Method for detecting defect of product and method for training artificial neural network therefor

Abstract

A method for detecting product defects and an artificial neural network learning method for the same are provided. An artificial neural network learning method for detecting product defects according to one embodiment includes: a step of acquiring a first image of a product; a step of acquiring a second image by performing preprocessing on the first image; a step of determining an inspection area in the second image; a step of acquiring a plurality of segments corresponding to the inspection area while rotating the second image by a predetermined angle; a step of acquiring correct answer or defective correct answer data for each of the plurality of segments; and a step of training an artificial neural network using a learning data set composed of the plurality of segments and correct answer data.

Description

Technology Field

[0001] The present invention relates to a method for detecting defects in articles and an artificial neural network learning method for the same. Background Technology

[0003] An artificial neural network refers to a network structured to resemble the connections of neurons in the human brain. Neurons functioning as mathematical models are interconnected to form a network, and it can be referred to as an artificial neural network to distinguish it from biological neural networks.

[0004] As the scope of artificial intelligence applications has recently increased, artificial neural networks, one of the algorithms for implementing AI, are also being applied in many technological fields. The problem to be solved

[0006] One embodiment aims to provide a method for training an artificial neural network used to detect defects in articles such as valve seats.

[0007] One embodiment aims to provide a method for detecting defects in an article using an artificial neural network. means of solving the problem

[0009] To solve the aforementioned technical problem, an artificial neural network learning method for detecting defects in an article according to an embodiment of the present invention may include: a step of acquiring a first image of an article; a step of acquiring a second image by performing preprocessing on the first image; a step of determining an inspection area in the second image; a step of acquiring a plurality of segments corresponding to the inspection area while rotating the second image by a predetermined angle; a step of acquiring correct answer or defective answer data for each of the plurality of segments; and a step of training an artificial neural network using a learning data set composed of the plurality of segments and correct answer data.

[0010] The step of obtaining a second image by performing preprocessing on the first image may include the step of performing masking on the first image.

[0011] The step of obtaining a second image by performing masking on the first image may include the step of performing masking on the outer side of side A and the inner side of side C of the article in the first image.

[0012] The step of performing masking on the outer side of side A and the inner side of side C of the article in the first image may include: the step of obtaining a center point of the article in the first image; the step of obtaining an outer diameter of side A of the article based on the center point; the step of obtaining an inner diameter of side C of the article based on the center point; and the step of performing masking on the outer region of the outer diameter of side A and the inner region of the inner diameter of side C.

[0013] The step of determining an inspection area in the second image may include determining a square containing all of the A, B, and C sides of the article in the second image as the inspection area.

[0014] The step of determining the inspection area as a square that includes all of the A, B, and C sides of the article in the second image above may include the step of determining the inspection area such that the outer diameter of the B side passes through the two vertices of the inspection area.

[0015] The step of determining the inspection area such that the outer diameter of the above-mentioned B-side passes through the two vertices of the inspection area may include the step of determining the inspection area such that the angle formed by two straight lines connecting the center point of the article to the two vertices is greater than or equal to the predetermined angle.

[0016] The above predetermined angle is 18 degrees, and the above plurality of segments may be 20.

[0017] A method for detecting defects in an article according to one embodiment may include: a step of acquiring a first image of an article; a step of acquiring a second image by performing preprocessing on the first image; a step of determining an inspection area in the second image; a step of acquiring a plurality of segments corresponding to the inspection area while rotating the second image by a predetermined angle; a step of detecting defects in the article from the plurality of segments using a trained artificial neural network; and a step of generating and outputting defect data for the detected defects.

[0018] The step of obtaining a second image by performing preprocessing on the first image may include the step of performing masking on the first image.

[0019] The step of obtaining a second image by performing masking on the first image may include the step of performing masking on the outer side of side A and the inner side of side C of the article in the first image.

[0020] The step of performing masking on the outer side of side A and the inner side of side C of the article in the first image may include: the step of obtaining a center point of the article in the first image; the step of obtaining an outer diameter of side A of the article based on the center point; the step of obtaining an inner diameter of side C of the article based on the center point; and the step of performing masking on the outer region of the outer diameter of side A and the inner region of the inner diameter of side C.

[0021] The step of determining an inspection area in the second image may include determining a square containing all of the A, B, and C sides of the article in the second image as the inspection area.

[0022] The step of determining the inspection area as a square that includes all of the A, B, and C sides of the article in the second image above may include the step of determining the inspection area such that the outer diameter of the B side passes through the two vertices of the inspection area.

[0023] The step of determining the inspection area such that the outer diameter of the above-mentioned B-side passes through the two vertices of the inspection area may include the step of determining the inspection area such that the angle formed by two straight lines connecting the center point of the article to the two vertices is greater than or equal to the predetermined angle.

[0024] The above predetermined angle is 18 degrees, and the above plurality of segments may be 20.

[0025] The defect data for the detected defect may include data for a segment containing the defect, the order of the segment containing the defect in the second image, and defect type data.

