Insulator infrared image fuzzy edge processing method with learning inspiration capability

Abstract

The invention discloses an insulator infrared image fuzzy edge processing method with a heuristic learning capability. The method comprises the following steps: training a heuristic-constraint system and an edge processing network; obtaining an infrared image of the insulator to be detected; exporting an infrared image temperature matrix of the insulator to be measured; processing a to-be-measured temperature matrix; calculating a primary edge feature matrix of the to-be-measured temperature matrix; normalizing the double matrixes to be measured; executing the edge processing network to obtain a processing result; according to the processing method, a weighted sample-label group driven by a feature normal form and a random vector is generated through a heuristic-constraint system, then the weighted sample-label group is used for continuously training an edge processing network, and the edge processing network is used for eliminating an insulator infrared image edge fuzzy region and realizing edge detection. The method has high technical advancement and wide applicability.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence systems in the field of power production and the technical field of visual inspection of power equipment, and in particular to a method for processing fuzzy edges of insulator infrared images based on an inspiration-constraint system and having inspiration learning capabilities. Background Art

[0002] Insulators are important basic insulation components in the power generation, transmission, transformation and distribution links of the power system. They play the role of electrical insulation and mechanical support between live conductors and power towers. Their insulation effectiveness directly affects the normal operation of the power grid. As the operating time increases, due to the combined effects of electromechanical loads and environmental factors, their insulation and mechanical properties will gradually decline, leading to aging and degradation of insulators, and even serious accidents such as explosions and string drops, causing large-scale power outages, threatening the safe and stable operation of the power grid.

[0003] The traditional insulator inspection method is manual tower inspection. In order to improve inspection efficiency, reduce labor intensity, and ensure the personal safety of inspectors, infrared thermal imaging inspection is currently the most effective alternative. Automatic positioning, identification, and segmentation of insulator areas in infrared images are the key to completing intelligent detection of insulator degradation status. Edge detection is an important method for obtaining the position and geometric shape of insulator strings, and is an important prerequisite for image segmentation.

[0004] In practical applications, due to the poor contrast of infrared images and single channel features, in most cases, the edge features of insulators in infrared images appear blurred. Currently, there is no effective method to eliminate the blurred area at the edge of insulator infrared images. Summary of the invention

[0005] The purpose of the present invention is to provide an insulator infrared image fuzzy edge processing method with heuristic learning ability to solve the problems of edge loss, fuzzy area and inaccurate edge position generated during insulator string infrared image edge detection proposed in the background technology.

[0006] To achieve the above object, the present invention provides a method for processing fuzzy edges of infrared images of insulators with heuristic learning capability, which is characterized by comprising the following steps:

[0007] S1, training heuristic-constraint system and edge processing network;

[0008] S2, using an infrared image shooting device to obtain an infrared image of the insulator to be tested;

[0009] S3, deriving the infrared image temperature matrix of the insulator to be tested obtained in step S2;

[0010] S4, processing the temperature matrix to be measured derived in step S3;

[0011] S5, calculating the primary edge feature matrix of the temperature matrix to be measured;

[0012] S6, normalization of the double matrix to be tested;

[0013] S7. Execute the edge processing network to obtain the processing result.

[0014] Furthermore, the step S1 is specifically as follows: collect a certain number of infrared images of insulators as a training sample set; use an infrared image acquisition device to derive a temperature matrix of the training sample set, and process the temperature matrix of the training sample set; calculate and normalize the primary edge feature matrix of the training sample set after training, and pass the normalized primary edge feature matrix of the training sample set to a paradigm generator, and the paradigm generator provides input paradigm samples to the heuristic network and the constraint network respectively; after receiving the input paradigm samples, the heuristic network and the constraint network are trained with boundary labels and sample-label groups as convergence targets respectively; use the sample judgment network and the label judgment network to perform adversarial training with the heuristic network and the constraint network respectively, and use the heuristic network to generate a qualified sample-label group for training the edge processing network.

