WO2023282043A1 - 検査方法、分類方法、管理方法、鋼材の製造方法、学習モデルの生成方法、学習モデル、検査装置及び鋼材の製造設備 - Google Patents
検査方法、分類方法、管理方法、鋼材の製造方法、学習モデルの生成方法、学習モデル、検査装置及び鋼材の製造設備 Download PDFInfo
- Publication number
- WO2023282043A1 WO2023282043A1 PCT/JP2022/024605 JP2022024605W WO2023282043A1 WO 2023282043 A1 WO2023282043 A1 WO 2023282043A1 JP 2022024605 W JP2022024605 W JP 2022024605W WO 2023282043 A1 WO2023282043 A1 WO 2023282043A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- defect
- inspection
- defect determination
- determination
- image
- Prior art date
Links
- 238000007689 inspection Methods 0.000 title claims abstract description 105
- 238000000034 method Methods 0.000 title claims abstract description 80
- 229910000831 Steel Inorganic materials 0.000 title claims abstract description 47
- 239000010959 steel Substances 0.000 title claims abstract description 47
- 238000004519 manufacturing process Methods 0.000 title claims abstract description 43
- 239000000463 material Substances 0.000 title claims abstract description 18
- 238000007726 management method Methods 0.000 title claims abstract description 16
- 238000012549 training Methods 0.000 title abstract description 14
- 230000007547 defect Effects 0.000 claims abstract description 276
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 28
- 238000001514 detection method Methods 0.000 claims abstract description 12
- 238000003384 imaging method Methods 0.000 claims abstract description 10
- 238000010801 machine learning Methods 0.000 claims description 32
- 230000008685 targeting Effects 0.000 claims description 5
- 238000012545 processing Methods 0.000 abstract description 31
- 238000000605 extraction Methods 0.000 abstract description 9
- 238000012216 screening Methods 0.000 abstract description 6
- 230000006872 improvement Effects 0.000 abstract description 2
- 230000009467 reduction Effects 0.000 abstract description 2
- 230000008569 process Effects 0.000 description 19
- 238000011156 evaluation Methods 0.000 description 9
- 238000007781 pre-processing Methods 0.000 description 6
- 238000003066 decision tree Methods 0.000 description 5
- 238000013135 deep learning Methods 0.000 description 5
- 238000010586 diagram Methods 0.000 description 5
- 230000006870 function Effects 0.000 description 5
- 230000002950 deficient Effects 0.000 description 4
- 238000004364 calculation method Methods 0.000 description 3
- 238000002790 cross-validation Methods 0.000 description 3
- 238000012986 modification Methods 0.000 description 3
- 230000004048 modification Effects 0.000 description 3
- 238000013461 design Methods 0.000 description 2
- 230000001627 detrimental effect Effects 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 239000000284 extract Substances 0.000 description 2
- 238000001914 filtration Methods 0.000 description 2
- 238000002372 labelling Methods 0.000 description 2
- 238000012417 linear regression Methods 0.000 description 2
- 238000007477 logistic regression Methods 0.000 description 2
- 230000003287 optical effect Effects 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 238000013138 pruning Methods 0.000 description 2
- 238000007637 random forest analysis Methods 0.000 description 2
- 238000005096 rolling process Methods 0.000 description 2
- 238000003860 storage Methods 0.000 description 2
- 238000012706 support-vector machine Methods 0.000 description 2
- 230000001629 suppression Effects 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 238000012795 verification Methods 0.000 description 2
- 230000002411 adverse Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 238000001816 cooling Methods 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 230000003111 delayed effect Effects 0.000 description 1
- 230000004069 differentiation Effects 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 238000011176 pooling Methods 0.000 description 1
- 238000003908 quality control method Methods 0.000 description 1
- 238000013526 transfer learning Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
- G01N21/84—Systems specially adapted for particular applications
- G01N21/88—Investigating the presence of flaws or contamination
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
- G01N21/84—Systems specially adapted for particular applications
- G01N21/88—Investigating the presence of flaws or contamination
- G01N21/8851—Scan or image signal processing specially adapted therefor, e.g. for scan signal adjustment, for detecting different kinds of defects, for compensating for structures, markings, edges
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0004—Industrial image inspection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0004—Industrial image inspection
- G06T7/001—Industrial image inspection using an image reference approach
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
- G01N21/84—Systems specially adapted for particular applications
- G01N21/88—Investigating the presence of flaws or contamination
- G01N21/8851—Scan or image signal processing specially adapted therefor, e.g. for scan signal adjustment, for detecting different kinds of defects, for compensating for structures, markings, edges
- G01N2021/8854—Grading and classifying of flaws
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
- G01N21/84—Systems specially adapted for particular applications
- G01N21/88—Investigating the presence of flaws or contamination
- G01N21/8851—Scan or image signal processing specially adapted therefor, e.g. for scan signal adjustment, for detecting different kinds of defects, for compensating for structures, markings, edges
- G01N2021/8883—Scan or image signal processing specially adapted therefor, e.g. for scan signal adjustment, for detecting different kinds of defects, for compensating for structures, markings, edges involving the calculation of gauges, generating models
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
- G01N21/84—Systems specially adapted for particular applications
- G01N21/88—Investigating the presence of flaws or contamination
- G01N21/8851—Scan or image signal processing specially adapted therefor, e.g. for scan signal adjustment, for detecting different kinds of defects, for compensating for structures, markings, edges
- G01N2021/8887—Scan or image signal processing specially adapted therefor, e.g. for scan signal adjustment, for detecting different kinds of defects, for compensating for structures, markings, edges based on image processing techniques
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N2201/00—Features of devices classified in G01N21/00
- G01N2201/12—Circuits of general importance; Signal processing
- G01N2201/129—Using chemometrical methods
- G01N2201/1296—Using chemometrical methods using neural networks
Definitions
- This disclosure relates to inspection methods, classification methods, management methods, steel manufacturing methods, learning model generation methods, learning models, inspection devices, and steel manufacturing equipment.