[0026] The above defect type data may include width size defects, appearance defects, and unclassified defects.

[0027] The step of detecting defects in the article from the plurality of segments using the above-mentioned trained artificial neural network includes the step of determining the article as normal when the trained artificial neural network detects a defect in one of the plurality of segments, and determining the article as defective when the network detects a defect in n consecutive segments among the plurality of segments, wherein n may be an integer greater than 1.

[0028] A product defect detection device according to one embodiment includes a processor and a memory connected to the processor, the memory is configured to store a program, the processor is configured to execute the program, and when the program is executed, steps of any one of the above-described embodiments may be implemented. Effects of the invention

[0030] According to the present invention, accurate and error-free product defect verification is made possible by using an artificial neural network. Furthermore, by using at least one camera and an artificial neural network, defect detection becomes possible solely through camera installation without configuring specific conditions such as a darkened environment. Additionally, by recognizing and analyzing defect patterns, it is possible to respond quickly when defects occur to prevent further defects. Moreover, the artificial neural network can be further trained through the accumulation of defect data, and various process and quality improvements can be achieved through the utilization of big data. Brief explanation of the drawing

[0032] FIG. 1 shows a vision inspection system according to one embodiment of the present invention. FIG. 2 shows an artificial neural network model and the operation of each neuron according to one embodiment of the present invention. FIG. 3 illustrates the training and inference of an artificial neural network according to one embodiment of the present invention. FIG. 4 shows an example of an image used for learning and inference of an artificial neural network according to an embodiment of the present invention. FIG. 5 shows an example of an image used for learning and inference of an artificial neural network according to an embodiment of the present invention. FIG. 6 shows an example of an image used for learning and inference of an artificial neural network according to an embodiment of the present invention. Figure 7 shows examples of images of a normal type valve seat. Figure 8 shows examples of images of defective valve seats. Figure 9 shows examples of images of defective valve seats. Figure 10 shows examples of images of defective valve seats. Figure 11 shows examples of images of defective valve seats. FIG. 12 shows a flowchart of a learning and inference method of an artificial neural network according to one embodiment. FIG. 13 shows a flowchart of an artificial neural network learning method according to one embodiment. FIG. 14 shows a flowchart of a method for detecting defects in articles according to one embodiment. FIG. 15 shows a flowchart of a defect target product report method according to one embodiment. FIG. 16 shows a flowchart of a learning model update method according to one embodiment. Specific details for implementing the invention

[0033] Preferred embodiments of the present invention are described in detail, and examples are shown in the accompanying drawings. The following detailed description with reference to the accompanying drawings is intended to describe preferred embodiments of the present invention rather than merely embodiments that may be implemented according to the embodiments of the present invention. The following detailed description includes details to provide a thorough understanding of the present invention, but the present invention is not required to have all of these details. The embodiments described below are not required to be used separately. Multiple embodiments or all embodiments may be used together, and specific embodiments may be used in combination.

[0034] Most terms used in this invention are selected from those commonly used in the field, but some terms are chosen at the applicant's discretion, and their meanings are described in detail in the following description as necessary. Accordingly, this invention should be understood based on the intended meaning of the terms, rather than their mere names or meanings.

[0036] In this specification, a defect inspection process for the B-side of a valve seat is described as an example. However, the present invention can be used for detecting defects in ring-type molds or similar shapes for various purposes, and the scope of the rights is not necessarily limited to the B-side of a specific valve seat.

[0037] A neural network refers to a network structured to resemble the connections of neurons in the human brain. Neurons functioning as a mathematical model are interconnected to form a network, and it can also be referred to as an artificial neural network to distinguish it from biological neural networks.

[0038] In the present invention, a computing system can implement a neural network and perform data operations. The computing system may refer to a computing processing part of any device capable of data processing, such as a computer, mobile phone, smart hub, laptop, or IoT device. In this specification, the computing system may be abbreviated as "system." The computing system can perform the present invention by executing data, such as an application stored in a storage device.

[0040] FIG. 1 shows a vision inspection system according to one embodiment of the present invention.

[0041] A vision inspection system performs inspection of a part of a process through networking. In particular, the vision inspection system analyzes images captured by a camera module and performs automated control of the process accordingly. The camera module may include a camera unit and a lighting unit. The camera unit may be a capturing unit including a CMOS image sensor or a CCD image sensor. The camera module may be linked with a defect detection device in which an artificial neural network operates.

[0042] A defect detection device included in a vision inspection system can determine whether a defect exists and the type of defect by analyzing an image acquired from a camera module. The camera module and the defect detection device may communicate remotely or use a communication bus. In one embodiment, the camera module and the defect detection device may be configured as a single module. The image analysis results or inspection status may be transmitted and stored on an internal network or transmitted and stored on a cloud server. Data transmitted to a cloud server may subsequently be provided as quality control data.

[0043] In the present invention, to increase the efficiency and accuracy of vision inspection, the defect detection device may use an artificial neural network. Data learning and inference of the artificial neural network will be described below.