[0015] Furthermore, processing the temperature matrix of the training sample set includes scaling and clipping the temperature matrix, and the scaling and clipping are calculated by the following formula (1):

[0016]

[0017] In formula (1), γ is the scaling factor, l sx is the preset image width, l sy is the preset image height, l x is the width of the temperature matrix, l y is the temperature matrix height, T r is the temperature matrix after scaling, T is the temperature matrix before scaling, T f is the trimmed temperature matrix.

[0018] Furthermore, the calculation method and normalization method of the primary edge feature matrix of the training sample set are respectively as follows (2) and (3):

[0019] B f =ΔT f (2);

[0020]

[0021] In the above formulas (2) and (3), B f is the primary edge feature matrix, Δ is the Laplace operator; B c is the normalized primary edge feature matrix.

[0022] Furthermore, the paradigm generator generates a paradigm example by the following formula:

[0023] N s =N(μ,σ 2 )(4);

[0024]

[0025] S=(1-α)B c +αN f (6);

[0026] In the above formulas (4) to (6), N s is a normally distributed random number, μ is the mean of the normal distribution, σ is the standard deviation of the normal distribution, N f is the truncated output of the random number, α is the noise level, and S is the generated paradigm.

[0027] Furthermore, the heuristic network and the constraint network are trained with boundary labels and sample-label groups as convergence targets respectively. The heuristic network is an artificial neural network whose input is the feature paradigm generated by the paradigm generator and whose output is the restored normalized temperature matrix and boundary labels; the constraint network is an artificial neural network whose input is the feature paradigm generated by the paradigm generator and whose output is the boundary labels; the heuristic network and the constraint network are implemented using convolutional neural networks or variational autoencoder methods.

[0028] Furthermore, the sample judgment network and the label judgment network are used to perform adversarial training with the inspiration network and the constraint network respectively, wherein the sample judgment network is an artificial neural network whose input is the samples output by the inspiration network and the real samples, and whose output is the real probability of the samples; the label judgment network is an artificial neural network whose input is the boundary labels output by the inspiration network, the boundary labels output by the constraint network and the real boundary labels, and whose output is the real probability of the labels; the sample judgment network and the label judgment network are implemented using convolutional neural network or BP neural network method.

[0029] Furthermore, the heuristic network generates qualified sample-label groups for training the edge processing network, wherein the qualified sample-label groups are sample-label groups that are judged as failed by both the sample judgment network and the label judgment network.

[0030] Furthermore, consistent with step S1, the processing of the temperature matrix to be measured in step S4 also includes scaling and clipping of the temperature matrix, and the scaling and clipping method is also calculated by formula (1); the primary edge feature matrix of the temperature matrix to be measured is calculated in step S5, which is also calculated using formula (2).

[0031] Furthermore, the normalization of the dual matrix to be tested in step S6 includes normalization of the clipped temperature matrix and normalization of the primary edge feature matrix, and the normalization method is as follows:

[0032]

[0033] In the above formula (7), T c is the normalized temperature matrix, t a is the preset starting temperature value, t b is the preset end point temperature, B c is the normalized primary edge feature matrix.

[0034] Furthermore, in step S7, the edge processing network is executed to obtain the processing result, and the input of the edge processing network is T obtained in step S6. c With B c , which is trained by the qualified sample-label group generated by the inspiration network, and the processing result is the probability distribution matrix of the edge position of the insulator; the edge processing network is implemented using convolutional neural network and variational autoencoder method.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] The fuzzy edge processing method of the infrared image of an insulator with heuristic learning ability of the present invention generates a weighted sample-label group driven by a feature paradigm and a random vector through an heuristic-constraint system, and then continuously trains an edge processing network using the weighted sample-label group, and uses the edge processing network to eliminate the edge fuzzy area of ​​the infrared image of the insulator and realize edge detection. The method of the present invention has high technical advancement and wide applicability.