- Patent Literature 1 discloses a surface inspection device that optically detects flaws on a surface using feature amounts.
- the feature quantity includes, for example, a feature quantity based on the size of the flaw, a feature quantity based on the type of flaw, a feature quantity based on the grade of the flaw, and the like.
- convolutional neural networks which have been increasingly used for image inspection in recent years, have higher discrimination performance than general machine learning methods.
- General machine learning methods are machine learning methods that use features, such as regression models, decision tree models, random forests, Bayesian inference models, Gaussian mixture models, and boosting methods of these machine learning models. be.
- Regression models are, for example, linear regression, logistic regression, multiple regression, support vector machines, nonlinear kernels, and the like.
- Patent Document 1 has discrimination performance that is inferior to that of a convolutional neural network. generate more. Such non-detection and over-detection lead to the outflow of defective products, and reduce the efficiency of the entire process because re-judgment is required in the post-inspection process as a countermeasure. Therefore, there is a need to reduce such underdetections and overdetections.
- the convolutional neural network has high discrimination performance, so it is possible to reduce the occurrence of undetected and over-detected cases.
- the convolutional neural network has a problem that it requires a large amount of calculation and cannot process a large number of defect candidates at once. For example, when used for the inspection or classification of steel materials continuously flowing through a production line, in order to realize continuous judgment by a convolutional neural network with a large amount of calculation, it is necessary to parallelize expensive processing equipment. there were. Alternatively, it was necessary to take measures such as lowering the production efficiency for inspection.
- the purpose of the present disclosure which has been made in view of such circumstances, is to provide an inspection method, a classification method, a management method, a steel manufacturing method, a learning model generation method, a learning model, and an inspection apparatus that can achieve both an improvement in detection accuracy and a reduction in processing time. and to provide steel manufacturing equipment.
- An inspection method includes An inspection method for detecting surface defects to be inspected, comprising: an imaging step of acquiring an image of the surface of the inspection target; an extracting step of extracting a defect candidate portion from the image; a selection step of selecting the extracted defect candidate part by a first defect determination; and an inspection step of detecting whether the surface defects are harmful or harmless by second defect determination using a convolutional neural network, targeting the defect candidate portions selected by the first defect determination.
- a classification method includes: A classification method for classifying surface defects to be inspected, an imaging step of acquiring an image of the surface of the inspection target; an extracting step of extracting a defect candidate portion from the image; a selection step of selecting the extracted defect candidate part by a first defect determination; and a classification step of classifying at least one of types and grades of the surface defects by second defect determination using a convolutional neural network, targeting the defect candidate portions selected by the first defect determination.
- a management method includes: A management step of classifying the inspection objects based on at least one of the types and grades of the surface defects classified by the above classification method.
- a steel manufacturing method includes: a production step for producing steel; and the inspection step in the above inspection method, In the inspection step, the steel manufactured in the manufacturing step is the inspection target.
- a steel manufacturing method includes: a production step for producing steel; and the management step in the above management method, In the managing step, the steel manufactured in the manufacturing step is subject to the inspection.
- a learning model generation method includes: A learning model generation method used in an inspection method for detecting surface defects of an inspection target, Defect candidate portions selected by the first defect determination for an image to be inspected acquired in advance are used as input performance data, and teacher data is used as output performance data for harmful or harmless results of the input performance data. , generating said learning model by a convolutional neural network.
- a learning model according to an embodiment of the present disclosure includes: It is generated by the learning model generation method described above.
- An inspection device includes: An inspection device for detecting or classifying surface defects to be inspected, an imaging device that acquires an image of the surface of the inspection target; A defect candidate portion is extracted from the image, the extracted defect candidate portion is selected by a first defect determination, and a convolutional neural network is used for the defect candidate portion selected by the first defect determination.
- a steel manufacturing facility includes: a manufacturing facility for manufacturing steel; and the above inspection device, The inspection device inspects the steel material manufactured by the manufacturing facility.
- an inspection method, a classification method, a management method, a steel manufacturing method, a learning model generation method, a learning model, an inspection device, and a steel manufacturing facility that can achieve both improved detection accuracy and shortened processing time are provided. can do.
- FIG. 1 is a diagram showing an example of the processing flow of surface inspection.
- FIG. 2 is a diagram showing an example of the flow of generating a determinator.
- FIG. 3 is a diagram illustrating a configuration example of an inspection apparatus according to one embodiment.
- FIG. 1 shows an example of the processing flow of surface inspection performed by the inspection method according to the present embodiment.
- an image of the surface of the object to be inspected is obtained (step S1, imaging step).
- the object to be inspected is a steel plate, but the object to be inspected is not limited to the steel plate.
- preprocessing is performed on the image (step S2).
- Preprocessing is performed for the purpose of removing disturbances from the obtained image and facilitating the detection of defect signals (signals indicating surface defects).
- Pre-processing may include, for example, inspection region extraction, brightness unevenness correction, frequency filtering, differentiation filtering, and the like.
- a defect candidate portion extraction process for extracting a defect candidate portion from the image is executed (step S3, extraction step).
- the defect candidate portion extraction process is a process of extracting and labeling a portion (defect candidate portion) having a different brightness by comparing with the surrounding portion by threshold processing.
- the defect candidate portion is extracted by threshold processing in the present embodiment, other processing may be used as long as the defect candidate portion can be specified.