[0045] FIG. 2 shows an artificial neural network model and the operation of each neuron according to one embodiment of the present invention.

[0046] FIG. 2(a) shows an artificial neural network used by a defect detection device according to an embodiment of the present invention. The artificial neural network includes neurons, and in FIG. 3(a), each circle represents a neuron (or node). A set of multiple neurons is called a layer. A layer that receives input data is referred to as an input layer, and a layer that outputs the operation result of the artificial neural network is referred to as an output layer.

[0047] Figure 2(b) illustrates the operation of a neuron in an artificial neural network. The neuron applies weights (w_0, w_1) to inputs (i_0, i_1) and performs an activation function (f(x)) operation to provide an output. The activation function can be a function whose value increases rapidly when stimulation exceeding a certain level is applied, similar to a real nerve cell. Since using a linear function may reduce analytical capability, a non-linear function may be used. The activation function does not use the sum of the input signals as is, but can convert them into an output signal. In other words, the activation function determines whether to activate the inputs to which weights have been applied.

[0049] FIG. 3 illustrates the training and inference of an artificial neural network according to one embodiment of the present invention.

[0050] The top of FIG. 3 shows the learning of an artificial neural network used by a defect detection device according to one embodiment.

[0051] A defect detection device according to one embodiment trains an artificial neural network to perform a specific task. Training may refer to supervised learning. That is, the defect detection device may configure an artificial neural network and train the neural network to perform a specific task. For example, a task of determining whether an image is a dog or a cat may be performed. The defect detection device repeatedly performs input data processing, comparison of output data with correct data, and weight adjustment operations by using multiple images and correct data (dog or cat) for the multiple images. The neural network trained in this way is ready to perform the task.

[0052] The bottom of Fig. 3 shows the inference of an artificial neural network.

[0053] A trained artificial neural network corresponds to a weighted database. An artificial neural network sufficiently trained to identify cat images can infer and identify a cat image when an arbitrary animal image is input.

[0055] FIG. 4 shows an example of an image used for learning and inference of an artificial neural network according to an embodiment of the present invention, FIG. 5 shows an example of an image used for learning and inference of an artificial neural network according to an embodiment of the present invention, and FIG. 6 shows an example of an image used for learning and inference of an artificial neural network according to an embodiment of the present invention.

[0056] In the case of valve seats, it was difficult to apply them to manufacturing processes because defect detection using existing image processing technology is not easy due to reasons such as uniform color and differences in visibility depending on lighting, as well as the variety of defect types and patterns. The present invention significantly improves defect detection performance by training an artificial neural network with a training data set consisting of previously detected defect sample data and normal sample data. The defect detection device can provide results with almost no error when detecting defects using the trained artificial neural network after training the artificial neural network with defect sample data and normal sample data. The defect sample data and normal sample data will be described later with reference to FIGS. 7 to 11.

[0057] The present invention not only simply detects defects but also classifies the types of defects to train an artificial neural network and determines the types of defects. In addition, since the artificial neural network is continuously trained using data from the use of the defect detection device and the detection process, more precise and accurate defect detection becomes possible as use progresses.

[0058] Referring to FIG. 4, a defect detection device according to one embodiment may acquire an image (100) from a camera module. For example, the image (100) may be an image of a valve seat of an engine. In the image (100), the valve seat may include ring-shaped surfaces A, B, and C between area (120) and area (130). At this time, surfaces A, B, and C may be referred to as A, B, and C, respectively, from the outermost to the inner side. Surfaces A, B, and C may be concentric circles sharing the same center point (110). Surfaces A, B, and C are arranged sequentially toward the center point (110), and the inner side of surface A and the outer side of surface B are connected, and the inner side of surface B and the inner side of surface C are interconnected. At this time, the valve seat may be referred to as a product.

[0059] The defect detection device may perform pre-processing on an image (100) for at least one of learning and inference of an artificial neural network. For example, pre-processing may include image masking, rotation, and segment extraction. The defect detection device enables efficient learning and / or inference of the artificial neural network through pre-processing.

[0060] The defect detection device can perform masking processing on areas excluding sides A, B, and C of the valve seat. Masking processing may refer to a process of filling with a black image. To set the inspection target area, the defect detection device can detect an area (120) corresponding to the outer diameter of side A and an area (130) corresponding to the inner side of side C. Areas (120) and (130) may share a center point (110). In one embodiment, the defect detection device can detect areas (120 and 130) based on the center point (110). In one embodiment, the defect detection device can detect areas (120 and 130) using an artificial neural network. In this case, the defect detection device can perform masking based on areas (120 and 130) at different locations for different images. In one embodiment, the defect detection device can set the areas (120 and 130) to be fixed. In this case, the defect detection device can perform masking based on the same location area (120 and 130) for different images.

[0061] The defect detection device can perform masking processing outside the area (120) and inside the area (130). The defect detection device can obtain the image (200) of FIG. 5 by performing masking processing on the image (100) of FIG. 4. The image (200) of FIG. 5 may include a center point (110) and areas (120 and 130) in the same way as the image (100) of FIG. 4.

