WO2021161823A1 - Dispositif de détermination, procédé de génération de modèle de restauration, et programme informatique - Google Patents

Dispositif de détermination, procédé de génération de modèle de restauration, et programme informatique Download PDF

Info

Publication number
WO2021161823A1
WO2021161823A1 PCT/JP2021/003445 JP2021003445W WO2021161823A1 WO 2021161823 A1 WO2021161823 A1 WO 2021161823A1 JP 2021003445 W JP2021003445 W JP 2021003445W WO 2021161823 A1 WO2021161823 A1 WO 2021161823A1
Authority
WO
WIPO (PCT)
Prior art keywords
restoration
feature amount
determination device
constraint condition
input data
Prior art date
Application number
PCT/JP2021/003445
Other languages
English (en)
Japanese (ja)
Inventor
悠希 松本
Original Assignee
住友電気工業株式会社
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 住友電気工業株式会社 filed Critical 住友電気工業株式会社
Priority to JP2022500323A priority Critical patent/JPWO2021161823A1/ja
Publication of WO2021161823A1 publication Critical patent/WO2021161823A1/fr

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis

Definitions

  • the present disclosure relates to a determination device, a restoration model generation method, and a computer program.
  • This application claims priority based on Japanese Application No. 2020-023446 filed on February 14, 2020, and incorporates all the contents described in the Japanese application.
  • Non-Patent Document 1 discloses defect detection using an autoencoder.
  • the autoencoder dimensionally compresses the input data (input image) using a neural network to generate the restored data (restored image).
  • the autoencoder learns so that the difference between the input data and the restored data becomes small. Since the autoencoder learned using only the image showing the non-defective product fails to restore the defective part, the defective product can be detected from the restoration error of the restored data.
  • Non-Patent Document 2 discloses that a regularization term is used to prevent overfitting in machine learning.
  • the regularization term in Non-Patent Document 2 is a constraint condition of the weight parameter to be multiplied by the input to the neuron.
  • the determination device of the present disclosure includes an intermediate layer that dimensionally compresses input data and outputs dimensionally compressed features, and uses a restoration model that generates restoration data based on the features and a restoration error of the restoration data.
  • the restoration model includes a determination device for determining an abnormality of the input data, and the restoration model is a neural network machine-learned using an objective function having a constraint condition of the feature amount.
  • the restoration model generation method of the present disclosure includes generating a restoration model by machine learning that generates restoration data used for determining an abnormality of the input data based on a feature amount in which the input data is dimensionally compressed. , The machine learning uses an objective function including the constraint condition of the feature amount.
  • the computer program of the present disclosure is a computer program for causing a computer to execute a restoration model generation process, and the generation process determines an abnormality of the input data based on a feature amount in which the input data is dimensionally compressed.
  • a restoration model for generating the restoration data used for the purpose is generated by machine learning, and the machine learning uses an objective function including a constraint condition of the feature amount.
  • FIG. 1 is a block diagram of a processing unit that operates as a determination process.
  • FIG. 2 is a hardware configuration diagram of the processing unit.
  • FIG. 3 is a block diagram of the restoration model.
  • FIG. 4 is an explanatory diagram of neurons.
  • FIG. 5 is a flowchart of the determination process.
  • FIG. 6 is a block showing a machine learning unit.
  • FIG. 7 is an explanatory diagram of the feature amount.
  • FIG. 8 is a block diagram showing a machine learning unit.
  • FIG. 9 is an explanatory diagram of the calculation of the inter-distribution distance index.
  • FIG. 10 is a frequency distribution of restoration errors according to a comparative example.
  • FIG. 11 is a frequency distribution of restoration errors according to the embodiment.
  • FIG. 12 is an explanatory diagram of the calculation of the restoration error.
  • an autoencoder capable of generating restored data with high accuracy can generally be said to have high performance.
  • an increase in the performance of a neural network is, for example, an increase in the number of layers of the neural network or an increase in the number of parameters such as weights between neurons.
  • an auto encoder that can generate restored data from input data with high accuracy
  • an image showing a defective product is also restored with high accuracy, and even if it is a defective product, the restoration error of the restored data becomes small.
  • the restoration error of a non-defective product is small and the restoration error of a defective product is large, the autoencoder can easily determine whether the product is a non-defective product or a defective product.
  • the restoration error of not only the non-defective product but also the defective product is small, it is difficult for the autoencoder to distinguish between the non-defective product and the defective product due to the restoration error.
  • the determination device includes a restoration model.
  • the restoration model includes an intermediate layer that dimensionally compresses the input data and outputs the dimensionally compressed features.
  • the restoration model generates restoration data based on the feature quantity.
  • the determination device includes a determination device.
  • the determination device determines an abnormality in the input data based on the restoration error of the restoration data.