- the entire image may be extracted as the defect candidate portion.
- a first defect determination process is executed to select (screen) the extracted defect candidate portions by the first defect determination (step S4, screening step). Unlike the second defect determination described below, the first defect determination does not use a convolutional neural network. By the first defect determination, the defect candidate portion is reduced to such an extent that the second defect determination to be described later can be processed. In this embodiment, in the first defect determination, selection is performed so as to exclude defect candidate portions that are reliably determined to be harmless.
- a feature amount may be extracted from the image, and the defect candidate portion may be selected using the feature amount.
- the first defect determination uses a learning model generated by machine learning using feature values extracted from past images of the surface of the inspection object, as in an embodiment described later, to determine defect candidate portions. You can select. The details of the learning model used in the first defect determination will be described later.
- a second defect determination process for detecting harmful or harmless surface defects by second defect determination using a convolutional neural network targeting the defect candidate portions selected by the first defect determination. is executed (step S5, inspection step).
- a harmless surface defect is a surface defect that does not make an inspection object defective.
- a harmless surface defect includes, for example, a defect detected as a signal whose physical characteristics are different from those of most healthy sites and made a defect candidate, but which actually does not adversely affect quality and is not a defect. Harmless surface defects also include, for example, defects that were small enough to meet quality standards.
- Detrimental surface defects refer to surface defects excluding harmless surface defects. The details of the learning model, which is a convolutional neural network, used in the second defect determination will be described later.
- the second defect determination process may be executed as a process of classifying surface defects to be inspected. That is, instead of the inspection step described above, after the selection step, the second defect determination process may be performed on the defect candidate portions selected by the first defect determination (step S5, classification step). .
- the second defect determination process classifies at least one of the types and grades of surface defects by second defect determination using a convolutional neural network.
- a classification method is then realized that classifies the surface defects to be inspected in more detail than the inspection method (harmful or harmless detection).
- Such a classification method may be applied to a management method including a management step of classifying inspection objects based on at least one of types and grades of detected surface defects of the inspection objects.
- a method of manufacturing a steel material includes a manufacturing step of manufacturing the steel material and the inspection step or management step described above, and the inspection step or management step manages the steel material manufactured in the manufacturing step as an inspection target.
- the manufacturing step of manufacturing the steel material may include known processes.
- the manufacturing step may include a rough rolling step, a finish rolling step, a cooling step, and the like, for example, when the inspection object is a steel plate.
- the result of the second defect determination process is output as defect information (step S6).
- Defect determination may be the detection of harmful or harmless surface defects, or classification of at least one of the type and grade of surface defects.
- the output defect information may be used for quality control of the inspection object in the control method and steel manufacturing method, as described above.
- FIG. 3 is a diagram illustrating a configuration example of an inspection apparatus according to one embodiment.
- the inspection device is controlled by a host system (eg, a process computer that manages the manufacturing process to be inspected).
- the inspection device may be provided as one of the devices constituting the manufacturing facility to be inspected.
- the inspection device may be configured with, for example, a computing device and a camera.
- the inspection device may also include a machine learning device, if desired.
- a camera is an example of an imaging device, and images the surface of an inspection object (a steel plate in the example of FIG. 3) irradiated with light from a light source.
- the arithmetic device executes preprocessing for an image, defect candidate portion extraction processing, first defect determination processing, and second defect determination processing. Also, the arithmetic device may output data relating to these processes to the host system.
- the machine learning device acquires teacher data from the arithmetic device and generates a learning model (second determiner) used in the second defect determination.
- the machine learning device acquires teacher data from the arithmetic device and also generates the learning model (first determiner) used in the first defect determination. do.
- the first determiner may be manually designed if the learning model is not used in the first defect determination.
- the generated first determiner and second determiner are stored in a storage unit accessible by the computing device, and stored by the computing device when the first defect determination and the second defect determination are executed. may be read from the department.
- FIG. 2 is a diagram showing an example of the flow of generating a determinator.
- the processing flow of the same surface inspection as in FIG. 1 is drawn with dotted lines.
- the dotted line processing flow indicates that the process is not executed at the same time as the determinator generation flow, and is written together to show the relationship with steps S11, S12, and S21 of the determinator generation flow.
- the surface inspection processing flow is executed using the generated determinator.
- the first determiner is generated as follows before the second determiner. First, an image is acquired, preprocessing is performed, and defect candidate portions are extracted by threshold processing. These processes correspond to steps S1 to S3 described above and may be executed by the techniques described above.
- the acquired image is an image (past image) of an inspection object in which actual defects have already been measured.
- the extracted defect candidate part and the actual defect are linked (step S11).
- Machine learning or manual design is performed based on the defect candidate portion that has been associated with the actual defect to generate a first determiner (step S12).
- the first determiner (learning model) is a model generated in advance by machine learning using feature amounts extracted from pre-captured images of the surface of an arbitrary inspection object. .
- the feature amount is a feature amount associated with the actual defect, and is used as teacher data.
- a determination method is used that makes it possible to determine all the defect candidate parts after selection in time in the second defect determination. That is, in the first defect determination, a determination method that does not require much calculation time is required. For example, feature amounts such as size, shape, orientation, and gradation are extracted from an image, and defect determination is performed using the feature amounts. In other words, it is preferable to use a determination method other than the convolutional neural network. In the determination using the feature amount, conditional branching may be created manually (manual design in step S12).
- judgments using feature values can be performed using general machine learning methods using feature values, such as regression models, decision tree models, random forests, Bayesian estimation models, mixed Gaussian models, or boosting methods of these models. may be used (machine learning in step S12).
- regression models include, for example, linear regression, logistic regression, multiple regression, support vector machines, nonlinear kernels, and the like.
