KR102658479B1 - Quality inspection system of unstructured data and mycelium leather and method performing thereof - Google Patents

Abstract

The quality inspection method of unstructured data and mycelial leather according to an embodiment of the present invention includes the steps of an analysis object recognition device recognizing an analysis object, generating and providing a test specimen image for the analysis object, and a quality inspection server recognizing the analysis object. When the device receives a test object image, preprocessing the test object image, the quality inspection server performs preprocessing on the test object image, and then extracting features from the preprocessed test object image, It includes a step of the quality inspection server learning a defective judgment model using the characteristics, and a step of the quality inspection server determining whether the inspection object image is defective through the defective judgment model.

Description

Quality inspection server and method of executing the same {QUALITY INSPECTION SYSTEM OF UNSTRUCTURED DATA AND MYCELIUM LEATHER AND METHOD PERFORMING THEREOF}

The present invention relates to a quality inspection server and a method of executing the same. More specifically, unlike predefined defect patterns, it is possible to detect defect patterns that actually occur and classify them into specific groups so that direct action can be taken for efficient process management. It relates to a quality inspection server for unstructured data and mycelial leather and a method of executing the same.

Recently, as consumer demands for products have increased and information and communication media such as the Internet and broadcasting have developed, the importance of quality control of products sold and distributed is emerging.

In particular, if a defect occurs in a product or its parts in home appliances such as TVs, DVDs, refrigerators, machine tools such as presses and milling, electronic devices such as personal computers and servers, and automobiles, a mass recall of already distributed products is required. ), in cases where quality control of the above-mentioned product or parts of the above-mentioned product fails, a huge financial loss is often incurred to the manufacturer or distributor.

Therefore, in order to measure and improve quality during the manufacturing process, the need for a quality management system to automatically collect quality information in real time and analyze statistically to continuously improve quality was requested.

Accordingly, in the relevant technical field, there is a need for the development of a flagpole to improve the quality of observation data and secure consistency by applying consistent quality standards to observation data collected at observation points.

Korean Patent Publication No. 10-2005-7020086 relates to an automated quality inspection system. More specifically, the process of organizing, monitoring, and accounting inspection of a project for producing a product or service is performed at each stage to produce a product or service. It includes providing a project description as an interconnection part, and providing custom databases for storing and accessing documents related to each step of the description part.

Korean Patent Publication No. 10-2012-0018034 relates to a method of automating image quality inspection for digital cinema. More specifically, it can automatically detect image quality deterioration and encoding errors that may occur during color space conversion and image compression. It's about how to be.

Korean Patent No. 10-2118487 relates to an automatic quality inspection system and method for container-based applications for continuous integration and distribution. More specifically, it relates to a system and method that can automatically inspect or verify the quality of container-based applications. The purpose is to provide continuous integration and distribution, and to ensure that the above systems and methods can also be used when providing microservices.

Korean Patent No. 10-1444245 relates to an automobile welding quality inspection device and an automobile welding quality inspection device using the same. More specifically, it secures and stores welding data performed on multiple welding devices in real time, and also secures and stores each welding data in real time. It is disclosed that it is possible to determine whether welding defects are performed in the device and to secure the history.

As described above, the principle of the existing quality checker calculates the position, shape, and color difference of the pixel unit between the vision-centered, predefined image and the image to be inspected. It is a method used in various industries with the advantage of being able to quickly and accurately obtain results regarding compliance with quality standards.

However, this dictionary definition lacks flexibility and has the problem that it is not suitable for inspection of complex shapes or shapes of fabrics, leather materials, and organic materials that change the shape of the inspection object.

Korean Patent Publication No. 10-2005-7020086 Korean Patent Publication No. 10-2012-0018034 Korean Patent No. 10-2118487 Korean Patent No. 10-1444245

The purpose of the present invention is to provide a quality inspection server and execution method that can recognize and deal with various quality errors and easily apply new error types to conduct inspections corresponding to areas such as leather materials, organic materials, and textiles. do.

In addition, the present invention provides a quality inspection server and its execution method that, unlike predefined defective patterns, can detect actually occurring defective patterns and classify them into specific groups, allowing efficient and direct action to be taken for process management. The purpose.

In addition, the purpose of the present invention is to provide a quality inspection server and a method of executing the same that can learn new types of defect patterns that occur during mass production and continuously update quality standards.

