ABNORMALITY DETECTION SYSTEM AND ABNORMALITY DETECTION METHOD

The abnormality determination system uses low-rank approximation and machine learning to generate estimation images for accurate abnormality detection in industrial products, effectively identifying local defects.

JP7586684B2Active Publication Date: 2024-11-19株式会社的
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
JP2020178009
Authority / Receiving Office
JP · JP
Patent Type
Patents
Current Assignee / Owner
Filing Date
2020-10-23
Publication Date
2024-11-19
Estimated Expiration
2040-10-23

AI Technical Summary

Technical Problem

Existing systems lack the ability to accurately determine local abnormalities in objects based on images, particularly in industrial products like wire harnesses.

Method used

An abnormality determination system that utilizes low-rank approximation through singular value decomposition on captured images, combined with machine learning to generate an estimation image and compare it with the original image for abnormality detection.

Benefits of technology

Enables accurate identification of local abnormalities by generating a reproducible estimation image for normal objects and highlighting noise in abnormal objects, allowing precise determination of abnormalities.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 0007586684000002
    Figure 0007586684000002
  • Figure 0007586684000003
    Figure 0007586684000003
  • Figure 0007586684000004
    Figure 0007586684000004
Patent Text Reader

Abstract

To provide an abnormality determination system and an abnormality determination method capable of appropriately determining the presence or absence of an abnormality in an object to be inspected.SOLUTION: The abnormality determination system includes: an image processing unit that generates an input image G1 by performing low-rank approximation based on singular value decomposition on an original image G0 in which an object to be inspected is captured; an image generation unit 4 that estimates an image before the low-rank approximation has been performed from the input image based on a result of machine learning and generates an image G2 of an estimation result; and a data providing unit that provides at least the image of the estimation result as data for determining the abnormality of the object to be inspected.SELECTED DRAWING: Figure 3
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] The present invention relates to an abnormality determination system and an abnormality determination method. [Background technology]

[0002] Conventionally, there are systems and methods for determining anomalies. Patent Document 1 discloses a technology for an anomaly detection method using a neural network, in which a fourth piece of data, the normality or abnormality of which is unknown, is input to the neural network, and the fourth piece of data is determined to be normal or abnormal. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2019-61577 A Summary of the Invention [Problem to be solved by the invention]

[0004] There is a demand for a technique that can appropriately determine whether an object to be inspected has an abnormality based on an image of the object to be inspected. For example, it is preferable to be able to determine whether or not a local abnormality exists in the object to be inspected.

[0005] An object of the present invention is to provide an abnormality determination system and an abnormality determination method that can appropriately determine the presence or absence of an abnormality in an object to be inspected. [Means for solving the problem]

[0006] The anomaly determination system of the present invention is characterized by comprising an image processing unit that performs low-rank approximation using singular value decomposition on an original image of an object to be inspected to generate an input image, an image generation unit that estimates an image before the low-rank approximation is performed from the input image based on the results of machine learning, and generates an image of the estimated result, and a data providing unit that provides at least the image of the estimated result as data for determining an abnormality in the object to be inspected. Effect of the Invention

[0007] The anomaly determination system according to the present invention has a data providing unit that provides at least an image of an estimated result as data for determining an anomaly in an object to be inspected. If the object to be inspected is normal, an image of an estimated result with high reproducibility is obtained. On the other hand, if an anomaly exists in the object to be inspected, the image of the estimated result will have low reproducibility. The anomaly determination system according to the present invention has the effect of being able to appropriately determine whether or not an anomaly exists in the object to be inspected. [Brief description of the drawings]