  • the reconstruction model is a neural network machine-learned using an objective function having the constraint condition of the feature quantity.
  • the constraint condition of the feature amount can include a constraint condition of the distance between the frequency distribution of the feature amount and the reference distribution of the feature amount. By using the constraint condition of the distance from the reference distribution, the frequency distribution of the feature amount can be made into a desired distribution.
  • the reference distribution may be a distribution having a frequency peak in a section to which a feature amount having a value of zero belongs. In this case, most of the features can be zero or near zero.
  • the determination device may further include an adjuster for adjusting the distribution shape of the reference distribution. In this case, it is possible to search for the optimum reference distribution.
  • the constraint condition of the feature amount can have a constraint condition of the distance of the feature amount.
  • the feature distance can be reduced and the feature dimension can be substantially reduced.
  • the objective function may further have a hyperparameter which is a regularization coefficient for the constraint condition of the feature quantity. In this case, the influence of the constraint condition of the feature amount can be adjusted.
  • the restoration model may be a neural network machine-learned using images of a plurality of products of the same standard that are determined to be normal products in terms of appearance shape.
  • the input data may be an image of the same standard product that has not been determined with respect to the appearance shape.
  • the determination device may be configured to determine the undetermined product of the same standard as a normal product or a defective product. In this case, defect detection (defective product detection) becomes possible.
  • a restoration model for generating restoration data used for determining an abnormality of the input data based on a feature amount in which the input data is dimensionally compressed is generated by machine learning. Can include doing. For the machine learning, an objective function including the constraint condition of the feature amount can be used.
  • the computer program according to the embodiment can cause the computer to execute the restoration model generation process.
  • the generation process can include generating a restoration model by machine learning that generates restoration data used for determining an abnormality of the input data based on a feature amount in which the input data is dimensionally compressed. For the machine learning, an objective function including the constraint condition of the feature amount can be used.
  • Computer programs are stored on computer-readable, non-temporary storage media.
  • FIG. 1 shows a processing unit 10 that operates as a determination device according to an embodiment.
  • the determination device according to the embodiment is used for defect detection of a product (product 100) of the same standard, which is a product produced in a factory and has not been determined whether it is a normal product or a defective product in terms of appearance shape.
  • the 100 that is the target of defect detection is, for example, a component such as a connector.
  • the connector is provided, for example, at the end of the wire harness. Wire harnesses are used, for example, for in-vehicle wiring.
  • the connector has a connector housing and a plurality of terminals (for example, four terminals) provided in the connector housing.
  • the connector is imaged by the image pickup device 50.
  • the image pickup apparatus 50 is installed so as to obtain an image in which a plurality of terminals inside the manufactured connector housing are captured, for example.
  • the processing unit 10 of the embodiment acquires an image from the image pickup apparatus 50, and executes a process for detecting a defect of the connector using the acquired image.
  • Defective connectors include, for example, damage to the connector housing, intrusion of foreign matter, and defective terminals.
  • the processing unit 10 is composed of a processor 11 and a storage device 12 connected to the processor 11 and a computer including the processor 11.
  • the processor 11 includes, for example, a CPU (Central Processing Unit) and a GPU (Graphics Processing Unit).
  • the computer program 13 is stored in the storage device 12.
  • the storage device 12 may be a magnetic storage device such as a hard disk device, an optical storage device such as an optical disk device, or a flash memory such as an SSD (Solid State Drive).
  • the storage device 12 may be referred to as a memory in the present specification.
  • the processing unit 10 also includes an interface 14.
  • the image pickup device 50 and the display device 60 are connected to the processing unit 10 via the interface 14.
  • the computer program 13 is read and executed by the processor 11.
  • the computer program 13 has a program code for causing the processor 11 to execute a process (determination process) for detecting a defect. Further, the computer program 13 has a program code for causing the processor 11 to execute a process (machine learning process) for generating a restoration model used for defect detection.
  • the restoration model may be referred to as a learning model.
  • the processing unit 10 shown in FIG. 1 a functional block showing an operation realized by the processor 11 executing the computer program 13 is shown.
  • the processing unit 10 that operates as a determination device includes a restoration model 20.
  • the restoration model 20 generates the restoration image AE (X) which is the restoration data based on the input image X which is the input data.
  • the restored image AE (X) shows the restored input image X.
