KR102576213B1 - Quality control system and method using learning module - Google Patents

Abstract

A quality control system using a learning model according to an embodiment of the present invention includes a transfer unit that transports tofu contained in a container with one side open, an input unit that photographs the tofu and obtains image data for it, and based on the image data, It includes a determination unit that determines whether the tofu is a normal product or a defective product, and an output unit that informs that the tofu is a defective product, and the determination unit determines that foreign substances other than tofu are contained in the container through a learning model learned using machine learning logic. If there is, it can be determined as a defective product, and if the corner of the head is uneven or dented, it can be determined as a defective product.

Description

Quality control system and method using learning module}

The present invention relates to a quality control system that uses an artificial intelligence learning module to quickly and accurately detect products containing foreign substances in containers or products with uneven tofu shapes in the tofu production process.

In general, the manufacturing process of tofu involves soaking soybeans, grinding them, filtering them, boiling soy milk, adding a coagulant, stirring it to coagulate it, forming and pressing it, cooling it, cutting it, and packaging it.

There are cases where foreign substances introduced into the product during the tofu manufacturing process are included. For example, during the manufacturing process of tofu products, foreign substances such as hair, bugs, yuba, carbide, and plastic may enter. Because these foreign substances can cause problems with product stability, it is important to detect products containing foreign substances in advance and manage them so that they are not shipped as is.

In addition, tofu products can be distributed on the market packaged in containers made of PP. In order to package tofu products in containers, they are molded in a mold and then cut to a certain size by a cutter with a blade formed on one side. You can. During this process, tofu products may not be manufactured in a consistent shape due to damage or breakage such as crumbling at the edges or uneven cutting surfaces. Even if defects in the appearance of these tofu products do not have a significant impact on stability, they can be a factor in reducing marketability. Products with poor marketability act as a factor in reducing product reliability, so it is important to detect such products in advance and manage them so that they are not shipped as is.

However, in the past, there was a problem of low inspection probability by visually inspecting whether the packaging container contained foreign substances or whether the shape of the tofu included in the packaging container was unmarketable. In addition, each inspector had different standards for classifying products as defective, which led to a problem of inconsistent product quality.

Accordingly, in products such as tofu products that easily contain foreign substances due to the manufacturing process or the environment or are difficult to manufacture in a certain form, there is a need for a method that can easily, accurately, and uniformly inspect defective products. .

The present invention seeks to secure the reliability of the product by easily and accurately inspecting the quality of tofu products before shipping using a machine learning model. Specifically, the purpose is to improve product quality and secure product reliability by detecting foreign substances such as metal and hair that have entered the product, as well as precisely inspecting the appearance of the product.

As a means to achieve the above-described task, a quality control system using a learning model according to an embodiment of the present invention includes a transfer unit for transporting tofu contained in a container with one side open, and image data for the tofu by photographing the tofu. It includes an input unit that obtains, a determination unit that determines whether the tofu is a normal product or a defective product based on the image data, and an output unit that informs that the tofu is a defective product, and the determination unit is a learning model learned using machine learning logic. Through this, if the container contains foreign substances other than tofu, it can be determined as a defective product, and if the edge of the tofu is uneven or dented, it can be determined as a defective product.

In the quality control system using a learning model according to an embodiment of the present invention, the learning model determines normal products and defective products among image data classified into five or more classes according to the degree to which the corners of the head are uneven or distorted. It can be learned excluding data of the following unclear classes.

A quality control system using a learning model according to an embodiment of the present invention includes a first learning model in which the learning model is learned through image data classified according to the type of foreign matter, and a first learning model that does not contain foreign matter and has a corner of the head. A second learning model may be included through image data classified according to the degree of unevenness or distortion.

A quality control system using a learning model according to an embodiment of the present invention can learn whether the corners of the head are even through a virtual boundary line formed based on the corners of the head and edge deviation data of the head.

A quality control system using a learning model according to an embodiment of the present invention is that the deviation data determines at least one of the area formed by the virtual boundary line and the edge of the head being spaced apart, the direction of the separation, and the maximum degree of separation. It can be included.

The quality control system using a learning model according to an embodiment of the present invention further includes a plurality of light source units that irradiate light at different angles toward the head, and the input unit turns on/off the plurality of light source units to change the A plurality of image data may be acquired, and the determination unit may determine whether the tofu is a normal product or a defective product based on the plurality of image data obtained with different illumination directions.