[0008] [Figure 1] FIG. 1 is a block diagram of an abnormality determination system according to an embodiment. [Diagram 2] FIG. 2 is a diagram showing the flow of image processing in the embodiment. [Diagram 3] FIG. 3 is a diagram showing an example of an image generated when an abnormality exists in an object to be inspected. [Figure 4] FIG. 4 is a diagram illustrating machine learning according to the embodiment. [Diagram 5] FIG. 5 is a diagram showing evaluation data of the examples. [Figure 6] FIG. 6 shows test data and result images of the embodiment. [Figure 7] FIG. 7 is a diagram showing an original image, an input image, and an estimated image for the third test data. [Figure 8] FIG. 8 is a diagram showing the first test data and the third test data. [Figure 9] FIG. 9 is a diagram showing the fifth result image and the sixth result image. [Figure 10] FIG. 10 is a diagram showing an example of a method for dividing an image. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0009] Hereinafter, an abnormality determination system and an abnormality determination method according to an embodiment of the present invention will be described in detail with reference to the drawings. Note that the present invention is not limited to the embodiment. In addition, the components in the following embodiments include those that a person skilled in the art can easily imagine or that are substantially the same.

[0010] [Embodiment] An embodiment will be described with reference to Fig. 1 to Fig. 10. This embodiment relates to an anomaly determination system and an anomaly determination method. Fig. 1 is a block diagram of an anomaly determination system according to an embodiment, Fig. 2 is a diagram showing a flow of image processing in the embodiment, Fig. 3 is a diagram showing an example of an image generated when an abnormality exists in an object to be inspected, Fig. 4 is a diagram explaining machine learning of an embodiment, Fig. 5 is a diagram showing evaluation data of an embodiment, Fig. 6 is a diagram showing test data and a result image of an embodiment, Fig. 7 is a diagram showing an original image, an input image, and an image of an estimated result for the third test data, Fig. 8 is a diagram showing the first test data and the third test data, Fig. 9 is a diagram showing the fifth result image and the sixth result image, and Fig. 10 is a diagram showing an example of an image division method.

[0011] As shown in FIG. 1, the abnormality determination system 1 according to this embodiment has an abnormality determination device 10. The abnormality determination device 10 has an acquisition unit 2, an image processing unit 3, an image generation unit 4, a data provision unit 5, a determination unit 6, and an output unit 7. The abnormality determination device 10 determines the presence or absence of an abnormality in an object to be inspected based on an image captured by a camera 20. The object to be inspected is, for example, an industrial product. The abnormality determination device 10 is placed, for example, in an inspection process in a factory that manufactures products. In this embodiment, the abnormality determination system 1 will be described for the case where the object to be inspected is a wire harness WH.

[0012] The camera 20 captures an image of an object to be inspected to generate an original image G0. The original image G0 is an image of the entire object to be inspected or a part of the object to be inspected. The camera 20 generates a color image having, for example, three colors of RGB.

[0013] An overview of the anomaly determination system 1 according to this embodiment will be described with reference to Fig. 2. Fig. 2 shows an original image G0, an input image G1, and an estimation result image G2. The original image G0 is an image of an object to be inspected, and is generated by a camera 20. The input image G1 is an image generated by performing low-rank approximation by singular value decomposition on the original image G0. In other words, the input image G1 is an image blurred by singular value decomposition.

[0014] The estimated image G2 is an image generated by the artificial intelligence based on the input image G1. The image generation unit 4, which is the artificial intelligence, estimates the image before the low-rank approximation is performed based on the result of machine learning, and generates the estimated image G2. In other words, the image generation unit 4 restores the image before the low-rank approximation is performed based on the input image G1. In the anomaly determination system 1, whether or not an abnormality exists in the object to be inspected is determined based on the original image G0 and the estimated image G2.

[0015] If the difference between the original image G0 and the estimated image G2 is large, it is determined that an abnormality exists in the object being inspected. On the other hand, if the difference between the original image G0 and the estimated image G2 is small, it is determined that an abnormality does not exist in the object being inspected.