The objects of the present invention are not limited to the objects mentioned above, and other objects and advantages of the present invention that are not mentioned can be understood by the following description and will be more clearly understood by the examples of the present invention. Additionally, it will be readily apparent that the objects and advantages of the present invention can be realized by the means and combinations thereof indicated in the patent claims.

To achieve this purpose, the quality inspection system for unstructured data and mycelial leather is an analysis object recognition device that recognizes the analysis object and generates and provides an image of the specimen for the analysis object. When the analysis target recognition device receives a test object image, preprocessing is performed on the test object image, features are extracted from the preprocessed test object image, a defect determination model is trained using the features, and the defect It includes a quality inspection server that determines whether the inspection object image is defective through a judgment model.

In addition, the quality inspection method of unstructured data and mycelial leather to achieve this purpose includes the steps of an analysis object recognition device recognizing the analysis object, generating and providing a test specimen image for the analysis object, and a quality inspection server using the analysis object recognition device. When receiving a test object image, preprocessing the test object image, performing preprocessing on the test object image by the quality inspection server, extracting features from the preprocessed test object image, and preprocessing the test object image. It includes the step of the inspection server learning a defective judgment model using the characteristics, and the quality inspection server determining whether the inspection object image is defective through the defective judgment model.

According to the present invention as described above, there is an advantage in that various quality errors can be recognized and dealt with, and new error types can be easily applied to conduct inspections corresponding to areas such as leather materials, organic materials, and textiles.

In addition, according to the present invention, unlike predefined defect patterns, it is possible to detect actual defect patterns and classify them into specific groups, so there is an advantage that direct action can be taken for efficient process management.

Additionally, the present invention has the advantage of being able to learn new types of defect patterns that occur during mass production and continuously update quality standards.

Figures 1 and 2 are diagrams for explaining a quality inspection system for unstructured data and mycelial leather according to an embodiment of the present invention.
Figure 3 is a block diagram for explaining the internal structure of a quality inspection server according to an embodiment of the present invention.
Figure 4 is a flowchart illustrating an embodiment of the method for inspecting the quality of unstructured data and mycelial leather according to the present invention.
Figures 5 to 7 are exemplary views for explaining the quality inspection process of unstructured data and mycelial leather according to an embodiment of the present invention.

The above-described objects, features, and advantages will be described in detail later with reference to the attached drawings, so that those skilled in the art will be able to easily implement the technical idea of the present invention. In describing the present invention, if it is determined that a detailed description of known techniques related to the present invention may unnecessarily obscure the gist of the present invention, the detailed description will be omitted. Hereinafter, preferred embodiments according to the present invention will be described in detail with reference to the attached drawings. In the drawings, identical reference numerals are used to indicate identical or similar components.

Figures 1 and 2 are diagrams for explaining a quality inspection system for unstructured data and mycelial leather according to an embodiment of the present invention.

Referring to Figures 1 and 2, the quality inspection system for unstructured data and mycelial leather includes an analysis object recognition device 100 and a quality inspection server 200.

The analysis object recognition device 100 recognizes the analysis object 10, generates a test object image for the analysis object, and provides the image to the quality inspection server 200.

The analysis object recognition device 100 includes a plurality of cameras capable of photographing the analysis object 10 located on the upper side of the plate 20. These plural cameras include side cameras 120 installed at both edges, and a plurality of upper cameras 110 that photograph the analysis object 10 from above. At this time, a plurality of cameras can generate ultra-high resolution (40 million pixels or more) images.

The upper camera 110 is designed passively so that the height value can be confirmed numerically. For example, when acquiring AI images and performing inspection functions, it can be confirmed that the upper camera has moved +120mm compared to the z-axis origin.

The upper camera 110 can capture high-resolution images, and can generate images of different types of test objects by photographing the analysis target 10 on the plate 20. For example, the upper camera 110 can generate a total of two types of test object images, one type 50cm x 50cm and one type 1m x 1m.

The position of the upper camera 110 can be vertically adjusted according to the lens focus, and the lens can be replaced.

The side camera 120 can generate different types of height data by photographing the analysis target 10 of the plate 20, and the height can be adjusted.