- the first defect determination need not be a determination using a feature amount.
- the first determiner is designed to determine unwanted defect candidates, eg, harmless surface defects, and exclude unwanted defect candidates from the defect candidates determined in the second defect determination.
- a defect candidate part is selected using the generated first determiner. This process corresponds to step S4 described above.
- a second determiner is generated by a convolutional neural network based on the selected defect candidate portion (step S21). In step S21, the defect candidate portion selected by the first defect determination for the acquired past image is used as input result data, and teacher data is used in which the harmful or harmless result of the input result data is used as output result data.
- a learning model is generated by The second determiner uses a convolutional neural network to determine defects that require further determination after the first defect determination (e.g., harmful defects), and the obtained result is the final determination result. is generated as
- the number of defect candidate parts to be subjected to the second defect determination is set to the threshold of the first determiner so as not to reach the upper limit in consideration of the defect determination processing capability of the second determiner. Or it is preferably adjusted by a parameter.
- the number of defect candidate parts selected by the first defect determination within the range not exceeding the capability of the second defect determination processing, the number of objects to be determined by the convolutional neural network with high discrimination performance is increased. , it is possible to improve the detection accuracy.
- the processing time is not delayed in the second defect determination. , the overall processing time can be shortened.
- the determination time and discrimination performance differ depending on the layer structure, but it is preferable to determine the layer structure in consideration of the permissible determination time and required discrimination performance.
- the judgment machine was generated. As shown in FIG. 3, a camera and a light source were installed in the thin steel sheet production line, and the entire image of the steel sheet surface was captured by the camera. A monochrome line camera was used. A line light source was used as the light source.
- the computing device acquired the above images and performed the necessary processing. Specifically, the computing device performed a process to remove luminance unevenness. Further, the arithmetic unit extracts defect candidate portions by performing threshold processing, labeling processing, and the like after normalizing the image with a frequency filter. From the extracted defect candidate portion, more than 100 feature amounts related to shape, orientation, size, gradation, etc. were calculated by image processing.
- the machine learning device associates the feature quantity of the extracted defect candidate part with the actual defect, attaches a label such as harmful or harmless, and prepares teacher data ("first One teacher data”).
- the machine learning device automatically constructed a first decision machine based on a binary decision tree through machine learning using the first teacher data.
- the computing device can perform the first defect determination in real time by using the first determiner generated by the machine learning device.
- the number of the first training data was 50 data linked as harmful (harmful data) and 400 data linked as harmless (harmless data).
- cross-validation was used to perform optimal pruning to prevent overfitting.
- the training data is randomly divided into 10 parts, 90% of which is used as learning data to construct a binary decision tree decision machine, and 10% is used as evaluation data for decision making. was done.
- evaluation was performed by mutual verification in which all the evaluation data was separated from the learning data.
- an over-learning suppression penalty is given to the number of branches to suppress over-learning, and the over-learning suppression penalty is set so that the evaluation result by cross-validation is the best, thereby achieving optimal pruning.
- the first discriminator was designed to reliably detect harmful data while leaving harmless data to some extent.
- the first classifier was designed to exclude only data that was reliably harmless. Therefore, the first determiner was adjusted so as not to determine harmful data as harmless in the first defect determination. Specifically, the adjustment is such that the first misclassification penalty for misclassifying harmful data as harmless during training is ten times the second misclassification penalty for misclassifying harmless data as harmful. set to size. With this setting, the first determiner was able to correctly determine all harmful data.
- defect candidate parts determined to be harmful in the first defect determination there are those that are linked as harmless in actual defects.
- Such harmless data is preferably correctly determined as harmless in the second defect determination.
- a second classifier was generated using deep learning (specifically, a convolutional neural network) to have better discrimination performance than the first classifier. Data different from the first training data is prepared, and the data of the defect candidate portion judged to be harmful by the first defect judgment is linked with the actual defect, and the training data (the second judgment used as "second training data”) to generate machines.
- the number of second training data was 58 harmful data and 390 harmless data.
- random division into learning data and verification data was performed so that all data could be evaluated.
- the second training data was normalized to 256 ⁇ 256 size by resizing.
- For the second training data in order to increase the number of data for learning, after resizing, vertically, horizontally, and vertically and horizontally inverted images are added, and each image is -10°, -5°, +5°. and additional images with +10° rotation. After adding the rotated image, the final second training data was obtained by cropping the 224 ⁇ 224 size from the center of the image in order to make all images have the same size. By such inversion and rotation processing, the number of second teacher data could be expanded twenty times.
- ResNet152 was used as the network structure for deep learning. GAP (Global Average Pooling) was performed with ResNet 152 in order to learn to discriminate between harmful data and harmless data into two classes. After that, remove the fully connected layer for classifying into 1000 classes, instead add 2 fully connected layers that output 128 nodes and a fully connected layer for classifying into 2 classes, and the network structure is constructed. was done. For the deep learning framework, tensorflow (version 2.0.0) using the ImageNet dataset was used.
- the arithmetic unit can perform the second defect determination in real time.
- the machine learning device was a device (another computer) provided with a processor different from the arithmetic device.
- the machine learning device communicated with the computing device and was able to send and receive the necessary data.
- the defect information that is finally determined to be harmful in the above two-stage defect determination is output to the host system via communication and used to determine whether the steel plate needs maintenance and whether it can be shipped.
- harmful surface defects may be determined by type and harmfulness (a value indicating the degree of harmfulness).
- the host system may take optimal measures according to the type and severity of harmful surface defects.
- the surface defect is a flaw as an example, but is not limited to a flaw.
- the surface inspection of the steel plate was performed according to the processing flow shown in FIG.