A quality control system using a learning model according to an embodiment of the present invention is a quality control system in which the learning model determines the size of the head through at least one of the size, location, and direction of a shaded area included in image data obtained by irradiating different illumination angles. Edges can be learned to be even.

In a quality control system using a learning model according to an embodiment of the present invention, the plurality of light source units may irradiate polarized light so as not to reflect on the liquid contained in the container.

A quality control method using a learning model according to an embodiment of the present invention includes a transport step of transporting tofu in a container with one side open in order to classify the tofu contained in the container into a normal product and a defective product, and photographing the tofu. An input step of acquiring image data, a determination step of determining whether the tofu is a normal product or a defective product based on the image data obtained in the input step, and an output step of outputting that the tofu is a defective product, The discrimination step is a learning step of learning the image data using machine learning logic, and determining if the container contains foreign substances other than tofu as a defective product through a learning model acquired through the learning step, and determining that the edge of the tofu is a defective product. It may include the step of determining that an uneven or distorted product is a defective product.

The quality control method using a learning model according to an embodiment of the present invention includes the learning step of classifying the head image data into five or more classes according to the degree to which the corners of the head are uneven or distorted, and a plurality of classified It may include a step of excluding a class in which it is unclear to determine the tofu as a normal product or an abnormal product among the classes, and learning using the remaining image data in addition to the image data of the excluded class.

The quality control method using a learning model according to an embodiment of the present invention includes the learning step of forming a boundary surface based on the corner of the head in the image data of the head, and the difference between the formed boundary surface and the corner of the head. It may include acquiring data, and learning using the obtained deviation data.

The quality control method using a learning model according to an embodiment of the present invention includes a first image data acquisition step in which the input step acquires image data by irradiating light at a first angle toward the head, and It may include a second image data acquisition step of acquiring image data by irradiating light at a second angle that is different from the first angle.

The quality control method using a learning model according to an embodiment of the present invention includes the learning step of distinguishing shaded areas of the first image data and the second image data, and using the differentiated shaded area data. It may include a learning step.

According to the present invention, it is possible to detect foreign substances such as metal and hair that have entered the product, as well as to easily and precisely perform a quality inspection on the appearance of the product. Accordingly, the quality of the product can be improved and the reliability of the product can be secured.

According to the present invention, it is possible to provide consistency in standards for detecting foreign substances and quality inspection for product appearance, thereby ensuring consistency in product quality.

Of course, the scope of the present invention is not limited by this effect.

1 is a block diagram of a quality control system using a learning model according to an embodiment of the present invention.
Figure 2 is a diagram for explaining normal products and defective products according to an embodiment of the present invention.
Figures 3 and 4 are diagrams for explaining a learning model for determining whether the edges of the head are even according to an embodiment of the present invention.
Figure 5 is a diagram for explaining data used in a learning model that determines whether the edges of the head are even according to an embodiment of the present invention.
Figure 6 is a diagram for explaining data used in a learning model that determines whether the edges of the head are even according to an embodiment of the present invention.
Figure 7 is a flowchart for explaining a quality control method using a learning model according to an embodiment of the present invention.

Hereinafter, specific embodiments of the present invention will be described with reference to the drawings. The detailed description below is provided to provide a comprehensive understanding of the methods, devices and/or systems described herein. However, this is only an example and the present invention is not limited thereto.

In describing the embodiments of the present invention, if it is determined that a detailed description of the known technology related to the present invention may unnecessarily obscure the gist of the present invention, the detailed description will be omitted. In addition, the terms described below are terms defined in consideration of functions in the present invention, and may vary depending on the intention or custom of the user or operator. Therefore, the definition should be made based on the contents throughout this specification. The terminology used in the detailed description is merely for describing embodiments of the present invention and should in no way be limiting. Unless explicitly stated otherwise, singular forms include plural meanings. In this description, expressions such as “comprising” or “comprising” are intended to indicate certain features, numbers, steps, operations, elements, parts or combinations thereof, and one or more than those described. It should not be construed to exclude the existence or possibility of any other characteristic, number, step, operation, element, or part or combination thereof.