The quality inspection server 200 is a server that receives the test object image for the analysis object 10 from the analysis object recognition device 100 and inspects the quality of the analysis object 10 using the test object image.

First, the quality inspection server 200 displays all pixels of the test object image as a single color value from 0 to 255. The reason for this is that by removing the color information of the test object image, there is no color information in the test object image, so it is easy to identify features and patterns.

Afterwards, the quality inspection server 200 converts the pixel value into a binary value according to the pixel value of the test object image.

In one embodiment, the quality inspection server 200 may convert the pixel value to 0 when the pixel value of the test object image is above a certain value, and if the pixel value is below a certain value, the pixel value may be converted to 0. It can be converted to 1.

In another embodiment, the quality inspection server 200 may convert the pixel value to 1 when the pixel value of the test object image is above a certain value, and if the pixel value is below a certain value, the pixel value may be converted to 1. can be converted to 0.

As described above, the quality inspection server 200 converts the pixel value of the test object image into a binary value to facilitate object detection and borderline detection.

Then, the quality inspection server 200 extracts feature points by applying a feature point extraction algorithm to the test object image. At this time, feature points can be extracted from the specimen image by applying a feature point extraction model to the specimen image for raw hide (i.e., leather) and the specimen image for raw materials (i.e., mycelial culture). At this time, the feature point extraction model refers to the Mask R-CNN model based on Faster R-CNN.

In one embodiment, the quality inspection server 200 may extract the boundary line of the test object image, identify the outline of the object, and extract feature points.

More specifically, the quality inspection server 200 may generate a plurality of bounding boxes to extract the shape and position size of the object based on the boundary line of the test object image and extract feature points of the plurality of bounding boxes.

In one embodiment, the quality inspection server 200 calculates the distance between the coordinate values by comparing the coordinate value of the first bounding box and the coordinate of the second bounding box among the plurality of bounding boxes, and when the distance value is less than a specific value. After merging the first bounding box and the second bounding box to create one bounding box, feature points of the bounding box can be extracted.

Then, the quality inspection server 200 learns the features extracted from the test object image to determine whether the analysis object 10 is good or bad.

More specifically, the quality inspection server 200 generates a bounding box matrix using a plurality of bounding boxes. At this time, the bounding box matrix records the binary value of the bounding box for each row. For example, the quality inspection server 200 generates a bounding box matrix by recording binary values of the bounding box on a row-by-row basis.

As described above, the quality inspection server 200 generates a bounding box matrix using a plurality of bounding boxes and then deep learns the characteristics of the bounding box matrix of the plurality of bounding boxes to generate a defective judgment model that determines whether it is defective. do.

To this end, the quality inspection server 200 identifies the characteristics of the bounding box matrix using a convolutional neural network (CNN).

More specifically, the quality inspection server 200 learns a defective judgment model by filtering each part of the bounding box represented by the bounding box matrix and recording values representing the characteristics of each coordinate part of the bounding box in the output matrix. do.

At this time, the quality inspection server 200 changes the numerical value representing each feature of the coordinates of the bounding box through a filter and records the corresponding numerical value in the output matrix to learn a defective judgment model.

In one embodiment, the quality inspection server 200 records a first binary value representing the characteristics of a defective part through a filter into an output matrix to generate a defect determination model, and then creates an inspection object image through the defect determination model. Determine whether or not it is defective.

In the above embodiment, the quality inspection server 200 determines whether the test object image is defective and provides it through a defect judgment model, and then compares whether the test object image is defective and whether the test object actually determined is defective. According to the comparison result, the first binary value expressing the characteristics of the defective part is changed to a second binary value through a filter, and the second binary value is recorded in the output matrix to generate a defective judgment model. The decision model determines whether the image of the test object is defective and provides it.

That is, the quality inspection server 200 compares the predicted defective decision result and the actual defective decision result of the corresponding analysis object 10, and filters according to the comparison result to create a first binary number representing the characteristics of the defective part. After changing the value to a second binary value, the second binary value is recorded in the output matrix to generate a defective judgment model, and then the defectiveness of the inspection object image is determined through the defective judgment model.

At this time, the first binary value and the second binary value are different values and can be 0 or 1. If the first binary value is 1, the second binary value is 0, and if the first binary value is 0, the second binary value is 0. The binary value is 1.