[0037] In addition to the above-described purposes, features and advantages, the present invention has other purposes, features and advantages. The present invention will be further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The accompanying drawings are used to provide a further understanding of the embodiments of the present invention and constitute a part of the specification. Together with the following specific implementations, they are used to explain the embodiments of the present invention, but do not constitute a limitation on the embodiments of the present invention. In the accompanying drawings:

[0039] Figure 1 It is a flow chart of a method for processing fuzzy edges of infrared images of insulators with heuristic learning capability according to the present invention;

[0040] Figure 2 is an infrared image of an insulator used in a preferred embodiment of the present invention;

[0041] Figure 3 This is the result obtained by the preferred embodiment of the present invention. DETAILED DESCRIPTION

[0042] The present invention will be described in detail below in conjunction with the various embodiments shown in the accompanying drawings, but it should be noted that these embodiments are not limitations of the present invention, and any equivalent transformations or substitutions in functions, methods, or structures made by ordinary technicians in the field based on these embodiments are all within the scope of protection of the present invention.

[0043] See also Figure 1 This embodiment provides a method for processing fuzzy edges of infrared images of insulators with heuristic learning capability, which is characterized by comprising the following steps:

[0044] S1. Training the heuristic-constraint system and edge processing network; the specific steps include:

[0045] A certain number of infrared images of insulators are collected as training sample sets; the temperature matrix of the training sample set is exported by using an infrared image acquisition device, and the temperature matrix of the training sample set is processed; the primary edge feature matrix of the training sample set after training is calculated and normalized, and the primary edge feature matrix of the training sample set after normalization is passed to the paradigm generator, and the paradigm generator provides input paradigm samples to the heuristic network and the constraint network respectively; after receiving the input paradigm samples, the heuristic network and the constraint network are trained with boundary labels and sample-label groups as convergence targets respectively; the sample judgment network and the label judgment network are used to perform adversarial training with the heuristic network and the constraint network respectively, and the qualified sample-label group generated by the heuristic network is used for the training of the edge processing network.

[0046] In a specific implementation, in step S1, processing the temperature matrix of the training sample set includes scaling and clipping the temperature matrix, and the scaling and clipping are calculated by the following formula (1):

[0047]

[0048] In formula (1), γ is the scaling factor, l sx is the preset image width, l sy is the preset image height, l x is the width of the temperature matrix, l y is the temperature matrix height, T r is the temperature matrix after scaling, T is the temperature matrix before scaling, T f is the temperature matrix after trimming. In this embodiment, l sx is 320, l sy 240, l x is 640, l y is 480.

[0049] In a specific implementation, the calculation method and normalization method of the primary edge feature matrix of the training sample set are respectively as follows (2) and (3):

[0050] B f =ΔT f (2);

[0051]

[0052] In the above formulas (2) and (3), B f is the primary edge feature matrix, Δ is the Laplace operator; B c is the normalized primary edge feature matrix.

[0053] In a specific implementation, the paradigm generator generates a paradigm example by the following formula:

[0054] N s =N(μ,σ 2 )(4);

[0055]

[0056] S=(1-α)B c +αN f (6);

[0057] In the above formulas (4) to (6), N s and N s is a normally distributed random number, μ is the mean of the normal distribution, σ is the standard deviation of the normal distribution, N f is the truncated output of the random number, α is the noise level, and S is the generated pattern. In this embodiment, α is 0.3, μ is 0.5, and σ is 0.4.