  • the restoration model 20 is composed of a neural network including an input layer 110, one or more intermediate layers 120, and an output layer 130.
  • the input layer 110, the intermediate layer 120, and the output layer 130 each have a plurality of neurons (units).
  • the input layer 110 accepts the image acquired from the image pickup apparatus 50 as the input image X.
  • the output layer 130 outputs the restored image AE (X).
  • the restoration model 20 according to the embodiment is an autoencoder.
  • the autoencoder generates features by dimensionally compressing the input data (encoding), and generates restored data by reconstructing the input data based on the features (decoding).
  • the restoration model 20 does not have to be an autoencoder, and it is sufficient if the restoration model 20 can generate the restoration data of the input data.
  • the intermediate layer 120 of the autoencoder generates m-dimensional feature quantities Z 1 , Z 2 , Z 3, ... By dimensionally compressing the input image X.
  • the m-dimensional feature quantities Z 1 , Z 2 , Z 3, ... are also referred to as feature quantity vectors.
  • the restoration model 20 constructs restoration data AE (X) based on the feature quantities Z 1 , Z 2 , Z 3, ....
  • the intermediate layer 120 that outputs the feature amount has a number of dimensions m that is smaller than the number of dimensions n of the input data.
  • the number of dimensions m corresponds to the number of neurons in the middle layer 120.
  • a plurality of feature quantities Z 1 , Z 2 , Z 3, ... Output from the intermediate layer 120 constitute an m-dimensional feature quantity vector (feature quantity space).
  • FIG. 4 shows neurons 150 (units) contained in the intermediate layer 120.
  • X 1 , X 2 , X 3 , ... Are the outputs of neurons contained in the layer (input layer 110 or other intermediate layer) in front of the intermediate layer 120, to the neurons 150 shown in FIG. It is also the input given.
  • w 1 , w 2 , w 3 , ... Are weight parameters multiplied by inputs X 1 , X 2 , X 3 , ..., And indicate the weights of connections between neurons.
  • Neuron 150 computes output Z 1 using an activation function that includes inputs X 1 , X 2 , X 3 , ..., weights w 1 , w 2 , w 3, ..., and bias.
  • Output Z 1 indicates a feature amount output from the neuron 150.
  • the restoration model 20 of FIG. 1 is generated by machine learning using only normal data which is an image showing a product 100 which is a non-defective product.
  • Machine learning is executed, for example, by the machine learning unit 40.
  • the restoration model 20 trained using the training data 70 including only the normal data can acquire the optimum feature space for restoring the non-defective image which is the normal data. In this feature space, the restoration model 20 is difficult to restore abnormal data which is a defective product image showing a defective product. The difference between the input data and the restored data is called the restoration error.
  • the processing unit 10 that operates as a determination device determines an abnormality in the input data by utilizing the fact that the restoration error of the abnormality data becomes large. It should be noted that a plurality of products of the same standard that are determined to be normal products in terms of appearance and shape are suitable for defect detection by an autoencoder. This is because a large number of products having the same appearance and shape are produced, so that a large amount of normal data (non-defective product images) showing normal products (non-defective products) having the same appearance and shape can be easily collected as learning data. Although there is a problem of overfitting, the large amount of learning data greatly contributes to improving the accuracy of autoencoder defect detection.
  • the processing unit 10 that operates as a determination device includes a determination device 30 for determining an abnormality in input data.
  • the determination device 30 includes a restoration error calculation unit 31 and a comparison unit 32.
  • the restoration error calculation unit 31 calculates the restoration error based on the input data and the restoration data.
  • the restoration error is calculated as, for example, a square error between the pixel value of the input data and the pixel value of the restored data.
  • the comparison unit 32 determines the presence or absence of an abnormality in the input data by comparing the restoration error with the threshold value.
  • the determination result of the determination device 30 is output to, for example, the display device 60.
  • FIG. 5 shows the procedure of the determination process executed by the processing unit 10 operating as the determination device.
  • the determination process the presence or absence of an abnormality in the input data, that is, the quality of the product 100, which is the defect detection target shown in the input image X, is determined.
  • step S11 of the determination process the processing unit 10 generates the restored image AE (X) from the input image X by using the restoration model 20.
  • the input image X is an image obtained by the image pickup apparatus 50 and shows the product 100. If the product 100 is a non-defective product, the restoration accuracy of the restored image AE (X) is high, and if the product 100 is a defective product, the restoration accuracy of the restored image AE (X) is low.
  • step S12 of the determination process the restoration error calculation unit 31 calculates the restoration error L con of the restored image AE (X).
  • the restoration error L con is calculated according to the equation (1) in FIG. 5 using the input image X and the restored image AE (X).