Figure 3 is a block diagram for explaining the internal structure of a quality inspection server according to an embodiment of the present invention.

Referring to FIG. 3 , the quality inspection server 200 includes a preprocessor 210, a feature point extraction unit 220, a defect determination model learning unit 230, and a defect determination unit 240.

First, the preprocessor 210 displays all pixels of the test object image as a single color value from 0 to 255. The reason for this is that by removing the color information of the test object image, there is no color information in the test object image, so it is easy to identify features and patterns.

Afterwards, the preprocessor 210 converts the pixel value into a binary value according to the pixel value of the test object image.

In one embodiment, the preprocessor 210 may convert the pixel value to 0 when the pixel value of the test object image is greater than or equal to a certain value, and may convert the pixel value to 1 when the pixel value is less than or equal to a certain value. It can be converted to .

In another embodiment, the preprocessor 210 may convert the pixel value to 1 when the pixel value of the test object image is above a certain value, and when the pixel value is below a certain value, the pixel value may be converted to 1. It can be converted to 0.

As described above, the pre-processing unit 210 converts the pixel value of the test object image into a binary value to facilitate object detection and borderline detection.

Then, the feature point extraction unit 220 extracts feature points by applying a feature point extraction algorithm to the test object image. At this time, feature points can be extracted from the specimen image by applying a feature point extraction model to the specimen image for raw hide (i.e., leather) and the specimen image for raw materials (i.e., mycelial culture). At this time, the feature point extraction model refers to the Mask R-CNN model based on Faster R-CNN.

In one embodiment, the feature point extraction unit 220 extracts the boundary line of the test object image, determines the outline of the object, and extracts the feature point.

More specifically, the feature point extraction unit 220 may generate a plurality of bounding boxes to extract the shape and position size of the object based on the boundary line of the object image and extract the feature points of the plurality of bounding boxes.

In one embodiment, the feature point extractor 220 calculates the distance between the coordinate values by comparing the coordinate values of the first bounding box and the coordinates of the second bounding box among the plurality of bounding boxes, and when the distance value is less than a specific value. After merging the first bounding box and the second bounding box to create one bounding box, feature points of the bounding box can be extracted.

Then, the defect determination unit 240 learns the features extracted from the test object image to determine whether the analysis object 10 is good or bad.

More specifically, the defective decision model learning unit 230 generates a bounding box matrix using a bounding box. At this time, the bounding box matrix records the binary value of the bounding box for each row. For example, the defective decision model learning unit 230 records the binary values of the bounding box on a row-by-row basis to generate a bounding box matrix.

As described above, the defective decision model learning unit 230 generates a bounding box matrix using a plurality of bounding boxes and then deep learns the characteristics of the bounding box matrix of the bounding box to generate a defective decision model that determines whether it is defective. do.

To this end, the defective decision model learning unit 230 determines the characteristics of the bounding box matrix using a convolutional neural network (CNN).

More specifically, the defective decision model learning unit 230 learns a defective decision model by filtering each part of the bounding box represented by the bounding box matrix and recording values representing the characteristics of the bounding box in the output matrix.
At this time, the defective decision model learning unit 230 changes the numerical value representing the characteristics of the plurality of bounding boxes through a filter and records the corresponding numerical value in the output matrix to learn the defective decision model.

delete

In one embodiment, the defective decision model learning unit 230 records the first binary value representing the characteristics of the corresponding part through a filter in an output matrix to generate a defective decision model, and then prints the test object image through the defective decision model. Determine whether or not the product is defective.

In the above embodiment, the defective decision model learning unit 230 determines whether the test object image is defective and provides it through a defective decision model, and then determines whether the test object image is defective and whether the actually determined test object is defective. Compare, change the first binary value representing the characteristics of the defective part through a filter according to the comparison result to a second binary value, and record the second binary value in the output matrix to generate a defect determination model. Afterwards, it determines whether the image of the test object is defective and provides it through a defect judgment model.

That is, the defect determination model learning unit 230 compares the predicted defect determination result and the actual defect determination result of the corresponding analysis object 10, and uses a filter according to the comparison result to express the characteristics of the part where the defect occurred. After changing the first binary value to a second binary value, the second binary value is recorded in the output matrix to create a defective judgment model.