[0016] In the original image G0 in FIG. 3, an abnormality exists in the wire harness WH, which is the object to be inspected. Specifically, the color of the coating of the electric wire W1 is different from the color of the normal electric wire W0 (see FIG. 2). In other words, the wrong electric wire W1 is connected to the connector 30. In this case, as shown in FIG. 3, noise occurs in the image G2 of the estimation result. For example, disturbance occurs in the area A4 in the lower right of the image G2 of the estimation result. The occurrence of noise in the image G2 of the estimation result allows it to be determined that an abnormality exists in the wire harness WH. According to the abnormality determination system 1 of this embodiment, the presence or absence of an abnormality in the object to be inspected can be appropriately determined.

[0017] The details of the abnormality determination system 1 will be described. The abnormality determination device 10 shown in FIG. 1 is configured by, for example, a computer having a memory, an arithmetic unit, a communication interface, and the like. The acquisition unit 2 acquires the original image G0 generated by the camera 20 via the communication interface. The acquisition unit 2 may acquire the original image G0 by communicating with the camera 20, or may acquire the original image G0 recorded in a server or the like. The original image G0 acquired by the acquisition unit 2 is stored in a memory or the like.

[0018] The image processing unit 3 generates an input image G1 by performing a low-rank approximation by singular value decomposition on the original image G0. As shown in the following [Equation 1], the image processing unit 3 decomposes the matrix A of the original image G0 into three matrices.

Equation

[0019] [Equation 1] The matrix S is a matrix in which the values of the diagonal components are the singular values σ1, σ2,..., σ i and the values of the elements other than the diagonal components are 0. The singular value σ i (i = 1, 2,..., r) satisfies the relationship of σ1 ≧ σ2 ≧... ≧ σ r . The image processing unit 3 sets an integer q (where q < r) and performs a low-rank approximation that sets the singular values σ q+1 and later to 0. Note that the image processing unit 3 performs the approximation without reducing the resolution of the image. The value of the integer q is determined so that when the input image G1 is viewed in pixel units, portions where the color and luminance vary remain. In the embodiments described later, q = 1. However, the value of q is not limited to 1. The value of q may be selected, for example, from the range of 1 or more and 5 or less. The inventor of the present application has confirmed that even when the value of q is 4 or 5, the presence or absence of an abnormality can be determined in the same manner as when q = 1.

[0020] ​​The image generating unit 4 estimates an image before low-rank approximation is performed from the input image G1, and generates an image G2 of the estimated result. The image generating unit 4 generates the image by machine learning, which will be described with reference to FIG. 4. FIG. 4 shows a first image T1 and a second image T2. The first image T1 is an image captured of a normal wire harness WH. The second image T2 is an image generated by performing low-rank approximation on the first image T1.

[0021] The image generating unit 4 generates the image by machine learning using the second image T2 as input data and the first image T1 as correct answer data. The machine learning method used is, for example, a convolutional neural network (CNN). The image generating unit 4 performs machine learning using a pair of the first image T1 and the second image T2 as teacher data so that the first image T1 can be estimated based on the second image T2.

[0022] The data providing unit 5 provides data for determining an abnormality of the object to be inspected. More specifically, the data providing unit 5 provides at least the image G2 of the estimation result. The data providing unit 5 of this embodiment provides a data set combining the original image G0 and the image G2 of the estimation result to the determination unit 6. The data providing unit 5 may provide data for determining an abnormality to an external device. The data providing unit 5 provides data to, for example, the monitor 50. The monitor 50 displays an image of the provided data. The data providing unit 5 may provide both the original image G0 and the image G2 of the estimation result to the monitor 50, or may provide only the image G2 of the estimation result to the monitor 50. The operator performs an abnormality determination based on the image displayed on the monitor 50.

[0023] The determination unit 6 determines the presence or absence of an abnormality in the object to be inspected based on the difference between the original image G0 and the estimated image G2. In this embodiment, the determination unit 6 determines the presence or absence of an abnormality using a peak signal-to-noise ratio (PSNR) as an evaluation index. The determination unit 6 calculates a PSNR value using the original image G0 as the original image and the estimated image G2 as the converted image. The PSNR value indicates the degree of reproducibility of the estimated image G2 relative to the original image G0. For example, when the PSNR value is less than a threshold, the determination unit 6 determines that an abnormality exists in the object to be inspected captured in the original image G0. The determination unit 6 provides the determination result to the output unit 7.