  • step S13 of the determination process the comparison unit 32 compares the restoration error L con with the threshold value T. If the restoration error L con is larger than the threshold value T, it is determined that the input image X is abnormal, that is, the product 100 shown in the input image X is defective (step S14). On the other hand, if the restoration error L con is smaller than the threshold value T, it is determined that the input image X is normal, that is, the product 100 shown in the input image X is a non-defective product (step S15). The display device 60 outputs the determination result of defective / non-defective product.
  • FIG. 6 shows a machine learning unit 40 that generates the restoration model 20 according to the embodiment.
  • the machine learning unit 40 generates a restoration model 20 by machine learning using learning data that does not include abnormal data that is a defective image indicating a defective product and includes only normal data that is a good image indicating a non-defective product.
  • the weight parameter w included in the reconstruction model 20 is optimized so as to minimize the objective function represented by the equation (2) in FIG.
  • the objective function represented by the equation (2) in FIG. 6 includes a restoration error L con and an L2 norm L Z (distance of the feature amount Z) of the feature amount Z. Since the objective function of the embodiment includes the restoration error L con , when the input image X (non-defective image) which is the training data is given to the autoencoder, the autoencoder accurately restores the input image X and the restoration error L con. Becomes smaller.
  • the objective function according to the embodiment further to include constraints L2 norm L Z feature amount Z (Distance feature quantity Z), it is possible to reduce the L2 norm L Z feature quantity Z.
  • the output is output from the intermediate layer 120 while improving the restoration performance of the restoration model 20 during machine learning.
  • the feature amount Z to be formed may be restricted.
  • the restoration model 20 can learn the constraint condition of the feature amount Z and the improvement of the restoration performance at the same time.
  • the objective function for machine learning includes only the restoration error L con , it is easy to fall into overfitting.
  • an objective function including a constraint condition of the weight parameter w to be multiplied by the input to the neuron as a regularization term as in Non-Patent Document 2.
  • the constraint condition of the weight parameter w for example, the L2 norm of the weight parameter w is used.
  • the optimal solution of the weight parameter w suitable for restoring normal data is due to the constraint condition of the weight parameter w. Search is hindered.
  • the objective function of the present embodiment has a regularization term that is not a constraint condition of the weight parameter w but a constraint condition of the feature quantity Z.
  • Constraint feature amount Z of this case as an example, a L2 norm L Z feature quantity Z.
  • the L2 norm L Z of the feature quantity Z is calculated as the L2 norm (Euclidean norm) of a vector having a plurality of feature quantity Z as elements.
  • the constraint condition of the feature amount Z may be an index indicating the L1 norm of the feature amount Z or another distance of the feature amount Z.
  • the weight parameter w of the restoration model 20 is optimized so that the L2 norm L Z of the feature quantity Z becomes small. Even if a constraint condition is added to the feature quantity Z, the weight parameter w that affects the magnitude of the feature quantity Z is not directly subject to the constraint condition. Therefore, the search range of the optimum solution of the weight parameter w is not restricted, and the optimization of the weight parameter w is easy.
  • CaseA of FIG. 7 shows a case where the objective function includes only the restoration error L con, as shown in the equation (3) in FIG.
  • the graph included in CaseA shows an example of the value of each feature amount output from each of the plurality of neurons contained in the intermediate layer 120.
  • the horizontal axis is the dimension of the feature amount Z, and values from 1 to m are taken.
  • the vertical axis indicates the value of the feature amount Z, and indicates the magnitude of each value of the m feature amounts.
  • the feature amount has m number of dimensions.
  • each of the m dimensions often has a feature amount with a non-zero value.
  • L2 norm L Z feature amount Z increases.
  • Case B of FIG. 7 shows a case where the objective function includes not only the restoration error L con but also the constraint condition of the feature amount Z as shown in the equation (2) in FIGS. 6 and 7. Due to the constraint condition for reducing the L2 norm L Z of the feature quantity Z, in the graph included in Case B, the values of some feature quantity Z are zero or near zero.
  • the feature amount Z has m number of dimensions, but when the feature amount Z takes zero or a value near zero, the number of dimensions of the feature amount Z is substantially reduced. That is, when a constraint condition is added to the feature amount Z, the dimension of the feature amount exists, but the magnitude of the feature amount becomes zero or a value near zero, so that the number of neurons that output the feature amount Z is reduced. The same effect as
  • the feature amount Z By reducing the number of dimensions of the feature amount Z to the extent necessary and sufficient for restoring normal data, it becomes easy to distinguish between normal data and abnormal data. Since the feature amount Z having a small number of dimensions constitutes only a feature amount space necessary and sufficient for restoring normal data, the restoration error of abnormal data is increased to obtain normal data and abnormal data. Make it easy to distinguish between.
  • the restoration model 20 is optimized only to the extent that normal data can be restored, so that the restoration performance of the defective image, which is abnormal data, is significantly deteriorated, and the restoration error is sufficient. growing.