In this specification, 'part' includes a unit realized by hardware, a unit realized by software, and a unit realized using both. Additionally, one unit may be realized using two or more pieces of hardware, and two or more units may be realized using one piece of hardware. Meanwhile, '~ part' is not limited to software or hardware, and '~ part' may be configured to reside in an addressable storage medium or may be configured to reproduce one or more processors. Therefore, as an example, '~ part' refers to components such as software components, object-oriented software components, class components, and task components, processes, functions, properties, and procedures. , subroutines, segments of program code, drivers, firmware, microcode, circuits, data, databases, data structures, tables, arrays, and variables. The functions provided within the components and 'parts' may be combined into a smaller number of components and 'parts' or may be further separated into additional components and 'parts'. Additionally, components and 'parts' may be implemented to regenerate one or more CPUs within a device or a secure multimedia card.

In cases where an embodiment can be implemented differently, a specific process sequence may be performed differently from the described sequence. For example, two processes described in succession may be performed substantially at the same time, or may be performed in an order opposite to that in which they are described.

Hereinafter, an embodiment of the present invention will be described in detail with reference to the attached drawings.

Figure 1 is a block diagram of a quality management system 100 using a learning model according to an embodiment of the present invention.

The quality control system 100 using a learning model according to an embodiment of the present invention may include a transfer unit 110, an input unit 120, a determination unit 130, and an output unit 140.

The transport unit 110 may be a device that transports the head, which is an inspection object, including a conveyor belt, a drive motor, and an encoder.

The conveyor belt is a part that can transport the mounted inspection object, and materials such as rubber, synthetic rubber, and steel cord can be used. A transport path through which the product to be inspected is transported is formed on the conveyor belt, and for smooth transport of the product to be inspected, the transport path can be formed as a straight path.

The conveyor belt can be circulated by a driving part for automation of the device. The driving part may be formed of a driving motor, and may be provided with an encoder that can detect the operation of the driving motor. The encoder can detect the rotational speed or direction of the drive motor, generate a detection signal, and control the drive motor based on the detection signal.

On the transfer unit 110, the product to be inspected may be moved in a state where a plurality of products are positioned horizontally with respect to the direction of movement. For example, two products to be inspected can be placed side by side and moved on a conveyor belt at preset intervals.

The transfer unit 110 may include a part that transports the product to be inspected to a location for inspection and a part that transports the product to a location for packaging after inspection. Additionally, the transfer unit 110 may include a component that inspects the inspection target product and excludes defective products. The configuration for excluding defective products may be a configuration that separates defective products from normal products, changes the line, or removes defective products from the transfer unit 110.

The input unit 120 may include a camera. The input unit 120 may acquire a video image of the inspection object being transferred on the transfer unit 110. The video image may be an image that includes the inspection target item and its surrounding area. For example, a video image may be acquired that includes tofu and a container containing tofu.

At least one camera may be provided, and in one embodiment, the camera may include at least one of a vision camera, an X-ray camera, and a hyperspectral camera.

A vision camera is a device that can acquire RGB images in the 400nm to 700nm range and may include, for example, a CCD camera or a CMOS camera. Through a vision camera, images of the shape, shape, and color of the inspection target article and images of foreign substances contained in the exterior of the inspection target article can be obtained.

An X-ray camera consists of an emitter and a receiver, and the X-rays emitted by the emitter can pass through the object to be inspected and then converted into an image by the receiver. Through an X-ray camera, images of foreign substances included in the inspection object, for example, foreign substances with high density such as metal, glass, stone, etc., can be obtained.

A hyperspectral camera is a device that can acquire hyperspectral images in the range of 400nm to 1700nm. Since it can acquire images divided by wavelength, it can acquire images of foreign substances contained in other components of the inspection target product. .

The input unit 120 may include a lighting device to make it easier to view video images or to obtain different video images depending on the situation. Here, the lighting device may correspond to the light source unit. In some cases, the light source unit may be included in the present invention as a separate component from the input unit 120.

The lighting device can irradiate light containing wavelengths in the visible and near-infrared regions, for example, 380 to 1700 nm. In one embodiment, the lighting device may be configured using special lighting such as a near infrared light emitting diode (NIR LED) or a halogen lamp.