[0024] The output unit 7 outputs the determination result by the determination unit 6. The abnormality determination device 10 displays the determination result on, for example, a monitor 50. The abnormality determination device 10 may notify the determination result by a lamp, a sound, or the like.

[0025] [Example] An example will be described. In the example, it was verified whether the judgment unit 6 can appropriately judge the presence or absence of an abnormality using evaluation data and test data. In the example, the resolution of each image is 256 pixels x 256 pixels. The resolution of the first image T1 and the second image T2 for learning is also the same. The image generation unit 4 of the example was generated by performing machine learning using 100 first images T1 and 100 second images T2 as teacher data.

[0026] FIG. 5 shows the evaluation data VD. The evaluation data VD is an image of a normal wire harness WH, and is not used as the teacher data of the image generating unit 4. The evaluation data VD is an image captured under the same conditions as when the first image T1 for learning was captured. Note that some images randomly extracted from a plurality of images captured under the same conditions may be the first image T1, and other images may be the evaluation data VD. In the embodiment, 20 input images G1 were generated using 20 different evaluation data VD as the original image G0. The image generating unit 4 generates 20 estimated image G2 from the 20 input images G1. Thus, 20 combinations of the original image G0 and the estimated image G2 are obtained. The determining unit 6 calculated the PSNR value for each of the 20 pairs of the original image G0 and the estimated image G2. As a result, in the evaluation data, the maximum PSNR was 21.15 [dB] and the minimum was 19.30 [dB].

[0027] 6, the test data includes first test data D1, second test data D2, third test data D3, and fourth test data D4. Each of the test data D1, D2, D3, and D4 is generated by capturing an image of the same wire harness WH as the wire harness WH captured as the teacher data.

[0028] The first test data D1 is an image captured in the same manner as the first image T1 of the teacher data. That is, the first test data D1 is image data similar to the evaluation data VD. The second test data D2 is an image of the wire harness WH captured by turning the connector 30 upside down. In other words, the second test data D2 is an image of the connector 30 captured from the back side.

[0029] The third test data D3 is an image of an end of the wire harness WH opposite to the connector 30. The wire harness WH has an electric wire W, a connector 30 connected to one end of the electric wire W, and a connector 40 connected to the other end of the electric wire W. The second test data D2 is an image of the front surface of the connector 40. The fourth test data D4 is an image of the wire harness WH with the connector 40 turned over. In other words, the fourth test data D4 is an image of the connector 40 taken from the back side.

[0030] In the embodiment, the PSNR was calculated using 20 pieces of each test data D1, D2, D3, and D4. For example, an input image G1 is generated using 20 pieces of first test data D1 as an original image G0. Then, based on the 20 pieces of input images G1, 20 pieces of estimated result images G2 are generated by the image generating unit 4. FIG. 6 shows a first result image R1, a second result image R2, a third result image R3, and a fourth result image R4. The first result image R1 is the estimated result image G2 corresponding to the first test data D1. Similarly, the result images R2, R3, and R4 are the estimated result images G2 corresponding to the test data D2, D3, and D4.

[0031] The determination unit 6 calculates PSNR for 20 pairs of the first test data D1 and the first result image R1. Similarly, 20 PSNR values ​​are calculated for each of the test data D2, D3, and D4.

[0032] For the first test data D1, the maximum value of PSNR was 21.01 [dB] and the minimum value was 19.32 [dB]. These values ​​are the same as the PSNR value for the evaluation data. For the test data D2, D3, and D4, the maximum value of PSNR was 19.17 [dB] and the minimum value was 17.64 [dB]. In other words, the distribution range of PSNR values ​​for the evaluation data and the first test data D1 (19.30 to 21.15) is discontinuous with the distribution range of PSNR values ​​for the test data D2, D3, and D4 (17.64 to 19.17). Therefore, the determination unit 6 can distinguish between an image of a normal wire harness WH and an image of a wire harness WH having an abnormality.