  • the feature amount Z is not clear, it is difficult to appropriately reduce the number of dimensions of the feature amount Z itself. Therefore, in the present embodiment, instead of taking the difficult method of reducing the number of dimensions m of the feature amount Z (the number of neurons in the intermediate layer 120) itself, the feature amount is maintained while maintaining the number of dimensions m of the feature amount Z. Add a constraint to the Z norm. As a result, the feature amount Z becomes sparse, and the number of dimensions m of the feature amount Z can be substantially reduced.
  • Equation (2) shown in FIGS. 6 and 7 has a hyperparameter ⁇ for the constraint condition of the feature amount Z.
  • the hyperparameter ⁇ is a regularization coefficient multiplied by the regularization term, and defines the magnitude of the influence of the regularization term.
  • the hyperparameter ⁇ takes a value from 0 to 1, for example.
  • the magnitude of the hyperparameter ⁇ is automatically adjusted during machine learning by the regulator 41 included in the processing unit 10.
  • the present inventor conducted an experiment of determining abnormality / normality of the input image X by using the processing unit 10 having the machine-learned restoration model 20 as shown in FIG.
  • the present inventor optimized the restoration model 20 by using 100 non-defective images as learning data and using the equation (2) as an objective function.
  • the present inventor used 4000 non-defective product images (normal data) and 40 defective product images (abnormal data) as verification data to be input to the restoration model 20.
  • all defective image images were determined to be abnormal, and the defective image was not determined to be normal.
  • the rate of over-detection in which a non-defective image is determined to be abnormal varies depending on the value of the hyperparameter ⁇ , but can be suppressed to a relatively low rate.
  • a non-defective product and a defective product are accurately distinguished, and in particular, since the defective product image is not determined to be normal, the shipment of the defective product is surely prevented.
  • FIG. 8 shows another example of machine learning by the machine learning unit 40.
  • the objective function represented by the equation (4) in FIG. 8 includes a restoration error L con and a constraint condition of the inter-distribution distance index L Z as a regularization term.
  • the restoration model 20 is optimized so as to reduce the restoration error L con and the inter-distribution distance index L Z.
  • the inter-distribution distance index L Z is an index showing the distance between the frequency distribution 48 (histogram) of the feature amount Z shown in FIG. 9 and the reference distribution 46.
  • Distribution distance index L Z represents the similarity of the two distributions 48, 46.
  • the inter-distribution distance index L Z Kullback-Leibler distance (KL distance) is used.
  • KL distance is an index showing the difference between the two distributions.
  • Distribution distance index L Z is not limited to the KL distance, Jensen-Shannon, it may be Wasserstein, or Histogram Intersection.
  • the restoration model 20 brings the shape of the frequency distribution 48 of the feature amount Z closer to the shape of the reference distribution 46 by learning so that the difference between the two distributions 48 and 46 becomes small based on the inter-distribution distance index L Z. be able to.
  • the processing unit 10 operating as the machine learning unit 40 executes a generation process 45 that generates a frequency distribution 48 in which the feature amount Z output from the intermediate layer 120 is histogramd during machine learning. do.
  • the frequency distribution 48 shows the relationship between the feature amount Z and the frequency of the feature amount Z (the number in which the feature amount Z is generated).
  • the frequency distribution 48 is represented as, for example, a distribution showing the number of feature quantities Z belonging to each interval when the value of the feature quantity Z is divided into a plurality of intervals.
  • Processing unit 10 which operates as a calculation unit 47 of the distribution between the distance index L Z is a frequency distribution 48 of the feature Z, the reference distribution 46, on the basis to calculate the inter-distribution distance index L Z.
  • the reference distribution 46 is, for example, a Gaussian distribution having a small variance.
  • the horizontal axis is the value of the feature amount Z
  • the vertical axis is the frequency (frequency) of a certain feature amount Z.
  • the frequency distribution 48 is a distribution in which a peak of frequency occurs in each section indicated by the horizontal axis to which the zero feature Z belongs.
  • the reference distribution 46 is preferably a distribution having a frequency peak in the interval to which the zero feature amount Z belongs (for example, a Gaussian distribution having a peak at zero). Further, since the variance is small, the number of feature quantities Z having zero or a value near zero increases in the frequency distribution 48 of the feature quantity Z. As a result, the number of dimensions of the feature amount Z is substantially reduced.
  • the frequency distribution of the desired shape can be obtained by appropriately setting the shape of the reference distribution 46. .. Therefore, the feature space is appropriately controlled through the adjustment of the reference distribution 46.
  • the shape of the reference distribution 46 is adjusted during machine learning by the processing unit 10 that operates as the regulator 41.
  • the regulator 41 has a first regulator 41A that adjusts the shape of the reference distribution 46. Since the reference distribution 46 is adjustable, it is possible to search for an appropriate reference distribution 46 in order to obtain a restoration model 20 suitable for determining abnormality / normality during machine learning.