- 98% of the defect candidate portions determined to be harmless in the defect candidate portion extraction process are determined to be harmless in the first defect determination, and 98% are determined to be harmless in the first defect determination.
- 99.97% of the defects were judged to be harmless in combination with the first defect judgment.
- the inspection method and inspection apparatus can both improve the detection accuracy and shorten the processing time by the above steps and configuration.
- the inspection method and inspection apparatus according to the present embodiment discriminate the detected defect candidates by a conventional machine learning method capable of high-speed processing using feature values etc. as the first defect determination. Limit the number. After that, since the convolutional neural network is applied for the second defect determination, it becomes possible to apply the convolutional neural network with high discrimination performance even to a large number of defect candidates. At this time, since the number of defect candidates limited by the first defect determination can be adjusted to the number that can be processed by the second defect determination or less, the overall processing time can be shortened.
- the inspection method and inspection apparatus according to the present embodiment can individually adjust the threshold value, parameter, layer structure, etc. of each determination machine in the two-stage defect determination. There is also an effect that the degree of freedom of adjustment increases compared to the case of having one.
- the purpose of the first defect determination is screening. Therefore, the defect candidate portion determined to be harmful in terms of the actual defect is reliably determined to be harmful, and the defect candidate portion determined to be harmless in terms of the actual defect is allowed to be erroneously determined as harmful to some extent. , preferably adjusted.
- the first defect determination by machine learning there is a trade-off between erroneously determining a harmful defect as harmless and erroneously determining a harmless defect as a harmful defect. If machine learning is performed without considering this relationship, it may not be possible to reliably determine that a defect candidate portion determined to be harmful in an actual defect is harmful.
- the judging machine For example, by machine learning using N pieces of data of the defect candidate portion judged to be harmful in the actual defect and M pieces of data of the defect candidate portion judged to be harmless in the actual defect as learning data, the judging machine Consider the case where a is created. Let us consider a case where learning is simply performed so that the number of data judged as correct answers by the judging machine among (N+M) pieces of data is the highest. In that case, even if a physical defect is judged to be either harmful or harmless, data out of the population tends to be erroneously judged. Therefore, it is difficult to reliably determine that a defect candidate portion that has been determined to be harmful in an actual defect is harmful.
- R 01 is the number of correct answers of defect candidates determined to be harmful in actual defects.
- R 02 is the number of correct answers for defect candidates determined to be harmless in actual defects.
- a judgment machine is generated by machine learning so that the larger ⁇ is, the higher the correct answer rate of defect candidates judged to be harmful in terms of actual defects is, and the smaller the ⁇ is, the higher the correct answer rate is of defect candidates judged to be harmless in terms of actual defects. be done. Therefore, by adjusting ⁇ , it is possible to generate a determiner that reliably and correctly determines that a defect candidate portion determined to be harmful in an actual defect is harmful.
- the same effect can be obtained by duplicating the N pieces of data of the defect candidate portion determined to be harmful in the actual defect at the time of machine learning to increase the number of learning data.
- ⁇ is adjusted by performing machine learning while changing ⁇ , actually classifying the performance of the obtained determination machine with test data, and evaluating the results so as to meet the requirements for the first defect determination. You can do it by looking for a suitable condition.
- the test data may be the same as the learning data, but is preferably different from the learning data in order to suppress over-learning.
- the correct answer rate of the defect candidate part judged to be harmful in the target physical defect can be used, the correct answer rate of the defect candidate part judged to be harmless in the actual defect is used as a constraint condition.
- An approach may be taken to solve as an optimization problem to maximize.
- the classes may be further divided and the screening conditions for each class may be determined.
- detrimental defects may be classified into three categories: severe, moderate, or light, depending on the impact the defect has.
- the condition is set such that the correct answer rate for defect candidate portions determined to be severely harmful in actual defects is 100%.
- the condition is set such that the correct answer rate for the defect candidate portion determined to be moderately harmful is 75%.
- conditions are set such that the percentage of correct answers for defect candidate portions determined to be mildly harmful, for example, is 50%.
- R 11 is the correct number of defect candidates judged to be severely harmful in physical defects.
- R 12 is the number of correct answers for defect candidates determined to be moderately harmful in physical defects.
- R 13 is the number of correct answers for defect candidates determined to be mildly harmful in physical defects.
- R 14 is the number of correct answers for defect candidates determined to be harmless in actual defects.
- a line camera was used in the above embodiment, but an area camera may be used. Also, a color camera may be used instead of a monochrome camera.
- the camera may capture images under a plurality of conditions and combine the image information for inspection.
- the first defect determination may be performed using feature amounts calculated separately or in combination from images of defect candidates obtained under different conditions. For example, two images obtained by an optical system under two kinds of specular reflection conditions and diffuse reflection conditions may be used, or an image of each channel of a color image may be used. Further, the results of concatenation or operations (eg, addition and subtraction) of defect candidate images obtained under different conditions may be used in the first defect determination.
- the feature amount may be used in addition to the image.
- the feature amount may be used in addition to the image.
- the surface inspection of steel plates is targeted, but it goes without saying that the method can be applied not only to steel plates but also to various steel materials, various material products, and industrial products.