[0033] FIG. 7 shows the original image G0, the input image G1, and the estimated image G2 for one of the third test data D3. In the estimated image G2, the entire image is distorted. In other words, the estimated image G2 contains noise that is clearly visible to the naked eye. The PSNR value for this data was 18.27 dB.

[0034] The above-described abnormality detection can be applied to, for example, detection of incorrect insertion of the wire harness WH. For example, when the connector 30 is connected to a counterpart device in an upside-down state, an image similar to the second test data D2 is captured by the camera 20. When the connector 40 is connected to a location where the connector 30 should be connected, an image similar to the third test data D3 or the fourth test data D4 is captured by the camera 20.

[0035] In the images of the test data D2, D3, and D4, localized abnormalities occur in various parts compared to the image of the first test data D1. FIG. 8 shows an image of the first test data D1 and an image of the third test data D3. As can be seen from FIG. 8, the position and color of each electric wire in the image of the third test data D3 are different from the position and color of each electric wire in the image of the first test data D1. In other words, the image of the third test data D3 is an image including multiple localized abnormalities. Therefore, the image generating unit 4 cannot accurately restore the image of the third test data D3, and noise generated in the image G2 of the estimation result becomes large as shown in FIG. 7.

[0036] In this way, the abnormality determination device 10 of the present embodiment can determine that an object has an abnormality based on an image of the object having a local abnormality. Similarly, the abnormality determination device 10 can accurately determine whether an object has a local abnormality for the second test data D2 and the fourth test data D4. Therefore, the abnormality determination device 10 can accurately detect a misconnection of the connectors 30, 40. In other words, the abnormality determination device 10 can appropriately determine whether a local abnormality exists in the wire harness WH.

[0037] Next, an example of determining an abnormality in a portion near the connector 30 will be described. As described above, in the original image G0 of FIG. 3, the color of the coating of the electric wire W1 is different from that of the normal electric wire W0 (see FIG. 2). The color of the electric wire W0 is green, and the color of the electric wire W1 is yellow. In the wire harness WH of FIG. 3, the electric wires other than the electric wire W1 are normal electric wires. That is, in the wire harness WH of FIG. 3, there is a local abnormality in that one electric wire W1 is different from the original electric wire W0. In this case, as shown in FIG. 3, noise that can be recognized at a glance occurs in the image G2 of the estimation result. Therefore, it is possible to determine a local abnormality based on the difference between the original image G0 and the image G2 of the estimation result.

[0038] FIG. 9 shows a fifth result image R5 and a sixth result image R6. The fifth result image R5 is the image G2 of the estimation result shown in FIG. 2, and the sixth result image R6 is the image G2 of the estimation result shown in FIG. 3. In the sixth result image R6, the shape of the electric wire is disturbed, and the disturbance of the shape of the electric wire is particularly noticeable in the lower right region A4. Comparing the sixth result image R6 shown in FIG. 9 with the third result image R3 shown in FIG. 7, the range in which the disturbance occurs is wide in the third result image R3. In the third result image R3, the entire image is disturbed, and it is considered that the noise is large. Therefore, the sixth result image R6 has a larger PSNR value than the third result image R3, and the accuracy of anomaly detection based on PSNR may be reduced.

[0039] The determination of an abnormality based on the original image G0 and the estimated image G2 may be performed, for example, by visual inspection by an operator. In this case, the abnormality determination device 10 provides the operator with a data set combining the original image G0 and the estimated image G2. The provided original image G0 and estimated image G2 are displayed side by side on the monitor 50, for example. The operator compares the original image G0 and the estimated image G2 to determine whether or not an abnormality exists.