  • the shape of the reference distribution 46 is adjusted, for example, by changing the type of distribution. Changes in the type of distribution include, for example, changing from a Gaussian distribution to another type of distribution.
  • the adjustment of the shape of the reference distribution 46 may include a change in the variance of the reference distribution 46. Adjusting the shape of the reference distribution 46 may include changing the magnitude of the peaks of the distribution.
  • the reference distribution 46 may be, for example, a distribution in which only the section to which the feature quantity Z of zero belongs has a non-zero frequency, and the other sections have a zero frequency, that is, a distribution in which the variance is zero.
  • the regulator 41 shown in FIG. 8 includes a second regulator 41B that adjusts the regularization coefficient ⁇ during machine learning, similarly to the regulator 41 shown in FIG.
  • FIG. 10 shows the frequency (frequency) distribution of the restoration error for the comparative example
  • FIG. 11 shows the frequency (frequency) distribution of the restoration error for the example.
  • the objective function used for machine learning of the restoration model 20 a function having a constraint condition of the weight parameter w as a regularization term was used in addition to the restoration error L con.
  • the objective function a function having a constraint condition of the inter-distribution distance index L Z as a regularization term in addition to the restoration error L con (see equation (4) in FIG. 8) was used.
  • 100 non-defective images were used as learning data.
  • the threshold value T to be compared with the restoration error due to abnormality / normality is set to a value at which abnormality (defective product) can be detected without omission.
  • the threshold value T is set to a value of about 800.
  • the threshold value T is set to a value of about 850.
  • the comparison unit 32 determines abnormal / normal using the threshold value T capable of detecting an abnormality (defective product) without omission, it erroneously determines that the normal (non-defective product) is abnormal (defective product).
  • the comparison unit 32 determines whether the abnormality (defective product) is abnormal / normal using the threshold value T capable of detecting the abnormality (defective product) without omission, it erroneously determines that the normal (non-defective product) is abnormal (defective product).
  • the comparison unit 32 may erroneously determine normal as abnormal or erroneously determine abnormal as normal.
  • the restoration error range in which normal exists (range from 150 to 850) and the restoration error range in which abnormality exists (range from 850 to 1900) are set. It is separated. Therefore, in the embodiment, even if the threshold value is set so that the abnormality (defective product) is detected without omission, the comparison unit 32 rarely erroneously determines that the normal (non-defective product) is abnormal (defective product). , Advantageous.
  • FIG. 12 shows the method used to calculate the restoration error L con shown in FIGS. 10 and 11 in Comparative Examples and Examples. This calculation is performed by the comparison unit 32 to obtain the restoration error L con used for determining abnormality / normality, but the comparison unit 32 may be performed to obtain the restoration error L con used for machine learning.
  • the comparison unit 32 generates restoration error data indicating an error in the pixel value of each pixel included in the input image X and the restored image AE (X). Subsequently, the comparison unit 32 calculates the total value (total error) of the errors in the regions A1, A2, A3, and A4 for each of the plurality of terminals. Since the input image X and the restored image AE (X) shown in FIG. 12 show four terminals, a total error is calculated for each of the four regions A1, A2, A3, and A4.
  • the largest error is adopted as the restoration error L con of the restored image AE (X) and is used for determining abnormality / normality.
  • the maximum value of the total error (partial restoration error) of each of the divided plurality of regions A1, A2, A3, and A4 is adopted as the restoration error L con used for determining abnormality / normality. Will be done. This facilitates the determination of abnormality / normality.
  • the restoration performance is lowered, and the restoration error may be large for the entire image.
  • the restoration error is caused by, for example, the illumination light reflected in the input image X. Even if the restoration error of the entire image becomes large due to the influence of light reflection or other noise, the restoration error in the divided regions A1, A2, A3, and A4 is relatively small if the data is normal. That is, since the restoration error of the entire image is the sum of the restoration errors of each of the plurality of regions A1, A2, A3, A4, the restoration error of each region A1, A2, A3, A4 is larger than the restoration error of the entire image. It becomes smaller.
  • the abnormality may occur in any one of the divided regions A1, A2, A3, and A4, and in that case, the restoration error of the regions A1, A2, A3, and A4 including the abnormality becomes very large. .. Therefore, even if the restoration error of the entire image becomes large due to the influence of light reflection or other noise, the restoration error L con used for determining abnormality / normality can be suppressed to a small value. As a result, the reconstruction error L con when it is abnormal, it is easy to distinguish, and reconstruction error L con when it is normal, abnormal / normal judgment is facilitated.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Image Analysis (AREA)
  • Image Processing (AREA)