Landscapes
- Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Theoretical Computer Science (AREA)
- Pathology (AREA)
- Chemical & Material Sciences (AREA)
- Analytical Chemistry (AREA)
- Biochemistry (AREA)
- General Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Immunology (AREA)
- Health & Medical Sciences (AREA)
- Quality & Reliability (AREA)
- Software Systems (AREA)
- Signal Processing (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Medical Informatics (AREA)
- Artificial Intelligence (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Image Analysis (AREA)
- Investigating Materials By The Use Of Optical Means Adapted For Particular Applications (AREA)
- Image Processing (AREA)
Abstract
Description
検査対象の表面欠陥の検出を行う検査方法であって、
前記検査対象の表面を撮像した画像を取得する撮像ステップと、
前記画像から欠陥候補部を抽出する抽出ステップと、
抽出された前記欠陥候補部を第一の欠陥判定によって選別する選別ステップと、
前記第一の欠陥判定によって選別された前記欠陥候補部を対象として、畳み込みニューラルネットワークを用いた第二の欠陥判定によって前記表面欠陥の有害又は無害の検出を行う検査ステップと、を含む。
検査対象の表面欠陥の分類を行う分類方法であって、
前記検査対象の表面を撮像した画像を取得する撮像ステップと、
前記画像から欠陥候補部を抽出する抽出ステップと、
抽出された前記欠陥候補部を第一の欠陥判定によって選別する選別ステップと、
前記第一の欠陥判定によって選別された前記欠陥候補部を対象として、畳み込みニューラルネットワークを用いた第二の欠陥判定によって前記表面欠陥の種類及び等級の少なくとも一つの分類を行う分類ステップと、を含む。
上記の分類方法によって分類された前記表面欠陥の種類及び等級の少なくとも一つに基づいて、前記検査対象を分類する管理ステップを含む。
鋼材を製造する製造ステップと、
上記の検査方法における前記検査ステップと、を含み、
前記検査ステップは、前記製造ステップで製造された前記鋼材を前記検査対象とする。
鋼材を製造する製造ステップと、
上記の管理方法における前記管理ステップと、を含み、
前記管理ステップは、前記製造ステップで製造された前記鋼材を前記検査対象とする。
検査対象の表面欠陥の検出を行う検査方法で用いられる学習モデルの生成方法であって、
予め取得された検査対象の画像に対して第一の欠陥判定によって選別された欠陥候補部を入力実績データとして、前記入力実績データの有害又は無害の結果を出力実績データとする教師データを用いて、畳み込みニューラルネットワークによって前記学習モデルを生成するステップを含む。
上記の学習モデルの生成方法によって生成される。
検査対象の表面欠陥の検出又は分類を行う検査装置であって、
前記検査対象の表面を撮像した画像を取得する撮像装置と、
前記画像から欠陥候補部を抽出し、抽出された前記欠陥候補部を第一の欠陥判定によって選別し、前記第一の欠陥判定によって選別された前記欠陥候補部を対象として、畳み込みニューラルネットワークを用いた第二の欠陥判定によって前記表面欠陥の欠陥判定を行う演算装置と、を備え、
前記第二の欠陥判定によって行われる前記表面欠陥の欠陥判定は、前記表面欠陥の有害若しくは無害の検出、又は、前記表面欠陥の種類及び等級の少なくとも一つの分類である。
鋼材を製造する製造設備と、
上記の検査装置と、を備え、
前記検査装置は、前記製造設備で製造された前記鋼材を前記検査対象とする。
本実施形態に係る検査方法は、検査対象の表面欠陥の検出を行う。図1は、本実施形態に係る検査方法で実行される表面検査の処理フローの例を示す。まず、検査対象の表面を撮像した画像が取得される(ステップS1、撮像ステップ)。後述する実施例において、検査対象は鋼板であるが、鋼板に限定されない。
上記の検査方法は、例えば検査装置によって実行される。図3は、一実施形態に係る検査装置の構成例を示す図である。検査装置は上位システム(一例として検査対象の製造プロセスを管理するプロセスコンピュータ)によって制御される。検査装置は、検査対象の製造設備を構成する装置の一つとして設けられてよい。
図2は、判定機の生成フローの例を示す図である。図2には、図1と同じ表面検査の処理フローが点線で描かれている。点線の処理フローは、判定機の生成フローと同時に処理が実行されないことを示し、判定機の生成フローのステップS11、S12及びS21との関係性を示すために併記されている。判定機の生成フローによって判定機が生成された後に、生成された判定機を用いて表面検査の処理フローが実行される。
以下に、図3の検査装置を用いて行われた検査方法の実施例(第1実施例)が説明される。本実施例では、事前に機械学習を用いて構築された判定機を用いて、欠陥判定処理を実施することで、有害である欠陥と無害なものを精度よく弁別することができた。
以下に、第一の判定機の調整方法に関する実施例(第2実施例)が説明される。
Claims (11)
- 検査対象の表面欠陥の検出を行う検査方法であって、
前記検査対象の表面を撮像した画像を取得する撮像ステップと、
前記画像から欠陥候補部を抽出する抽出ステップと、
抽出された前記欠陥候補部を第一の欠陥判定によって選別する選別ステップと、
前記第一の欠陥判定によって選別された前記欠陥候補部を対象として、畳み込みニューラルネットワークを用いた第二の欠陥判定によって前記表面欠陥の有害又は無害の検出を行う検査ステップと、を含む、検査方法。 - 前記第一の欠陥判定は、
前記画像から特徴量を抽出し、前記特徴量を用いて選別する、請求項1に記載の検査方法。 - 前記第一の欠陥判定は、学習モデルを用いて選別し、
前記学習モデルは、任意の検査対象の表面を予め撮像した画像から抽出された特徴量を用いて、機械学習によって予め生成されたモデルである、請求項1に記載の検査方法。 - 検査対象の表面欠陥の分類を行う分類方法であって、
前記検査対象の表面を撮像した画像を取得する撮像ステップと、
前記画像から欠陥候補部を抽出する抽出ステップと、
抽出された前記欠陥候補部を第一の欠陥判定によって選別する選別ステップと、
前記第一の欠陥判定によって選別された前記欠陥候補部を対象として、畳み込みニューラルネットワークを用いた第二の欠陥判定によって前記表面欠陥の種類及び等級の少なくとも一つの分類を行う分類ステップと、を含む、分類方法。 - 請求項4に記載の分類方法によって分類された前記表面欠陥の種類及び等級の少なくとも一つに基づいて、前記検査対象を分類する管理ステップを含む、管理方法。
- 鋼材を製造する製造ステップと、
請求項1から3のいずれか一項に記載の検査方法における前記検査ステップと、を含み、
前記検査ステップは、前記製造ステップで製造された前記鋼材を前記検査対象とする、鋼材の製造方法。 - 鋼材を製造する製造ステップと、
請求項5に記載の管理方法における前記管理ステップと、を含み、
前記管理ステップは、前記製造ステップで製造された前記鋼材を前記検査対象とする、鋼材の製造方法。 - 検査対象の表面欠陥の検出を行う検査方法で用いられる学習モデルの生成方法であって、
予め取得された検査対象の画像に対して第一の欠陥判定によって選別された欠陥候補部を入力実績データとして、前記入力実績データの有害又は無害の結果を出力実績データとする教師データを用いて、畳み込みニューラルネットワークによって前記学習モデルを生成するステップを含む、学習モデルの生成方法。 - 請求項8に記載の学習モデルの生成方法によって生成される、学習モデル。
- 検査対象の表面欠陥の検出又は分類を行う検査装置であって、
前記検査対象の表面を撮像した画像を取得する撮像装置と、