[0062] The defect detection device can extract segments while rotating the image (200). A segment may refer to a sub-image. The defect detection device can acquire multiple segments corresponding to the area (210) while rotating the image (200) at a predetermined angle while keeping the area (210) fixed. The area (210) can be configured to inspect the largest area of ​​the valve seat. For example, the defect detection device can configure the area (210) to be square, include side A, side B, and side C of the valve seat, have the outer diameter (outer arc) of side B pass through the two vertices of the area (210), and have the angle (intermediate angle, in-between angle) formed by two straight lines connecting the two vertices from the center point (110) be greater than or equal to the predetermined angle. However, it is not necessarily limited to this, and the inner diameter (inner arc) of side B may be set to pass through the two vertices of the region (210), or the arc (inner diameter or outer diameter) of side A or side C may pass through the two vertices of the region (210). In one embodiment, the defect detection device may set the region (210) to be an angle.

[0063] In FIG. 5, the defect detection device is described as rotating the image (200) clockwise, but it is not necessarily limited to this and can also be implemented by rotating the image (200) counterclockwise to acquire multiple segments.

[0064] In one embodiment, the predetermined angle may be 18 degrees. The defect detection device may obtain the segment (300) of FIG. 6 by rotating the image (200) of FIG. 5 around the center point (110) and extracting the segment corresponding to the region (210). The segments (300) may include first to twentieth segments (300_1, 300_2, 300_3, 300_4, ... 300_20). The first to twentieth segments (300_1, 300_2, 300_3, 300_4, ... 300_20) may include the same or similar size and shape at the same location, and if the segment includes a defective region, there may be a part that differs from the normal region. An artificial neural network can learn and infer the characteristics of the defect based on these differing parts. The normal region may represent a good product.

[0065] A defect detection device can induce learning toward quality by training a classification model with defects having various patterns and shapes compared to good products. The classification model may include width inspection (e.g., width of side B) and external defects (defect types such as thinning, breakage, and contamination). Learning toward quality refers to a learning method that determines a product as defective when it has a pattern or shape different from a good product, enabling the identification of newly occurring defects even with a small amount of defect data.

[0066] The defect detection device may use segments (300) obtained from a single image (100) as a training data set for training an artificial neural network, or as an input data set for inference of the artificial neural network. By using segments (300) rather than the image (100) for training and inference, the defect detection device enables more sophisticated and detailed defect detection. For example, the defect detection device can detect features of 10 pixels or more out of 50,176 pixels in each segment (300) with an image size of 224*224. 10 pixels may correspond to an actual size of 0.15 mm. This is merely one example, and depending on the camera and lighting, the capture and verification may be performed more finely or more clearly.

[0068] Figure 7 shows examples of images of a normal type valve seat, and Figures 8 to 11 show examples of images of a defective type valve seat.

[0069] The above valve seat may include a surface A, a surface B, and a surface C. The outer diameter of the valve seat may be defined as surface A, the inner diameter as surface C, and the area between surface A and surface C as surface B. A defect detection device according to one embodiment may detect a defect in surface B of the valve seat, but is not necessarily limited thereto, and may also detect a defect in at least one of surface A and surface C.

[0070] A defect detection device according to one embodiment can train an artificial neural network using normal type valve seat data as shown in FIG. 7 and defective type valve seat data as shown in FIG. 8 to 11. The defect detection device can train an artificial neural network using supervised learning that uses a plurality of images and corresponding ground truth data. In one embodiment, the ground truth data may indicate normal or defective. In one embodiment, the ground truth data may include defect types. For example, defect types may include types such as B-side damage, B-side width size defect, machining tool damage, and incomplete machining.

[0071] Referring to FIG. 7, the normal type of valve seat determined by a defect detection device according to an embodiment of the present invention is illustrated.

[0072] For a normal type valve seat, the width of the B-side may be constant. For example, for a normal type valve seat, the width of the B-side may not deviate by 10% (0.15mm) from the standard 1.5mm. In addition, the normal type valve seat may not show signs of damage on the B-side, and may not have defects such as broken machining tools or unmachined parts.

[0073] Referring to FIGS. 8 to 11, the defect type of the valve seat determined by the defect detection device according to an embodiment of the present invention is illustrated.

[0074] Figure 8 shows examples of valve seats with damaged B-side. Valve seat damage can occur due to various types of defects, such as material loss, breakage, or contamination. Figure 9 shows examples of valve seats with defective B-side width dimensions. An example of defective B-side width dimensions may refer to cases where the B-side width deviates from a critical range. For example, the critical range may be set to 10% (0.15 mm) based on 1.5 mm. Figure 10 shows examples of valve seats with damaged machining tools. Due to the machining tool damage, the boundaries of A-side, B-side, and C-side of the valve seat may be unclear, or the valve seat may be incompletely machined. Figure 11 shows examples of unmachined valve seats. In the case of unmachined valve seats, the mold may remain in a circular state.