Abstract

L'invention concerne un dispositif de détermination comprenant : un modèle de restauration contenant une couche intermédiaire destinée à réduire la dimension de données d'entrée et à émettre en sortie une caractéristique dont la dimension est réduite et configurée pour générer des données restaurées en fonction de cette caractéristique ; et un dispositif de détermination configuré pour déterminer une anomalie dans les données d'entrée en fonction d'une erreur de restauration dans les données restaurées. Le modèle de restauration est un réseau neuronal qui est entraîné par apprentissage automatique au moyen d'une fonction objective avec une contrainte pour la caractéristique donnée.
PCT/JP2021/003445 2020-02-14 2021-02-01 Dispositif de détermination, procédé de génération de modèle de restauration, et programme informatique WO2021161823A1 (fr)

Priority Applications (1)

Application Number Priority Date Filing Date Title
JP2022500323A JPWO2021161823A1 (fr) 2020-02-14 2021-02-01

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
JP2020023446 2020-02-14
JP2020-023446 2020-02-14

Publications (1)

Publication Number Publication Date
WO2021161823A1 true WO2021161823A1 (fr) 2021-08-19

Family

ID=77292193

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/JP2021/003445 WO2021161823A1 (fr) 2020-02-14 2021-02-01 Dispositif de détermination, procédé de génération de modèle de restauration, et programme informatique

Country Status (2)

Country Link
JP (1) JPWO2021161823A1 (fr)
WO (1) WO2021161823A1 (fr)

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2019003615A (ja) * 2017-04-27 2019-01-10 ダッソー システムズＤａｓｓａｕｌｔ Ｓｙｓｔｅｍｅｓ オートエンコーダの学習
JP2019106111A (ja) * 2017-12-14 2019-06-27 キヤノン株式会社 情報処理装置、情報処理方法、及びプログラム
JP2019139482A (ja) * 2018-02-09 2019-08-22 株式会社デンソーアイティーラボラトリ 情報推定装置及び情報推定方法

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2019003615A (ja) * 2017-04-27 2019-01-10 ダッソー システムズＤａｓｓａｕｌｔ Ｓｙｓｔｅｍｅｓ オートエンコーダの学習
JP2019106111A (ja) * 2017-12-14 2019-06-27 キヤノン株式会社 情報処理装置、情報処理方法、及びプログラム
JP2019139482A (ja) * 2018-02-09 2019-08-22 株式会社デンソーアイティーラボラトリ 情報推定装置及び情報推定方法

Non-Patent Citations (6)