A plurality of lighting devices may be included in different positions to vary the direction of irradiating light to the inspection object. The camera included in the input unit 120 can acquire a plurality of video images for one inspection target product in response to a plurality of lighting devices being turned on/off sequentially or in a preset combination.

The video image of the object to be inspected obtained from the input unit 120 may be transmitted to the determination unit 130.

The determination unit 130 may include an image analysis unit 131 to determine whether the inspection object is normal. Here, the image analysis unit 131 may be configured to process commands of a computer program by performing basic arithmetic, logic, and input/output operations. Commands may be provided to the image analysis unit 131 by the memory unit 132 or the communication unit 133. Here, the image analysis unit 131 may mean, for example, a data processing device built into hardware that has a physically structured circuit to perform a function expressed as a code or command included in a program.

Examples of data processing devices built into hardware include microprocessor, central processing unit (CPU), processor core, multiprocessor, and application-specific integrated (ASIC). Circuit), FPGA (Field Programmable Gate Array), etc., but the scope of the present invention is not limited thereto.

The video analysis unit 131 can analyze video data using a learning model learned using machine learning logic. Specifically, it is possible to determine whether the inspection target product is a normal product or a defective product using a learning model learned with deep learning logic of the R-CNN method. Here, the deep learning logic of the R-CNN (Regions with Convolution Neural Network) method adds the concept of dividing the search part regionally to the concept of CNN. Rather than searching all images in the shooting area, the deep learning logic of the R-CNN (Regions with Convolution Neural Network) method adds the concept of dividing the search part regionally. is to explore. Specifically, the learning model can be learned using deep learning logic as a specific area containing a container containing tofu.

The learning model is trained on image data of tofu (specifically, tofu contained in a container with one side open), which is the product to be inspected, and classifies it as a defective product if there is a foreign substance in the container. If the edge of the tofu is uneven or dented, it is classified as a defective product. It can be distinguished as:

The learning model is stored in the memory unit 132 in the form of at least one of application software executable in the image analysis unit 131 or an executable file executable through application software, or can be communicated through the communication unit 133. Can be included on an external server.

The output unit 140 can output the results determined through the determination unit 130 so that the user can see them. The output unit 140 may display a defective product through a video image or notify that the product to be inspected is a defective product through a notification sound. In some cases, the output unit 140 may correspond to an electronic device that the user can separately carry.

Specifically, the output unit 140 can distinguish between cases where the product to be inspected is determined to be a defective product because it contains foreign substances, and cases where it is determined to be a defective product because the corners of the head are uneven or distorted. In some cases, the output unit 140 may distinguish and display cases in which the product to be inspected is determined to be a defective product due to the presence of foreign substances.

Figure 2 is a diagram for explaining normal products and defective products according to an embodiment of the present invention. Specifically, Figure 2 shows video images taken with tofu, a product to be inspected, contained in a container, divided into normal products (a) and defective products (b to h).

Tofu, which is a product subject to inspection, can be placed in a container and commercialized in a state that is cut into a square shape when viewed from one side (ex, top side). The video image may correspond to an image taken with tofu submerged in liquid in a container with one side open. Specifically, the first video image (a) is a video image corresponding to a normal product, and the second to eighth video images (b to h) are video images corresponding to a defective product.

A normal product can correspond to a product with even edges when viewed from the top and a product with no foreign substances inside the container. Defective products can be divided into cases where the edges of the tofu are uneven or dented and cases where foreign substances are contained inside the container. Specifically, the second video image (b) shows an example in which one corner of the head (area B) is uneven and distorted. The third to eighth video images (c to h) show an example in which a foreign substance is contained in the head. Specifically, the third video image (c) shows an example in which metal is present in region C within the container. The fourth video image (d) shows an example in which there is hair in area D within the container. The fifth video image (e) shows an example in which bugs are present in area E within the container. The sixth video image (f) shows an example in which yuba (cephalic skin) is present in area F within the container. The seventh video image (g) shows an example in which carbide is present in the G region within the container. The eighth video image (h) shows an example in which plastic is present in the H area within the container.

Video data can be divided into classes into the above eight cases and a learning model can be formed through machine learning. The learning model learned through image data divided into classes distinguishes normal products from normal products and determines cases where foreign substances other than tofu are contained in the container as defective products, and determines cases where the corners of the tofu are uneven or distorted as defective products. You can.