[0075] A defect detection device according to an embodiment of the present invention can generate a learning model for each defect type by learning images resulting from each of the defect types. Additionally, if a new type of defect is detected that has not previously been learned by the defect detection device, the defect detection device can classify the defect separately as a new type of defect. The data for the new type of defect can then be used as a training data set for an artificial neural network.

[0076] Meanwhile, the defect detection device of the present invention can identify a defect pattern. A defect pattern refers to a case where a specific type of defect is repeatedly detected. That is, when a specific type of defect appears repeatedly, the defect detection device can recognize this as the occurrence of a defect pattern.

[0077] Defect patterns can occur, in particular, due to issues with cutting devices, cleaning devices, and manufacturing processes. Therefore, defect patterns are reported to the system and can be improved through inspection and corrective actions regarding the relevant devices and processes. If a specific defect occurs but lacks a pattern, the defect detection device identifies it as a circular defect and can tag the detected defect area. Of course, even when a defect pattern is detected, the defect detection device can not only report the pattern but also perform tagging on the detected defect area. By recognizing defect patterns and providing them to the user, the defect detection device enables the improvement of defects that recur for specific reasons and prevents additional defects.

[0078] According to the present invention, accurate and error-free valve seat defect detection is made possible by using an artificial neural network. Furthermore, by using at least one camera and an artificial neural network, defect detection becomes possible solely through camera installation without configuring specific conditions such as a darkened environment. Additionally, by recognizing and analyzing defect patterns, it is possible to respond quickly when a defect occurs to prevent further defects. Moreover, the artificial neural network can be further trained through the accumulation of defect data, and various process and quality improvements can be achieved through the utilization of big data.

[0080] FIG. 12 shows a flowchart of a learning and inference method of an artificial neural network according to one embodiment.

[0081] Referring to FIG. 12, a defect detection device according to one embodiment can generate a vision-based deep learning defect learning model (S1110). The defect detection device can generate a vision-based deep learning defect learning model using the artificial neural network described above and a training data set consisting of defective sample data and normal sample data. That is, the defect detection device can train the artificial neural network in step (S1110). The configuration for generating the vision-based deep learning defect learning model will be explained in more detail with reference to FIG. 13.

[0082] The defect detection device can detect vision-based deep learning defects using a learning model (S1120). That is, the defect detection device can detect defects using a trained artificial neural network. A configuration for detecting vision-based deep learning defects using a learning model will be explained in more detail with reference to FIG. 14.

[0083] The defect detection device can determine whether a defect has been detected (S1130). At this time, the defect detection device can make a final defect determination if a predetermined condition is satisfied. In one embodiment, the defect detection device can make a final defect determination if it detects a defect through cross-verification. For example, among the segments, the first segment and the second segment may contain the same defective part. Accordingly, if the defect detection device detects a defect in the first segment and a defect in the second segment, it can make a final defect determination. In addition, in one embodiment, the defect detection device can make a final defect determination if it detects a defect n times consecutively. n can be an integer greater than 1. In the case of a valve seat, it is a common pattern for defects to occur consecutively. Accordingly, if the defect detection device detects a defect in consecutive segments, it can make a final defect determination. The defect detection device can lower the false defect rate and increase verification power by using cross-inspection and / or continuous inspection. If no defect is detected, it can perform a step (S1160) of determining whether to terminate the inspection.

[0084] When a defect is detected, the defect detection device can report the defective product (S1140). The defect detection device can obtain information about the defective valve seat from the segment where the defect was detected. The information may include an identifier of the valve seat, such as a serial number, the date and time of manufacture, the place of manufacture, etc. The defect detection device can transmit the information about the defective valve seat to a cloud server or output it to a display. The configuration for reporting the defective product will be explained in more detail with reference to FIG. 15.

[0085] The defect detection device can update a learning model using a defect image (S1150). For example, the defect detection device can update a learning model of an artificial neural network using a defect image as defect sample data. Additionally, the defect detection device can train the artificial neural network on the defect type of the defect image. A configuration for updating a learning model using a defect image will be described in more detail with reference to FIG. 16.

[0086] The defect detection device can determine whether the inspection has been completed (S1160). For example, if there are additional images or segments to inspect, the defect detection device can perform a defect detection step (S1120). Once the inspection is completed up to the last image, the defect detection device can terminate the learning and inference of the artificial neural network.

[0088] FIG. 13 shows a flowchart of an artificial neural network learning method according to one embodiment.

[0089] Referring to FIG. 13, a defect detection device according to one embodiment can set a masking area in a learning target product image (S1210). The masking area may be an area excluding surface A, surface B, and surface C of the valve seat in the image. For example, the defect detection device can detect the center point of the product. The defect detection device can set the inner diameter (surface C) at the center point of the product. The defect detection device can set the outer diameter (surface A) at the center point of the product. The defect detection device can mask the area excluding surface A, surface B, and surface C (the area outside the inner diameter and inside the outer diameter) in the image (i.e., the area inside the inner diameter and outside the outer diameter) with black. The center point, outer diameter, and inner diameter are indicated by reference numerals 110, 120, and 130 in FIG. 4 and FIG. 5, respectively.