* Cited by examiner, † Cited by third party
Title
MASAYA MORI; KATSUKO T. NAKAHIRA; HIROTAKA TAKAHASHI; TAKAHIRO TANAKA: "Toward Detection of Small Signals from Noisy Data Using Variational Autoencoder Based Anomaly Detection", FIT2019 - 18TH INFORMATION SCIENCE AND TECHNOLOGY FORUM, 20 August 2019 (2019-08-20), JP, pages 81 - 84, XP009530554 *
MASUMI ISHIKAWA: "Towards understandable deep learning in stacked autoencoders", IEICE TECHNICAL REPORT, vol. 118, no. 470, NC2018-62, 1 March 2019 (2019-03-01), JP, pages 99 - 104, XP009530557, ISSN: 0913-5685 *
RYOSUKE TACHIBANA , TAKASHI MATSUBARA , KUNIAKI UEHARA: "4G4-4 Label estimation and semi-supervised learning by hostile networks in deep learning", 2016 ANNUAL MEETING OF THE JAPANESE SOCIETY FOR ARTIFICIAL INTELLIGENCE (30TH) PROCEEDINGS, 6 June 2016 (2016-06-06), JP, pages 1 - 4, XP009530556, DOI: 10.11517/pjsai.JSAI2016.0_4G44 *
TAKAYUKI OKATANI, MASAKI SAITO: "Deep learning", SIG TECHNICAL REPORTS, vol. 2013-CVIM-185, no. 19, 15 February 2013 (2013-02-15), JP, pages 1 - 17, XP009530561, DOI: http://id.nii.ac.jp/1001/00088178/ *
TETSURO TATSUOKA: "Chapter 2 Challenge real-time anomaly detection of continuous waveform by autoencoder", INTERFACE, vol. 44, no. 6, 1 June 2018 (2018-06-01), JP , pages 117 - 125, XP009530559, ISSN: 0387-9569 *
YUKI ABE, TAKUMA SENO, SHOYA MATSUMORI, MICHITA IMAI: "3Rin2-30: Generative Adversarial Networks toward Representation Learning for Image Captions", 2019 PROCEEDINGS OF THE 33RD ANNUAL CONFERENCE OF JSAI, 4 June 2019 (2019-06-04), JP, pages 1 - 4, XP009530555, DOI: 10.11517/pjsai.JSAI2019.0_3Rin230 *

Also Published As

Publication number Publication date
JPWO2021161823A1 (fr) 2021-08-19

Similar Documents

Publication Publication Date Title
US20190304079A1 (en) Unsupervised neighbor-preserving embedding for image stream visualization and anomaly detection
JP7074460B2 (ja) 画像検査装置および方法
JP4183185B2 (ja) 診断装置、検出装置、制御方法、検出方法、プログラム、及び記録媒体
US11087452B2 (en) False alarm reduction system for automatic manufacturing quality control
KR20200039049A (ko) 외관불량 검사방법 및 외관불량 검사 시스템
CN110619618A (zh) 一种表面缺陷检测方法、装置及电子设备
CN110648290A (zh) 一种基于sure参数优化的双核非局部均值图像去噪方法
KR101992970B1 (ko) 딥러닝과 노이즈 제거 기술을 이용한 표면 결함 검출 장치 및 방법
CN111079539B (zh) 一种基于异常追踪的视频异常行为检测方法
CN112037166A (zh) 一种表面缺陷检测方法及检测装置
JP7426261B2 (ja) 学習装置、画像検査装置、学習済みパラメータ、学習方法、および画像検査方法
US20220103869A1 (en) Techniques for limiting the influence of image enhancement operations on perceptual video quality estimations
CN114862857B (zh) 一种基于两阶段学习的工业产品外观异常检测方法及系统
JP7392488B2 (ja) 遺留物誤検出の認識方法、装置及び画像処理装置
WO2021161823A1 (fr) Dispositif de détermination, procédé de génération de modèle de restauration, et programme informatique
CN116310473A (zh) 一种基于误差缓解的量子神经网络的图像分类方法
CN117078679B (zh) 一种基于机器视觉的散热风扇自动组装线生产检测方法
CN114707551A (zh) 面向工业互联网设备故障诊断的多模态数据融合方法
JP7459697B2 (ja) 異常検知システム、学習装置、異常検知プログラム、学習プログラム、異常検知方法、および学習方法
US11508039B2 (en) Image processing apparatus, image processing method, and storage medium
JP2022029262A (ja) 画像処理装置、画像処理方法、画像処理プログラム、および学習装置
CN112767331B (zh) 基于零样本学习的图像异常检测方法
CN116796187A (zh) 一种输电线局部放电检测方法
CN115641611A (zh) 人体姿态估计方法、装置、设备及存储介质
CN113487571A (zh) 一种基于图像质量评价的自监督异常检测方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21753991

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 2022500323

Country of ref document: JP

Kind code of ref document: A

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21753991

Country of ref document: EP

Kind code of ref document: A1