前記画像から欠陥候補部を抽出し、抽出された前記欠陥候補部を第一の欠陥判定によって選別し、前記第一の欠陥判定によって選別された前記欠陥候補部を対象として、畳み込みニューラルネットワークを用いた第二の欠陥判定によって前記表面欠陥の欠陥判定を行う演算装置と、を備え、
前記第二の欠陥判定によって行われる前記表面欠陥の欠陥判定は、前記表面欠陥の有害若しくは無害の検出、又は、前記表面欠陥の種類及び等級の少なくとも一つの分類である、検査装置。 - 鋼材を製造する製造設備と、
請求項10に記載の検査装置と、を備え、
前記検査装置は、前記製造設備で製造された前記鋼材を前記検査対象とする、鋼材の製造設備。
Priority Applications (5)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP2022552999A JP7459957B2 (ja) | 2021-07-08 | 2022-06-20 | 検査方法、分類方法、管理方法、鋼材の製造方法、学習モデルの生成方法、学習モデル、検査装置及び鋼材の製造設備 |
CN202280044784.0A CN117597579A (zh) | 2021-07-08 | 2022-06-20 | 检查方法、分类方法、管理方法、钢材的制造方法、学习模型的生成方法、学习模型、检查装置以及钢材的制造设备 |
KR1020237043410A KR20240008931A (ko) | 2021-07-08 | 2022-06-20 | 검사 방법, 분류 방법, 관리 방법, 강재의 제조 방법, 학습 모델의 생성 방법, 학습 모델, 검사 장치 및 강재의 제조 설비 |
EP22837460.9A EP4361615A1 (en) | 2021-07-08 | 2022-06-20 | Inspection method, classification method, management method, steel material manufacturing method, training model generation method, training model, inspection device, and steel material manufacturing facility |
BR112023027343A BR112023027343A2 (pt) | 2021-07-08 | 2022-06-20 | Método de inspeção, método de classificação, método de gerenciamento, método de produção de material de aço, método de geração de modelo de treinamento, modelo de treinamento, dispositivo de inspeção e instalação de produção de material de aço |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP2021113864 | 2021-07-08 | ||
JP2021-113864 | 2021-07-08 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2023282043A1 true WO2023282043A1 (ja) | 2023-01-12 |
Family
ID=84800209
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/JP2022/024605 WO2023282043A1 (ja) | 2021-07-08 | 2022-06-20 | 検査方法、分類方法、管理方法、鋼材の製造方法、学習モデルの生成方法、学習モデル、検査装置及び鋼材の製造設備 |
Country Status (6)
Country | Link |
---|---|
EP (1) | EP4361615A1 (ja) |
JP (1) | JP7459957B2 (ja) |
KR (1) | KR20240008931A (ja) |
CN (1) | CN117597579A (ja) |
BR (1) | BR112023027343A2 (ja) |
WO (1) | WO2023282043A1 (ja) |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2010249624A (ja) * | 2009-04-15 | 2010-11-04 | Jfe Steel Corp | 走行材の表面品質判定装置および表面品質判定方法 |
JP2011214903A (ja) * | 2010-03-31 | 2011-10-27 | Denso It Laboratory Inc | 外観検査装置、外観検査用識別器の生成装置及び外観検査用識別器生成方法ならびに外観検査用識別器生成用コンピュータプログラム |
KR20150074942A (ko) * | 2013-12-24 | 2015-07-02 | 주식회사 포스코 | 스캡 결함 검출 장치 및 방법 |
WO2020175666A1 (ja) * | 2019-02-28 | 2020-09-03 | 大日本印刷株式会社 | カラーフィルタ検査装置、検査装置、カラーフィルタ検査方法および検査方法 |
CN111667465A (zh) * | 2020-05-22 | 2020-09-15 | 广东顺德募优网络科技有限公司 | 一种基于远红外图像的金属洗手盆缺陷检测方法 |
JP2020187657A (ja) * | 2019-05-16 | 2020-11-19 | 株式会社キーエンス | 画像検査装置 |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPH09257723A (ja) | 1996-03-26 | 1997-10-03 | Nkk Corp | 表面検査装置 |
-
2022
- 2022-06-20 JP JP2022552999A patent/JP7459957B2/ja active Active
- 2022-06-20 BR BR112023027343A patent/BR112023027343A2/pt unknown
- 2022-06-20 KR KR1020237043410A patent/KR20240008931A/ko unknown
- 2022-06-20 WO PCT/JP2022/024605 patent/WO2023282043A1/ja active Application Filing
- 2022-06-20 EP EP22837460.9A patent/EP4361615A1/en active Pending
- 2022-06-20 CN CN202280044784.0A patent/CN117597579A/zh active Pending
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2010249624A (ja) * | 2009-04-15 | 2010-11-04 | Jfe Steel Corp | 走行材の表面品質判定装置および表面品質判定方法 |
JP2011214903A (ja) * | 2010-03-31 | 2011-10-27 | Denso It Laboratory Inc | 外観検査装置、外観検査用識別器の生成装置及び外観検査用識別器生成方法ならびに外観検査用識別器生成用コンピュータプログラム |
KR20150074942A (ko) * | 2013-12-24 | 2015-07-02 | 주식회사 포스코 | 스캡 결함 검출 장치 및 방법 |
WO2020175666A1 (ja) * | 2019-02-28 | 2020-09-03 | 大日本印刷株式会社 | カラーフィルタ検査装置、検査装置、カラーフィルタ検査方法および検査方法 |
JP2020187657A (ja) * | 2019-05-16 | 2020-11-19 | 株式会社キーエンス | 画像検査装置 |
CN111667465A (zh) * | 2020-05-22 | 2020-09-15 | 广东顺德募优网络科技有限公司 | 一种基于远红外图像的金属洗手盆缺陷检测方法 |