[0090] The defect detection device can set an inspection target segment in a masked product image (S1220). The inspection target segment can be set to inspect the widest area of ​​the valve seat. For example, the outer diameter of side B can be set to pass through the two vertices of the segment. The inspection target segment is indicated by reference numeral 210 in FIG. 5.

[0091] The defect detection device can acquire multiple segments by rotating the image while maintaining the segment to be inspected (S1230). The defect detection device can rotate the image at a predetermined angle. For example, the predetermined angle is 18 degrees, and the defect detection device can acquire 20 segments.

[0092] The defect detection device can perform supervised learning (S1240). The defect detection device can train an artificial neural network with a training data set consisting of defective sample data and normal sample data. The contents of FIGS. 7 to 11 may be applied to the defective sample data and normal sample data.

[0093] In one embodiment, the defect detection device can learn an algorithm for determining whether a valve seat is normal or defective. For example, the defect detection device can learn defect patterns from defective sample data that is different from normal sample data.

[0094] In one embodiment, the defect detection device can learn an algorithm to determine the defect type of the valve seat. The defect type may include width surface size defects (e.g., width surface inspection in the range of +10%, -10% for a width of 1.5 mm on side B), appearance defects (e.g., defects, cracks, contamination, etc. on side A, side B, or side C).

[0095] The defect detection device can determine whether the learning has been completed (S1250). For example, if there are additional images to learn, the defect detection device can perform the step of setting a masking area (S1210). When the learning is completed up to the last image, the defect detection device can terminate the learning of the artificial neural network.

[0097] FIG. 14 shows a flowchart of a valve seat defect detection method according to one embodiment.

[0098] Referring to FIG. 14, a defect detection device according to one embodiment can set a masking area in an image of a product to be inspected (S1310). The masking area may be an area excluding sides A, B, and C of the valve seat in the image. The defect detection device may set the masking area using an artificial neural network. That is, the defect detection device may train an artificial neural network for the masking area setting of step (S1110) of FIG. 12 and step (S1210) of FIG. 13, and set the masking area from the image using the trained artificial neural network. In another embodiment, the defect detection device may set the masking area by detecting a center point, an inner diameter, and an outer diameter. Accordingly, the defect detection device may apply a black masking process to the area excluding sides A, B, and C (the area outside the inner diameter and inside the outer diameter) in the image (i.e., the area inside the inner diameter and outside the outer diameter). The center point, outer diameter, and inner diameter are indicated by reference numbers 110, 120, and 130 in FIGS. 4 and 5, respectively.

[0099] The defect detection device can set an inspection target segment in a masked product image (S1320). The inspection target segment can be set to inspect the widest area of ​​the valve seat. For example, the outer diameter of side B can be set to pass through the two vertices of the segment. The inspection target segment is indicated by reference numeral 210 in FIG. 5.

[0100] The defect detection device can acquire multiple segments by rotating the image while maintaining the segment to be inspected (S1330). The defect detection device can rotate the image at a predetermined angle. For example, the predetermined angle is 18 degrees, and the defect detection device can acquire 20 segments.

[0101] The defect detection device can perform defect inspection on an image of a product to be inspected (S1340). That is, the defect detection device can perform defect inspection on multiple segments obtained from the image of the product to be inspected. The defect detection device can inspect multiple segments using a trained artificial neural network.

[0102] In one embodiment, the defect detection device can determine whether a valve seat is normal or defective from an image of a product to be inspected using a trained artificial neural network. The trained artificial neural network can perform the determination through inference.

[0103] In one embodiment, the defect detection device can determine the type of defect of the valve seat from an image of a product to be inspected using a trained artificial neural network. The defect types may include width surface size defects (e.g., width surface inspection within a range of +10% and -10% for a width of 1.5 mm on side B), appearance defects (e.g., defects, cracks, contamination, etc. on side A, side B, or side C).

[0104] The defect detection device can determine whether the inspection has been completed (S1350). The unit of inspection may be one image (e.g., 20 segments). For example, if there are additional images to inspect, the defect detection device may perform the step of setting a masking area (S1310). Once the inspection is completed up to the last image, the defect detection device may terminate the inference of the artificial neural network.

[0106] FIG. 15 shows a flowchart of a defect target product report method according to one embodiment.

[0107] Referring to FIG. 15, a defect detection device according to one embodiment can acquire defect data (S1410). The defect data may include sub-image (segment) data where a defect occurred, the order of the corresponding sub-image in the image, and defect category (defect type) data. The defect detection device can track the location of the defect based on the order. The defect detection device can output the location of the defect as two-dimensional coordinate information in the x and y directions. The defect type may be a width size defect, an external shape defect, an unclassified defect, etc.

[0108] The defect detection device can add defect data to the defective product information (S1420). That is, the defect detection device can output information on whether the valve seat is normal or defective, and if defective, the type of defect and the location of the defect, etc.