[0058] In a specific implementation, the heuristic network and the constraint network are trained with boundary labels and sample-label groups as convergence targets, respectively. The heuristic network is an artificial neural network whose input is the feature paradigm generated by the paradigm generator and whose output is the restored normalized temperature matrix and boundary labels; the constraint network is an artificial neural network whose input is the feature paradigm generated by the paradigm generator and whose output is the boundary labels; the heuristic network and the constraint network are implemented using convolutional neural networks or variational autoencoder methods. The sample judgment network and the label judgment network are used to perform adversarial training with the heuristic network and the constraint network respectively, wherein the sample judgment network is an artificial neural network whose input is the sample output by the heuristic network and the real sample, and whose output is the real probability of the sample; the label judgment network is an artificial neural network whose input is the boundary label output by the heuristic network, the boundary label output by the constraint network and the real boundary label, and whose output is the real probability of the label; the sample judgment network and the label judgment network are implemented using convolutional neural networks or BP neural network methods. The qualified sample-label group generated by the heuristic network is used for the training of the edge processing network, wherein the qualified sample-label group is the sample-label group judged by both the sample judgment network and the label judgment network.

[0059] S2. Use infrared image shooting equipment to obtain an infrared image of the insulator to be tested; the infrared image of the insulator to be tested is the image to be tested, and its quality should meet the requirements of relevant industry standards for infrared thermal imaging detection of insulators.

[0060] S3. Exporting the infrared image temperature matrix of the insulator to be tested; specifically, the exporting of the infrared image temperature matrix of the insulator to be tested can be achieved through the relevant extended functions provided by the infrared image shooting device.

[0061] S4, processing the temperature matrix of the infrared image of the insulator to be tested derived in step S3, specifically including matrix scaling and cropping, and the specific processing method also adopts the following formula (1):

[0062]

[0063] In formula (1), γ is the scaling factor, l sx is the preset image width, l sy is the preset image height, l x is the width of the temperature matrix, l y is the temperature matrix height, T r is the temperature matrix after scaling, T is the temperature matrix before scaling, T f is the temperature matrix after trimming. In this embodiment, l sx is 320, l sy 240, l x is 640, l y is 480.

[0064] S5, calculating the primary edge feature matrix of the temperature matrix to be measured; specifically, calculating the primary edge feature matrix of the cropped temperature matrix in step S4, and the primary edge feature matrix calculation method also adopts the following formula:

[0065] B f =ΔT f (2);

[0066] In the above formula, B f is the primary edge feature matrix, and Δ is the Laplacian operator.

[0067] S6, normalizing the dual matrix to be tested; specifically including normalizing the cropped temperature matrix described in step S5 and normalizing the primary edge feature matrix, the normalization method is as follows:

[0068]

[0069] In the above formula (7), T c is the normalized temperature matrix, t a is the preset starting temperature value, t b is the preset end point temperature, B c is the normalized primary edge feature matrix. In this embodiment, t a 5℃, t b is 45℃.

[0070] S7, execute the edge processing network to obtain the processing result, wherein the input of the edge processing network is T obtained in step S6 c With B c , which is trained by the qualified sample-label group generated by the inspiration network, and the processing result is the probability distribution matrix of the edge position of the insulator; the edge processing network is implemented using methods such as convolutional neural network and variational autoencoder.

[0071] The above steps of the embodiment of the present invention are used to Figure 2 The infrared image of the insulator in the image is processed, and the edge processing result is as follows Figure 3 shown.

[0072] In the present invention, each artificial neural network is not limited to the several methods mentioned in the present invention, and other methods that meet the functional requirements of the present invention can be applied.

[0073] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for processing fuzzy edges of infrared images of insulators with heuristic learning capability, characterized in that: The steps include: S1, training heuristic-constraint system and edge processing network; S2, using an infrared image shooting device to obtain an infrared image of the insulator to be tested; S3, deriving the temperature matrix of the infrared image of the insulator to be tested obtained in step S2; S4, processing the temperature matrix to be measured derived in step S3; S5, calculating the primary edge feature matrix of the temperature matrix to be measured; S6, normalization of the double matrix to be tested; S7. Execute the edge processing network to obtain the processing result.