In some cases, the learning model learned through image data is a first learning model learned through image data classified according to whether or not it contains foreign substances, and a first learning model learned through image data that does not contain foreign substances and is classified according to the degree to which the corners of the head are uneven or distorted. It may include a second learning model learned through image data. Specifically, the video data includes the first video image (a) and the second video image (b) as one class, and the third to eighth video images (c to h) are divided into respective classes. The first learning model can be formed using this. Also, only the first video image (a) and the second video image (b) can be divided into respective classes and used to form a second learning model.

The learning model can determine whether a foreign substance is included through the first learning model. If foreign substances are included, it is determined to be a defective product, and if foreign substances are not included, the second learning model can be used to determine whether the edges of the tofu are even. If the corners of the tofu are uneven or dented, it can be judged as a defective product, and if the corners of the tofu are even, it can be judged as a normal product.

When the first learning model is formed by dividing the image data into the presence or absence of foreign substances and their type, and the second learning model is formed by dividing the image data without foreign substances by whether the edges of the head are even, a more effective learning model because the objects to be judged are different. This can be formed. Specifically, the first learning model learns based on the container, and the second learning model learns based on the head, so it may be desirable to learn them separately.

However, products are classified into normal and abnormal products depending on the degree to which the edges of the head are uneven or distorted, but the boundary of judgment may be unclear. To solve this problem, it is explained below.

Figures 3 and 4 are diagrams for explaining a learning model for determining whether the edges of the head are even according to an embodiment of the present invention. Specifically, Figure 3 shows an example of image data classification used in a learning model that determines whether the tofu is even. FIG. 4 is a diagram for explaining a learning process using image data divided as shown in FIG. 3.

Distinguishing between good and defective products based on whether the edges of the tofu are even or not is a relative evaluation, so it is difficult to make a dichotomous distinction. In fact, even when conventional inspectors judged with the naked eye, each inspector had different judgment standards, resulting in a lack of uniformity in product completeness.

To solve this problem, the image data is divided into five or more classes according to the degree to which the corners of the head are uneven or distorted, and after excluding the image data of classes where it is unclear to distinguish between normal and defective products among the classified classes, You can form a learning model by learning machine learning with logic. Here, the image data used may be image data that does not contain foreign substances.

Figure 3 shows an example in which image data is divided into five classes according to the degree to which the corners of the head are uneven or distorted. Classes A to E are examples classified according to the degree to which the edges of the head are uneven or distorted. Specifically, class A is an example of image data with the most even edges of the head, and class E is an example of image data with the most uneven and distorted edges of the head. Class A and Class B are classified as normal products, Class D and Class E are classified as defective products, and Class C is excluded and can be learned using machine learning logic. The learning model learned in this way has a clear difference between class B and class D, making it possible to more clearly distinguish between normal and defective products. At this time, if the performance of the learned model is poor, the image data corresponding to class C can be selected again and learned using machine learning logic in the same way.

In practice, it is important to make the difference between class B and class D clear, so only class B and class D can be trained with machine learning logic. Referring to FIG. 4, in all the image data 510 (in some cases, the image data used here may be image data that does not contain foreign substances), an image data cluster 520 is used to determine whether the edges of the head are even. ) can be selected and learned (530) using machine learning logic. The image data cluster 520 for determining whether the edges of the head are even can be divided into an image data cluster 521 corresponding to class B and an image data cluster 522 corresponding to class D. Similarly, at this time, if the performance of the learned model is poor, the image data corresponding to class C can be selected again and learned using machine learning logic in the same way.

Classification of video data can be done through user labeling and grouping through self-learning. At this time, grouping may be easier when there is no foreign matter in the image data. It may be desirable for the user to determine and specify classes excluded from grouped image data.

Figure 5 is a diagram for explaining data used in a learning model that determines whether the edges of the head are even according to an embodiment of the present invention. Specifically, Figure 5(a) is an example of image data that is determined to be a defective product because the edges of the head are uneven, and Figure 5(b) is a diagram to explain a method of determining this as a defective product.

Video data can be processed before being trained with machine learning logic. The image data is pre-processed with the deviation data of the edge of the head and the virtual boundary line (e1 to e4) formed based on the corner of the head (p1 to p4), and the machine learning logic uses the pre-processed data to form a learning model. can do.