[0109] In one embodiment, the defect detection device can determine whether the valve seat is a recyclable product based on defect data. For example, if k or fewer sub-image data contain defects and the defect category is an external defect, the defect detection device can determine the valve seat as a recyclable product. k is an integer greater than or equal to 1 and may be set to 1 or 2 depending on the embodiment.

[0111] FIG. 16 shows a flowchart of a learning model update method according to one embodiment.

[0112] Referring to FIG. 16, a defect detection device according to one embodiment can acquire defect data (S1510). The defect data may be applied in the same way as described with reference to step (S1410) of FIG. 15.

[0113] The defect detection device can determine whether there is an unclassified category (S1520). That is, the defect detection device can determine whether there is a new type of defect that is not classified as a width-size defect or an external appearance defect.

[0114] If there are no unclassified categories, the defect detection device can update the category-specific learning model (S1530). For example, the defect detection device can update the category-specific learning model using data classified as at least one of width-size defects and appearance defects. The updated learning model may be an advanced model.

[0115] If there are unclassified categories, the defect detection device can exclude the unclassified categories and update the category-specific learning model (S1540). For example, the defect detection device can update the category-specific learning model using data classified as at least one of width surface size defects and appearance defects.

[0116] The defect detection device can report defect data of an unclassified category (S1550). The defect detection device can output information regarding the location of defects in the defective valve seat, etc. In one embodiment, when correct answer data for an unclassified category is input, the defect detection device can train an artificial neural network based on the defect data of the unclassified category and the correct answer data.

[0117] The defect detection device can determine whether to apply an update model to an existing learning model (S1560). For example, the defect detection device can determine whether to apply an update model to an existing learning model based on a command from the outside.

[0118] If the defect detection device determines that the update model is not to be applied to the existing learning model, it can proceed to step (S1510). The defect detection device can acquire new defect data (S1510).

[0119] If the defect detection device determines that it is applying an update model to an existing learning model, it can apply the update model to the existing learning model (S1570). The defect detection device can apply the update model to the existing learning model in real time or at preset time intervals based on a command from an external source.

[0121] The embodiments described above are combinations of the components and features of the present invention in a specific form. Each component or feature should be considered optional unless otherwise explicitly stated. Each component or feature may be implemented in a form not combined with other components or features. Additionally, it is possible to construct embodiments of the present invention by combining some components and / or features. The order of operations described in the embodiments of the present invention may be changed. Some components or features of one embodiment may be included in another embodiment, or may be replaced with corresponding components or features of another embodiment. It is obvious that embodiments may be constructed by combining claims that do not have an explicit citation relationship in the claims, or that new claims may be included by amendment after filing.

[0122] Embodiments according to the present invention may be implemented by various means, for example, hardware, firmware, software, or a combination thereof. In the case of implementation by hardware, one embodiment of the present invention may be implemented by one or more ASICs (application specific integrated circuits), DSPs (digital signal processors), DSPDs (digital signal processing devices), PLDs (programmable logic devices), FPGAs (field programmable gate arrays), processors, controllers, microcontrollers, microprocessors, etc.

[0123] In the case of implementation by firmware or software, an embodiment of the present invention may be implemented in the form of a module, procedure, function, etc., that performs the functions or operations described above. The software code may be stored in memory and executed by a processor. The memory may be located inside or outside the processor and may exchange data with the processor by various means already known.

[0124] It is obvious to those skilled in the art that the present invention may be embodied in other specific forms without departing from the essential features of the invention. Accordingly, the foregoing detailed description should not be interpreted restrictively in all respects but should be considered exemplary. The scope of the invention shall be determined by a reasonable interpretation of the appended claims, and all modifications within the equivalent scope of the invention are included within the scope of the invention.