[0040] The abnormality determination device 10 may request a final determination by an operator when the PSNR value is within a predetermined range. For example, when the PSNR value is higher than a threshold value for determining an abnormality and lower than a reference value for determining normality, the abnormality determination device 10 may request an operator to make a visual determination. In this manner, the burden on the operator regarding the determination work is reduced. For example, the reference value is statistically determined based on the distribution of PSNR values ​​obtained from a plurality of evaluation data. The reference value may be the median or average value of the PSNR values ​​obtained from the evaluation data.

[0041] The abnormality determination device 10 may calculate the PSNR for each region of the image. For example, in the sixth result image R6 shown in Fig. 9, image disturbance is noticeable in the lower right region A4, which corresponds to the fourth quadrant. This region is the region where the abnormal electric wire W1 is captured. In other words, when a local abnormality exists, it is considered that noise becomes large in the region where the abnormality is captured.

[0042] The abnormality determination device 10 may perform abnormality determination by dividing the estimated image G2 into a plurality of regions. For example, the estimated image G2 is divided into four regions A1, A2, A3, and A4 as shown in FIG. 10. The dividing lines are, for example, the vertical center line of the image and the horizontal center line of the image. The original image G0 is also divided into four regions in the same manner. The determination unit 6 calculates the PSNR value for each of the regions A1, A2, A3, and A4 by comparing them with the corresponding region of the original image G0. The determination unit 6 determines that the object to be inspected has an abnormality when, for example, at least one of the four calculated PSNR values ​​is less than a threshold value.

[0043] As described above, the anomaly determination system 1 according to this embodiment includes an image processing unit 3, an image generating unit 4, and a data providing unit 5. The image processing unit 3 performs low-rank approximation by singular value decomposition on an original image G0 in which an object to be inspected is captured, to generate an input image G1. The image generating unit 4 estimates an image before the low-rank approximation is performed from the input image G1 based on the result of machine learning, and generates an image G2 of the estimated result. The data providing unit 5 provides at least the image G2 of the estimated result as data for determining an abnormality of the object to be inspected. According to the anomaly determination system 1 of this embodiment, an abnormality of the object to be inspected can be appropriately determined based on the image G2 of the estimated result.

[0044] The anomaly determination system 1 of this embodiment further includes a determination unit 6. The determination unit 6 determines the presence or absence of an anomaly in the object to be inspected based on the difference between the original image G0 and the estimated image G2. The data providing unit 5 provides the original image G0 and the estimated image G2 to the determination unit 6. An index indicating the difference between the original image G0 and the estimated image G2 is, for example, PSNR, but is not limited to this. According to the anomaly determination system 1 of this embodiment, it is possible for the determination unit 6 to automatically determine an anomaly. The determination unit 6 can determine the presence or absence of a local anomaly in the object to be inspected.

[0045] The image generating unit 4 of this embodiment is generated by machine learning using the first image T1 as correct answer data and the second image T2 as input data. The first image T1 is an image captured of a normal object to be inspected. The second image T2 is an image generated by performing low-rank approximation on the first image T1. Therefore, the image generating unit 4 can accurately estimate an image before low-rank approximation is performed from the input image G1 for a normal object to be inspected. On the other hand, the image generating unit 4 cannot accurately estimate an image before low-rank approximation is performed from the input image G1 for an object to be inspected that has an abnormality. As a result, the presence or absence of an abnormality in the object to be inspected can be easily determined.

[0046] The anomaly determination method according to this embodiment includes the steps of generating an input image G1, generating an image G2 of an estimated result, and determining whether or not there is an anomaly. In the step of generating the input image G1, low-rank approximation using singular value decomposition is performed on an original image G0 in which an object to be inspected is captured, to generate the input image G1. In this embodiment, the image processing unit 3 executes the step of generating the input image G1.