At this time, the first binary value and the second binary value are different values and can be 0 or 1. If the first binary value is 1, the second binary value is 0, and if the first binary value is 0, the second binary value is 0. The binary value is 1.

The defect determination unit 240 determines whether the image of the test object is defective through the defect determination model generated by the defect determination model learning unit 230.

Figure 4 is a flowchart illustrating an embodiment of the method for inspecting the quality of unstructured data and mycelial leather according to the present invention.

Referring to FIG. 4, the analysis object recognition device 100 recognizes the analysis object and generates and provides a test object image for the analysis object (step S310).

When the quality inspection server 200 receives a test object image from the analysis object recognition device, it performs preprocessing on the test object image (step S320).

In one embodiment for step S320, the quality inspection server 200 displays all pixels of the test object image with a single color value from 0 to 255. The reason for this is that by removing the color information of the test object image, there is no color information in the test object image, so it is easy to identify features and patterns.

Afterwards, the quality inspection server 200 converts the pixel value into a binary value according to the pixel value of the test object image (step S320).

In one embodiment, the quality inspection server 200 may convert the pixel value to 0 when the pixel value of the test object image is above a certain value, and if the pixel value is below a certain value, the pixel value may be converted to 0. It can be converted to 1.

In another embodiment, the quality inspection server 200 may convert the pixel value to 1 when the pixel value of the test object image is above a certain value, and if the pixel value is below a certain value, the pixel value may be converted to 1. can be converted to 0.

As described above, the quality inspection server 200 converts the pixel value of the test object image into a binary value to facilitate object detection and borderline detection.

The quality inspection server 200 performs preprocessing on the test object image and then extracts features from the preprocessed test object image (step S340).

In one embodiment of step S340, the quality inspection server 200 may extract the boundary line of the test object image, identify the outline of the object, and extract feature points.

More specifically, the quality inspection server 200 may generate a plurality of bounding boxes to extract the shape and position size of the object based on the boundary line of the test object image and extract feature points of the plurality of bounding boxes.

In one embodiment, the quality inspection server 200 calculates the distance between the coordinate values by comparing the coordinate value of the first bounding box and the coordinate value of the second bounding box among the plurality of bounding boxes, and the distance value is a specific value. In the following cases, the first bounding box and the second bounding box are merged to create one bounding box, and then feature points for extracting feature points of the bounding box can be extracted.

The quality inspection server 200 learns a defective judgment model using the features (step S350).

In one embodiment of step S350, the quality inspection server 200 generates a bounding box matrix using a bounding box and generates a defective judgment model that determines whether it is defective by deep learning the characteristics of the bounding box matrix of the bounding box. do. At this time, the bounding box matrix records the binary value of the bounding box for each row. For example, the quality inspection server 200 generates a bounding box matrix by recording binary values of the bounding box on a row-by-row basis.

To this end, the quality inspection server 200 identifies the characteristics of the bounding box matrix using a convolutional neural network (CNN).

More specifically, the quality inspection server 200 learns a defect determination model by filtering each part of the bounding box represented by the bounding box matrix and recording values representing the characteristics of the defective part in the output matrix.

At this time, the quality inspection server 200 changes the numerical value representing the characteristics of the corresponding bounding box through a filter and records the numerical value in the output matrix to learn a defective judgment model.

The quality inspection server 200 determines whether the image of the test object is defective through the defect determination model (step S350).

Figures 5 to 7 are exemplary views for explaining the quality inspection process of unstructured data and mycelial leather according to an embodiment of the present invention.

Referring to FIGS. 5 to 7, the quality inspection server 200 converts all pixels of the test object image (FIGS. 5(1), 6(1), and 7(1)) into a single color value from 0 to 255. to generate a gray image (Figure 5(2), Figure 6(2), Figure 7(2)). The reason for this is that by removing the color information of the test object image, there is no color information in the test object image, so it is easy to identify features and patterns.

Afterwards, the quality inspection server 200 converts the pixel value into a binary value according to the pixel value of the test object image.

In one embodiment, the quality inspection server 200 may convert the pixel value to 0 when the pixel value of the test object image is above a certain value, and if the pixel value is below a certain value, the pixel value may be converted to 0. It can be converted to 1.

In another embodiment, the quality inspection server 200 may convert the pixel value to 1 when the pixel value of the test object image is above a certain value, and if the pixel value is below a certain value, the pixel value may be converted to 1. can be converted to 0.