2. The processing method according to claim 1, characterized in that: The step S1 specifically includes: collecting a certain number of infrared images of insulators as a training sample set; using an infrared image acquisition device to derive a temperature matrix of the training sample set, and processing the temperature matrix of the training sample set; calculating and normalizing the primary edge feature matrix of the training sample set after training, and passing the normalized primary edge feature matrix of the training sample set to a paradigm generator, and the paradigm generator provides input paradigm samples to the heuristic network and the constraint network respectively; after receiving the input paradigm samples, the heuristic network and the constraint network are trained with boundary labels and sample-label groups as convergence targets respectively; using the sample judgment network and the label judgment network to perform adversarial training with the heuristic network and the constraint network respectively, and the qualified sample-label group generated by the heuristic network is used for training the edge processing network.

3. The processing method according to claim 2, characterized in that: Processing the temperature matrix of the training sample set includes scaling and clipping the temperature matrix. The scaling and clipping are calculated by the following formula (1): In formula (1), γ is the scaling factor, l sx is the preset image width, l sy is the preset image height, l x is the temperature matrix width, l y is the temperature matrix height, T r is the temperature matrix after scaling, T is the temperature matrix before scaling, T f is the trimmed temperature matrix.

4. The processing method according to claim 3, characterized in that: The calculation method and normalization method of the primary edge feature matrix of the training sample set are respectively as follows (2) and (3): B f =ΔT f (2); In the above formulas (2) and (3), B f is the primary edge feature matrix, Δ is the Laplace operator; B c is the normalized primary edge feature matrix.

5. The processing method according to claim 2, characterized in that: The paradigm generator generates paradigm samples by the following formula: N s =N(μ,σ 2 )(4); S=(1-α)B c +αN f (6); In the above formulas (4) to (6), N s is a normally distributed random number, μ is the mean of the normal distribution, σ is the standard deviation of the normal distribution, N f is the truncated output of the random number, α is the noise level, and S is the generated paradigm.

6. The processing method according to claim 2, characterized in that: The heuristic network and the constraint network are trained with boundary labels and sample-label groups as convergence targets respectively. The heuristic network is an artificial neural network whose input is the feature paradigm generated by the paradigm generator and whose output is the restored normalized temperature matrix and boundary labels; the constraint network is an artificial neural network whose input is the feature paradigm generated by the paradigm generator and whose output is the boundary labels; the heuristic network and the constraint network are implemented using convolutional neural networks or variational autoencoder methods.

7. The processing method according to claim 2, characterized in that: The sample judgment network and the label judgment network are used to perform adversarial training with the inspiration network and the constraint network respectively, wherein the sample judgment network is an artificial neural network whose input is the sample output by the inspiration network and the real sample, and whose output is the real probability of the sample; the label judgment network is an artificial neural network whose input is the boundary label output by the inspiration network, the boundary label output by the constraint network and the real boundary label, and whose output is the real probability of the label; The sample judgment network and the label judgment network are implemented using convolutional neural network or BP neural network method.

8. The processing method according to claim 2, characterized in that: The qualified sample-label group generated by the inspiration network is used for training the edge processing network, wherein the qualified sample-label group is the sample-label group that both the sample judgment network and the label judgment network fail to judge.

9. The processing method according to claim 2, characterized in that: The normalization of the dual matrix to be tested in step S6 includes normalization of the clipped temperature matrix and normalization of the primary edge feature matrix, and the normalization method is as follows: In the above formula (7), T c is the normalized temperature matrix, t a is the preset starting temperature value, t b is the preset end point temperature, B c is the normalized primary edge feature matrix.

10. The processing method according to claim 9, characterized in that: In step S7, the edge processing network is executed to obtain the processing result, and the input of the edge processing network is T obtained in step S6. c With B c , which is trained by the qualified sample-label group generated by the inspiration network, and the processing result is the probability distribution matrix of the edge position of the insulator; the edge processing network is implemented using convolutional neural network and variational autoencoder method.

Images