[0047] In the step of generating an image G2 of the estimated result, an image before low-rank approximation is performed is estimated from the input image G1 based on the result of machine learning, and an image G2 of the estimated result is generated. In this embodiment, the image generating unit 4 executes the step of generating the image G2 of the estimated result. In the step of determining the presence or absence of an abnormality, the presence or absence of an abnormality in the object to be inspected is determined based on the image G2 of the estimated result. The step of determining the presence or absence of an abnormality is executed by, for example, the determining unit 6. In this case, in the step of determining the presence or absence of an abnormality, the presence or absence of an abnormality is determined based on the difference between the original image G0 and the image G2 of the estimated result.

[0048] The step of determining whether or not there is an abnormality may be executed by an operator. In this case, the presence or absence of an abnormality may be determined based on the image G2 of the estimation result without using the original image G0. The operator can determine whether or not there is an abnormality based on disturbances in the image G2 of the estimation result. In the step of determining whether or not there is an abnormality, the operator may visually compare the original image G0 with the image G2 of the estimation result to determine whether or not there is an abnormality.

[0049] [Modifications of the embodiment] The object to be inspected is not limited to an industrial product. The abnormality determination system 1 of the present embodiment may be applied to, for example, the inspection of food. The abnormality determination system 1 can determine the presence or absence of local abnormalities in various objects.

[0050] The camera 20 may capture an image of an object to be inspected and generate a monochrome image. An abnormality in the object to be inspected appears as an abnormality in the luminance distribution in the original image G0. Therefore, the abnormality determination system 1 can appropriately determine the presence or absence of an abnormality in the object to be inspected based on the monochrome image.

[0051] The contents disclosed in the above-mentioned embodiments and modifications can be implemented in appropriate combinations. [Explanation of symbols]

[0052] 1. Anomaly Judgment System 2 Acquisition part 3. Image Processing Section 4. Image Generation Section 5 Data Provision Department 6 Judgment section 7 Output section 10 Abnormality determination device 20 Camera 30,40 Connector 50 Monitors WH Wire harness A1,A2,A3,A4 area D1: First test data, D2: Second test data, D3: Third test data D4: Fourth test data G0: original image, G1: input image, G2: estimated image R1: First result image, R2: Second result image, R3: Third result image R4: Fourth result image, R5: Fifth result image, R6: Sixth result image T1: First image, T2: Second image Data for VD evaluation W0,W1 Electric wire

Claims

1. an image processing unit that performs low-rank approximation using singular value decomposition on an original image of an object to be inspected to generate an input image; an image generation unit that estimates an image before the low-rank approximation is performed from the input image based on a result of machine learning, and generates an image of the estimation result; a data providing unit that provides at least an image of the estimation result as data for determining an abnormality in the object to be inspected; An abnormality determination system comprising:

2. a determination unit that determines the presence or absence of an abnormality in the object to be inspected based on a difference between the original image and the image of the estimation result, The data providing unit provides the original image and the image of the estimation result to the determining unit. The abnormality determination system according to claim 1 .

3. The image generating unit generates an image by machine learning using a first image obtained by capturing an image of the object to be inspected, which is normal, as correct answer data and a second image generated by performing the low-rank approximation on the first image as input data. The abnormality determination system according to claim 1 .

4. A step of performing low-rank approximation using singular value decomposition on an original image of an object to be inspected to generate an input image; A step of estimating an image before the low-rank approximation is performed from the input image based on a result of the machine learning, and generating an image of the estimated result; determining whether or not there is an abnormality in the object to be inspected based on an image of the estimation result; Including, In the step of determining the presence or absence of an abnormality, the presence or absence of an abnormality is determined based on a difference between the original image and the image of the estimation result, or the presence or absence of an abnormality is determined based on image distortion in the image of the estimation result. The abnormality determination method according to the present invention is characterized in that

Citation Information

Patent Citations

  • Identification processing device, identification system, identification method, and program

    JP2017097718A

  • Abnormality determination method and program

    JP2019061577A

  • Image processing apparatus, camera, movable body, and image processing method

    JP2020071532A