As described above, the quality inspection server 200 converts the pixel value of the test object image into a binary value to facilitate object detection and borderline detection.

The quality inspection server 200 performs preprocessing on the test object image and then extracts features (FIGS. 5(4), 6(4), and 7(4)) from the preprocessed test object image.

More specifically, the quality inspection server 200 can generate a bounding box to extract the shape and position size of the object based on the boundary line of the test object image and extract the feature points of the bounding box (FIG. 5(5), FIG. 6(5), Figure 7(5)).

In one embodiment, the quality inspection server 200 compares the coordinate values of the bounding boxes to calculate the distance between the coordinate values, and if the distance value is less than a certain value, merges the coordinate values of the bounding boxes to create one bounding box. After that, the feature points of the bounding box can be extracted (Figures 5(6), 6(6), and 7(6)).

Afterwards, the quality inspection server 200 outputs a bounding box on the test object image (FIGS. 5 (result), 6 (result), and 7 (result)).

Although the present invention has been described with reference to limited embodiments and drawings, the present invention is not limited to the above embodiments, and various modifications and variations can be made by those skilled in the art from these descriptions. Accordingly, the spirit of the present invention should be understood only by the scope of the claims set forth below, and all equivalent or equivalent modifications thereof shall fall within the scope of the spirit of the present invention.

100: analysis object recognition device,
110: upper camera,
120: side camera,
200: quality inspection server,
210: preprocessing unit,
220: feature point extraction unit,
230: Defect judgment model learning unit,
240: Defect judgment unit

Claims (2)

All pixels of the test object image are displayed as a single color value, and if the pixel value of the test object image is above a certain value, the pixel value can be converted to either 1 or 0, and if the pixel value is less than a certain value, the pixel value is converted to a single color value. A preprocessor that converts the value to one of 1 and 0;
In order to extract the shape and position size of the object based on the boundary line of the object image, a plurality of bounding boxes are each created to extract feature points of each of the plurality of bounding boxes, and the coordinates of the first bounding box among the plurality of bounding boxes are Compare the coordinate values of the value and the second bounding box to calculate the distance between the coordinate values, and if the distance value is less than a certain value, merge the first bounding box and the second bounding box to create one bounding box. a feature point extraction unit that extracts feature points of the bounding box;
Each part of the bounding box represented by the bounding box matrix is filtered, and the first binary value representing the characteristics of the bounding box is recorded in the output matrix to create a defective decision model. Then, the defective decision model is used to generate a defective judgment model for After determining whether or not it is defective and providing it, comparing the defectiveness of the image of the test object with the defectiveness of the actually determined test object, a first binary value expressing the characteristics of the part where the defect occurred is filtered according to the comparison result. a defective decision model learning unit that changes to a second binary value and then records the second binary value in an output matrix to learn a defective decision model; and
Characterized by comprising a defect determination unit that determines whether the test object image is defective through the defect determination model generated by the defect determination model learning unit.
Quality inspection server. In the quality inspection method executed on the quality inspection server,
All pixels of the test object image are displayed as a single color value, and if the pixel value of the test object image is above a certain value, the pixel value can be converted to either 1 or 0, and if the pixel value is less than a certain value, the pixel value is converted to a single color value. Preprocessing the value by converting it to another of 1 and 0;
In order to extract the shape and position size of the object based on the boundary line of the object image, a bounding box is created to extract feature points of the bounding box, and the coordinate value of the first bounding box and the second bounding box among the plurality of bounding boxes are Compare coordinate values to calculate the distance between coordinate values, and if the distance value is less than a certain value, merge the first bounding box and the second bounding box to create one bounding box, and then extract the feature points of the bounding box. Extracting feature points;
generating a defective decision model by recording first binary values representing features of each part of the bounding box represented by the bounding box matrix through a filter into an output matrix;
After determining and providing the defective image of the test object through the defect determination model, whether the test object image is defective and whether the actually determined test object is defective is compared, and the defect is determined through a filter according to the comparison result. generating a defect determination model by changing a first binary value representing the characteristics of the generated part into a second binary value and then recording the second binary value in an output matrix; and
Characterized in that it includes the step of determining whether the test object image is defective through the defect determination model.
Quality inspection methods.