Claims

Claim 1 A step of acquiring a first image of an article having a ring-shaped mold; a step of acquiring a second image by performing preprocessing on the first image; a step of determining an inspection area in the second image; a step of acquiring a plurality of segments corresponding to the inspection area while rotating the second image by a predetermined angle; and a step of acquiring correct or defective correct data for each of the plurality of segments. A method for training an artificial neural network for detecting defects in an article, comprising the step of training an artificial neural network using a training data set composed of the plurality of segments and correct data, wherein the step of determining an inspection area in the second image includes determining a square as the inspection area that includes all of the outer surface A of the article having the ring-shaped mold, the area B between surface A and surface C, and the inner surface C of the article having the ring-shaped mold in the second image, and the step of determining a square as the inspection area that includes all of the outer surface A of the article having the ring-shaped mold, the area B between surface A and surface C, and the inner surface C of the article having the ring-shaped mold in the second image includes determining the inspection area such that the outer diameter of surface B passes through two vertices of the inspection area. Claim 2 An artificial neural network learning method for detecting product defects according to claim 1, wherein the step of obtaining a second image by performing preprocessing on the first image includes the step of performing masking on the first image. Claim 3 An artificial neural network learning method for detecting article defects according to claim 2, wherein the step of performing masking on the first image includes the step of performing masking on the outer side of surface A, which is the outer surface of the article having the ring-shaped mold, and the inner side of surface C, which is the inner surface of the article having the ring-shaped mold, in the first image. Claim 4 In claim 3, the step of performing masking on the outer side of side A, which is the outer side of the article having the ring-shaped mold in the first image, and the inner side of side C, which is the inner side of the article having the ring-shaped mold, comprises: a step of obtaining a center point of the article having the ring-shaped mold in the first image; a step of obtaining an outer diameter of side A, which is the outer side of the article having the ring-shaped mold, based on the center point; a step of obtaining an inner diameter of side C, which is the inner side of the article having the ring-shaped mold, based on the center point; and a step of performing masking on the outer region of the outer diameter of side A and the inner region of the inner diameter of side C, an artificial neural network learning method for article defect detection. Claim 5 delete Claim 6 delete Claim 7 An artificial neural network learning method for detecting defects in an article, wherein, in claim 1, the step of determining the inspection area such that the outer diameter of the B-side passes through the two vertices of the inspection area includes the step of determining the inspection area such that the angle formed by two straight lines connecting the center point of the article having the ring-shaped mold to the two vertices is greater than or equal to the predetermined angle. Claim 8 An artificial neural network learning method for detecting product defects according to claim 1, wherein the predetermined angle is 18 degrees and the plurality of segments is 20. Claim 9 A step of acquiring a first image of an article having a ring-shaped mold; a step of acquiring a second image by performing preprocessing on the first image; a step of determining an inspection area in the second image; a step of acquiring a plurality of segments corresponding to the inspection area while rotating the second image by a predetermined angle; and a step of detecting defects of the article having the ring-shaped mold from the plurality of segments using a trained artificial neural network. A method for detecting defects in an article, comprising the step of generating and outputting defect data for a detected defect, wherein the step of determining an inspection area in the second image includes determining a square as the inspection area that includes all of surface A, which is the outer surface of the article having the ring-shaped mold in the second image, surface B, which is the area between surface A and surface C, and surface C, which is the inner surface of the article having the ring-shaped mold in the second image, and the step of determining a square as the inspection area that includes all of surface A, which is the outer surface of the article having the ring-shaped mold in the second image, surface B, which is the area between surface A and surface C, and surface C, which is the inner surface of the article having the ring-shaped mold in the second image, wherein the inspection area is determined such that the outer diameter of surface B passes through two vertices of the inspection area. Claim 10 A method for detecting defects in an article according to claim 9, wherein the step of obtaining a second image by performing preprocessing on the first image includes the step of performing masking on the first image. Claim 11 A method for detecting article defects according to claim 10, wherein the step of performing masking on the first image comprises the step of performing masking on the outer side of surface A, which is the outer surface of the article having the ring-shaped mold in the first image, and the inner side of surface C, which is the inner surface of the article having the ring-shaped mold. Claim 12 In claim 11, the step of performing masking on the outer side of side A, which is the outer side of the article having the ring-shaped mold in the first image, and on the inner side of side C, which is the inner side of the article having the ring-shaped mold, comprises: a step of obtaining a center point of the article having the ring-shaped mold in the first image; a step of obtaining an outer diameter of side A, which is the outer side of the article having the ring-shaped mold, based on the center point; a step of obtaining an inner diameter of side C, which is the inner side of the article having the ring-shaped mold, based on the center point; and a step of performing masking on the outer region of the outer diameter of side A and the inner region of the inner diameter of side C, a method for detecting article defects. Claim 13 delete Claim 14 delete Claim 15 A method for detecting defects in an article according to claim 9, wherein the step of determining the inspection area such that the outer diameter of the B-side passes through the two vertices of the inspection area comprises the step of determining the inspection area such that the angle formed by two straight lines connecting the center point of the article having the ring-shaped mold to the two vertices is greater than or equal to the predetermined angle. Claim 16 A method for detecting product defects according to claim 9, wherein the predetermined angle is 18 degrees and the plurality of segments is 20. Claim 17 A method for detecting article defects according to claim 9, wherein the defect data for the detected defect includes data for a segment containing the defect, the order of the segment containing the defect in the second image, and defect type data. Claim 18 In claim 17, the above defect type data includes width and size defects, appearance defects, and unclassified defects, in a method for detecting defects in articles. Claim 19 A method for detecting defects in an article having a ring-shaped mold from a plurality of segments using the learned artificial neural network according to claim 9, wherein the learned artificial neural network determines the article having the ring-shaped mold as normal when it detects a defect in one of the plurality of segments, and determines the article having the ring-shaped mold as defective when it detects a defect in n consecutive segments among the plurality of segments, wherein n is an integer greater than 1. Claim 20 A device for detecting defective articles, comprising a processor and a memory connected to the processor, wherein the memory is configured to store a program and the processor is configured to execute the program, and when the program is executed, the steps of the method of any one of claims 1 to 4, 7 to 12, and 15 to 19 are implemented.

Images