
Labeling model generation device, characterized in that. The method of claim 1,
The labeling function generation unit,
A crowd that displays a suspected part defect area on the processed image information based on user input information using a labeling user interface (UI), generates defect pattern information based on the suspected part defect area, and defines it as the labeling function Sourcing Department
Labeling model generation device, characterized in that it further comprises. The method of claim 1,
The data preprocessing unit,
As the pre-processing technique, a method for generating a labeling model, wherein a method of correcting light reflection, a method of setting an ROI, or a method of enhancing contrast. The method of claim 1,
The model generation unit,
Based on the Development Dataset, a probabilistic model is trained by combining the labeling functions to generate a generative model, or a voting value corresponding to the result of each of the labeling functions on whether a part is defective or not. Labeling model generation device, characterized in that for generating a voting model based on. delete The method of claim 2,
The crowd sourcing unit,
An apparatus for generating a labeling model, comprising generating at least one defect pattern information based on a plurality of suspected part defect regions, and defining a region corresponding to the defect pattern information in the processed image information as the labeling function. . The method of claim 2,
The crowd sourcing unit,
A labeling model, characterized in that the representative defect image information including representative defect pattern information is selected among the processed image information displaying the part defect suspicious region, and the labeling function is automatically generated using the representative defect image information. Generating device. delete In the method for the labeling model generation device to generate the labeling model,
An image collection process of collecting image information of parts from an image photographing device;
A data pre-processing process of generating processed image information by performing a pre-processing technique on the part image information;
A labeling function generation process of determining whether a part is defective by applying an image processing technique or a crowdsourcing technique to the processed image information, and generating at least one labeling function according to a determination result; And
Including a model generation process of generating a labeling model for obtaining a weak label by combining the labeling function,
The labeling function generation process,
An image processing process of classifying a component defect candidate group by filtering the processed image information based on the image processing technique, and generating the labeling function using only defect image information satisfying a certain criterion among the component defect candidate groups,
The image processing process,
Based on an edge detector, feature information for checking whether a part is defective in the processed image information is extracted,
Classify the component defect candidate group using the feature information,
Pixel information corresponding to the broken area of the line included in the characteristic information is defined as a parameter,
For each image of the component defect candidate group, the length of the line measured based on the parameter and a threshold for determining whether the component is defective is compared to determine whether the component is defective, and according to the determination result Generating defective image information, and generating the labeling function using the defective image information
Labeling model generation method, characterized in that. The method of claim 9,
The labeling function generation process,
A crowd that displays a suspected part defect area on the processed image information based on user input information using a labeling user interface (UI), generates defect pattern information based on the suspected part defect area, and defines it as the labeling function Sourcing process
Labeling model generation method, characterized in that it further comprises. A computer program stored in a computer-readable recording medium to execute each process included in the labeling model generation method according to claim 9 or 10.

Images