Specifically, the pre-processed deviation data may be data including at least one of the area formed by the virtual boundary line and the edge of the head being spaced apart, the direction of the separation, and the maximum degree of separation. Referring to FIG. 5(b), area A is an area formed by the edge of the head crossing a virtual boundary line, and corresponds to an area formed by the virtual boundary line of area B being spaced apart from the corner of the head. The pre-processed deviation data may be data that separates the area data of regions A and B. Additionally, the pre-processed deviation data may be data containing the maximum distance between the virtual boundary line and the corresponding edge of the head.

The learning module learned using the pre-processed deviation data can determine the head to be defective when the edge of the head exceeds or falls short of the virtual boundary and the area formed is larger than the preset area. Alternatively, if the difference between the area formed beyond the virtual boundary line and the area formed below the virtual boundary line is greater than a preset difference, it can be determined as defective tofu. Alternatively, if the maximum distance between the edges of the head corresponding to the virtual boundary line is greater than or equal to a preset length, the head can be determined to be defective.

Figure 6 is a diagram for explaining data used in a learning model that determines whether the edges of the head are even according to an embodiment of the present invention. Specifically, Figure 6(a) is an example of image data that is determined to be a defective product due to uneven edges of the head, and Figure 6(b) shows a plurality of light source units (L1 to L4) with different illumination directions turned on in a preset order. This is a diagram to explain an embodiment of acquiring image data by turning /off.

The plurality of light source units (L1 to L4) may be positioned differently to irradiate light to the inspection target product at different angles. The camera can acquire a plurality of image data by turning on/off the plurality of light source units (L1 to L4) to provide different illumination directions to one inspection target product. At this time, the plurality of light source units are sequentially turned on/off according to the shutter speed of the camera, and the camera can respectively acquire image data corresponding to the light source unit that is turned on.

Machine learning logic can form a learning model using image data obtained with different lighting directions. Specifically, the learning model preprocesses image data obtained by examining different lighting angles to represent at least one of the size, location, and direction of the shaded area included in the image data, and the machine learning logic uses the preprocessed shaded data. You can form a learning model using .

The learning module learned using the pre-processed shading data determines that it is a normal product if there is no shading in the image data obtained by changing the lighting direction, and determines it to be a normal product if any shading is formed larger than the preset area. Or, if the shadow is dispersed even if the overall shadow area is large, it can be judged as a normal product, or if the shadow is long in the inner or outer direction, it can be judged as a defective product.

Referring to Figure 6(b), the distorted corner of the head in area A displays the shade more clearly and darkly through irradiation from the L1 light source, and the image data obtained through this can better determine the degree to which the corner of the head is distorted. You can.

At this time, the plurality of light source units used may irradiate polarized light so as not to be reflected by the liquid contained in the container.

Figure 7 is a flowchart for explaining a quality control method using a learning model according to an embodiment of the present invention.

The quality control method using a learning model includes a transport step (710) of transporting tofu in a container with one side open, an input step (720) of acquiring image data by photographing the tofu, and based on the image data acquired in the input step. It may include a determination step 730 to determine whether the tofu is a normal product or a defective product, and an output step 740 to output that the tofu is a defective product. After the output step 740, a verification step 750 may be included to verify the determined result.

The determination step (730) is a learning step (731) that learns the image data using machine learning logic, and determines a case where foreign substances other than tofu are contained in the container as a defective product through the learning model acquired through the learning step (731). And, if the corner of the head is uneven or dented, it may include a step 732 of determining that it is a defective product.

The learning step 731 is a step of classifying head image data into 5 or more classes according to the degree to which the corners of the head are uneven or distorted, and a class in which it is unclear to determine whether the head is a normal product or an abnormal product among the plurality of classified classes. It may include a step of excluding, and a step of learning using the remaining image data in addition to the image data of the excluded class. If the results in the verification step 750 are not satisfactory, the excluded image data can be modified and the learning step 761 can be performed again.

In another embodiment, the learning step 731 includes forming a boundary surface based on the corner of the head from image data of the head, obtaining deviation data between the formed boundary surface and the corner of the head, and obtaining the obtained deviation data. It may include a learning step.