Also Published As
Publication number | Publication date |
---|---|
JP7459957B2 (ja) | 2024-04-02 |
EP4361615A1 (en) | 2024-05-01 |
CN117597579A (zh) | 2024-02-23 |
KR20240008931A (ko) | 2024-01-19 |
BR112023027343A2 (pt) | 2024-03-12 |
JPWO2023282043A1 (ja) | 2023-01-12 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US10803573B2 (en) | Method for automated detection of defects in cast wheel products | |
Shipway et al. | Automated defect detection for fluorescent penetrant inspection using random forest | |
CN109543760A (zh) | 基于图像滤镜算法的对抗样本检测方法 | |
Islam et al. | Capsule image segmentation in pharmaceutical applications using edge-based techniques | |
CN114463843A (zh) | 一种基于深度学习的多特征融合鱼类异常行为检测方法 | |
TWI707299B (zh) | 光學檢測二次圖像分類方法 | |
JP2021143884A (ja) | 検査装置、検査方法、プログラム、学習装置、学習方法、および学習済みデータセット | |
KR102666787B1 (ko) | 인공지능을 이용한 결함 검사방법, 장치 및 프로그램 | |
CN111161233A (zh) | 一种用于冲孔皮革缺陷检测方法及系统 | |
Hashmi et al. | Computer-vision based visual inspection and crack detection of railroad tracks | |
CN206897873U (zh) | 一种基于检测产品特性的图像处理与检测系统 | |
WO2023282043A1 (ja) | 検査方法、分類方法、管理方法、鋼材の製造方法、学習モデルの生成方法、学習モデル、検査装置及び鋼材の製造設備 | |
JP2020077158A (ja) | 画像処理装置及び画像処理方法 | |
Mansano et al. | Inspection of metallic surfaces using local binary patterns | |
Wang et al. | Detection of capsule foreign matter defect based on BP neural network | |
Niskanen et al. | Experiments with SOM based inspection of wood | |
CN113267506A (zh) | 木板ai视觉缺陷检测装置、方法、设备及介质 | |
Rale et al. | Comparison of different ANN techniques for automatic defect detection in X-Ray images | |
Xu et al. | A multitarget visual attention based algorithm on crack detection of industrial explosives | |
CN115830403B (zh) | 一种基于深度学习的自动缺陷分类系统及方法 | |
Hernandez et al. | Automated defect detection in aluminium castings and welds using neuro-fuzzy classifiers | |
CN111724352B (zh) | 一种基于核密度估计的贴片led瑕疵标注方法 | |
CN217332186U (zh) | 木板ai视觉缺陷检测装置 | |
Momot et al. | Analysis of the influence of the threshold level of binarization on the efficiency of segmentation of images of surface defects in steel by the U-NET network | |
TWI745767B (zh) | 光學檢測二次圖像分類方法 |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
ENP | Entry into the national phase |
Ref document number: 2022552999 Country of ref document: JP Kind code of ref document: A |
|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 22837460 Country of ref document: EP Kind code of ref document: A1 |
|
ENP | Entry into the national phase |
Ref document number: 20237043410 Country of ref document: KR Kind code of ref document: A |
|
WWE | Wipo information: entry into national phase |
Ref document number: 1020237043410 Country of ref document: KR |
|
WWE | Wipo information: entry into national phase |
Ref document number: 18571253 Country of ref document: US |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2301008337 Country of ref document: TH |
|
WWE | Wipo information: entry into national phase |
Ref document number: 202280044784.0 Country of ref document: CN |
|
REG | Reference to national code |
Ref country code: BR Ref legal event code: B01A Ref document number: 112023027343 Country of ref document: BR |
|
WWE | Wipo information: entry into national phase |
Ref document number: MX/A/2024/000342 Country of ref document: MX |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2022837460 Country of ref document: EP |
|
ENP | Entry into the national phase |
Ref document number: 2022837460 Country of ref document: EP Effective date: 20240123 |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
ENP | Entry into the national phase |
Ref document number: 112023027343 Country of ref document: BR Kind code of ref document: A2 Effective date: 20231222 |