In another embodiment, the learning step 731 includes a first image data acquisition step of acquiring image data by irradiating light at a first angle toward the head, and irradiating light at a second angle different from the first angle toward the head. This may include a second image data acquisition step of acquiring image data. Additionally, the learning step 731 may include distinguishing shaded areas of the first image data and the second image data, and learning using the differentiated shaded area data.

Although representative embodiments of the present invention have been described in detail above, those skilled in the art will understand that various modifications can be made to the above-described embodiments without departing from the scope of the present invention. . Therefore, the scope of the present invention should not be limited to the described embodiments, but should be determined by the claims and equivalents of the claims as well as the claims described later.

100: Quality management system
110: transfer unit
120: input unit
130: Determination unit
140: output unit

Claims (13)

A transfer unit that transfers tofu contained in a container with one side open;
An input unit that photographs the head and obtains image data about it;
a determination unit that determines whether the tofu is a normal product or a defective product based on the image data; and
An output unit notifying that the tofu is a defective product,
The learning model is
It includes a first learning model learned through image data classified according to whether or not it contains foreign substances, and a second learning model learned through image data classified according to the degree to which the corner of the head is uneven or distorted and does not contain foreign substances. do,
The determination unit
Through a learning model learned using machine learning logic, if the container contains foreign substances other than tofu, it is determined as a defective product, and if the edge of the tofu is uneven or dented, it is determined as a defective product,
The second learning model is
Among the image data classified into five or more classes according to the degree to which the edges of the head are uneven or distorted, the model is learned by excluding data from classes that are unclear to distinguish between normal and defective products,
The determination unit
The first learning model is used to determine whether foreign substances are included, and if foreign substances are included, it is determined as a defective product. If the foreign substances are not included, the second learning model is used to determine whether the edge of the head is uneven or dented as a defective product. Characterized by a quality control system using a learning model. delete delete According to paragraph 1,
The learning model is a model that learns whether the corners of the head are even through a virtual boundary formed based on the corners of the head and deviation data of the corners of the head.
According to paragraph 4,
The deviation data is characterized in that it includes at least one of the area formed by the virtual boundary line and the edge of the head being spaced apart, the direction of the separation, and the maximum degree of separation.
According to paragraph 1,
The quality control system is
It further includes a plurality of light source units that irradiate light at different angles toward the head,
The input unit acquires a plurality of image data by turning on/off the plurality of light source units,
A quality control system using a learning model, wherein the determination unit determines whether the tofu is a normal product or a defective product based on a plurality of image data obtained with different illumination directions.
According to clause 6,
The learning model is a model that learns whether the edges of the head are even through at least one of the size, position, and direction of the shaded area included in the image data obtained by examining different lighting angles. Quality management system.
According to clause 6,
A quality control system using a learning model, wherein the plurality of light sources irradiate polarized light so as not to reflect on the liquid contained in the container.
A transfer step of transporting tofu in a container with one side open;
An input step of acquiring image data by photographing the head;
A determination step of determining whether the tofu is a normal product or a defective product based on the image data obtained in the input step; and
An output step of outputting that the tofu is a defective product,
The determination step is
A learning step of learning the image data using machine learning logic;
Using the learning model obtained through the learning step, determining that the container contains foreign substances other than tofu as a defective product;
If the container does not contain foreign substances other than tofu, determining that the edge of the tofu is uneven or dented as a defective product,
The learning step is
Classifying the head image data into five or more classes according to the degree to which corners of the head are uneven or distorted;
Excluding a class in which it is unclear to determine the tofu as a normal product or an abnormal product among the plurality of classified classes; and
A quality control method using a learning model, comprising: learning using the remaining image data in addition to the image data of the excluded classes. delete According to clause 9,
The learning step is
forming a boundary surface based on a corner of the head from the image data of the head;
Obtaining deviation data between the formed boundary surface and the edge of the head; and
A quality control method using a learning model, comprising: learning using the obtained deviation data.
According to clause 9,
The input step is
A first image data acquisition step of acquiring image data by irradiating light at a first angle toward the head; and
A quality control method using a learning model, comprising a second image data acquisition step of acquiring image data by irradiating light toward the head at a second angle different from the first angle. According to clause 12,
The learning step is
distinguishing shaded areas of the first image data and the second image data; and
A quality control method using a learning model, comprising: learning using the classified shaded area data.

Images