Insulator Detection Method Based on Background Classification and Transfer Learning

Abstract

The present invention relates to the technical field of power inspection image processing, and discloses an insulator detection method based on background classification and transfer learning. By constructing an insulator detection data set, the insulator detection data set is preprocessed. A single-stage object detector is used to perform transfer learning training on the preprocessed insulator detection data set to obtain an initial insulator detection model. The insulator detection data set is divided into multiple data sets according to background categories, and the corresponding background categories are labeled in each data set. The initial insulator detection model is used to perform transfer learning training on each data set respectively to obtain insulator object detection models under different background categories. The accuracy rates of the insulator object detection models under different background categories for detecting the corresponding data sets are calculated until the accuracy rates of insulator detection under each background reach the set values, thereby improving the accuracy rate of insulator detection.

Description

Technical Field

[0001] The present invention relates to the technical field of power inspection image processing, and particularly to an insulator detection method based on background classification and transfer learning. Background Art

[0002] The inspection of distribution network lines is an effective means to ensure the safe and reliable operation of the power system. In distribution network lines, insulators are widely used devices with dual functions of electrical insulation and mechanical support. The condition monitoring of insulators is one of the most important and difficult tasks in the inspection of transmission lines. Traditional manual inspection is time-consuming and laborious, and can be replaced by drone inspection to achieve more automated and efficient inspection. However, the aerial images taken by drones contain cluttered backgrounds and various types of insulators. External interference factors such as changes in perspective, different lighting conditions, and partial occlusion make it very difficult to detect insulators. Existing detection methods mainly extract the features of aerial images through image processing means to distinguish insulators from complex backgrounds, such as color, shape, and texture features, etc. However, the accuracy of insulator detection is greatly reduced. Summary of the Invention

[0003] The present invention provides an insulator detection method based on background classification and transfer learning, which solves the technical problem of reducing the accuracy of insulator detection.

[0004] In view of this, the present invention provides an insulator detection method based on background classification and transfer learning, including the following steps:

[0005] Step 1: Obtain several distribution network inspection line images, use the labelme software to mark the positions of insulators on each distribution network inspection line image with rectangular frames, and generate annotation files to form an insulator detection data set. The annotation files include target position coordinates and category information;

[0006] Step 2: Preprocess the insulator detection data set. The preprocessing methods include histogram equalization algorithm, image filtering algorithm, and image sharpening;

[0007] Step 3: Use a single-stage object detector to perform transfer learning training on the preprocessed insulator detection data set to obtain an initial insulator detection model. The single-stage object detector uses a ResNeSt convolutional network as the backbone network and a feature pyramid BiFPN as the feature extraction network. The ResNeSt convolutional network is pre-trained in the ImageNet deep learning network;

[0008] Step 4: Divide the insulator detection data set into multiple data sets according to background categories, and label the corresponding background categories in each data set;

[0009] Step Five: Use the initial insulator detection model to perform transfer learning training on each dataset respectively to obtain an insulator target detection model under different background categories;

[0010] Step Six: Calculate the accuracy of the insulator target detection model under different background categories in detecting the corresponding dataset. If the accuracy is lower than the preset accuracy threshold, update the network parameters of the insulator target detection model under the corresponding background category, and execute Step Five until the accuracies of all insulator target detection models reach the preset accuracy threshold, and output the best insulator target detection model.

[0011] Preferably, Step One specifically includes:

[0012] Obtain several distribution network inspection line images based on a drone;

[0013] Use the labelme software to set and load the distribution network inspection line images, use a rectangular box to frame and label the insulators in the distribution network inspection line images, generate an annotation file, and form an insulator detection dataset.

[0014] Preferably, the process of preprocessing the insulator detection dataset using the histogram equalization algorithm is specifically as follows:

[0015] According to the number of pixels and the gray level depth of the distribution network inspection line image before histogram equalization, calculate the frequency density of the distribution network inspection line image after histogram equalization as,

[0016]

[0017] where, H B (D) represents the frequency density of the distribution network inspection line image after histogram equalization, A0 represents the number of pixels of the distribution network inspection line image before histogram equalization, and L represents the gray level depth of the distribution network inspection line image before histogram equalization;

[0018] Let the gray value D A of the distribution network inspection line image A before histogram equalization be mapped to the gray value D B in the distribution network inspection line image after histogram equalization through the gray value transformation function, that is, D B = f(D A ), f(D A ) represents the gray value transformation function, then the number of pixels of the gray value D A in [0, D A is equal to the number of pixels of the gray value D B in [0, D B , that is,

[0019]

[0020] Solving the above equation gives the gray value transformation function f(D A ), as

[0021]

[0022] Discretize the gray value transformation function f(D A ), to obtain the discretized gray value transformation function f′(D A ). Input each pixel in the distribution network inspection line image A before histogram equalization into the discretized gray value transformation function f′(D A ) to obtain the gray value D B of the distribution network inspection line image B after histogram equalization, thereby realizing histogram equalization for each distribution network inspection line image in the insulator detection dataset. Among them, the gray value D B is

[0023]

[0024] Preferably, the process of preprocessing the insulator detection dataset using an image filtering algorithm is specifically as follows:

[0025] Filter the distribution network inspection line image in the insulator detection dataset based on a median filter. Let the pixel value at the i-th row and j-th column in the distribution network inspection line image be p(i,j), then the pixel value at this position after being filtered by the median filter is

[0026] med({p(i + x,j + y)|x ∈ {-1,0,1},y ∈ {-1,0,1}})

[0027] In the formula, med(·) is a function to take the median of the set elements, and p(i + x,j + y) is the pixel value after being filtered by the median filter;

[0028] Use a Gaussian filter to filter and denoise the distribution network inspection line image after being filtered by the median filter. The pixel value after Gaussian filtering is

[0029]

[0030] In the formula, σ is the given standard deviation of the pixel value.

[0031] Preferably, the process of preprocessing the insulator detection dataset using image sharpening is specifically as follows:

[0032] Calculate the second-order partial derivative of the pixel value of the distribution network inspection line image in the insulator detection dataset as

[0033]

[0034]

[0035] Wherein, p0(i,j) is the gray value of the pixel at the i-th row and j-th column of the image before the Laplacian operator acts on the pixel of the distribution network inspection line image;

[0036] The calculated gray value after the Laplacian operator acts on the pixel of the distribution network inspection line image is

[0037]

[0038] The calculated pixel value of the distribution network inspection line image after image sharpening is

[0039]

[0040] Wherein, k is the coefficient of the diffusion effect.

[0041] Preferably, the method further includes:

[0042] Divide the input X of the ResNeSt convolutional network into K groups of cardinality units along the dimension of the input channel, divide each group of cardinality units into R groups of sub-cardinality, and divide the input channel into G groups of sub-channels, G = KR;

[0043] Based on the ResNeSt convolutional network, each group of sub-channels is sequentially passed through a 1×1 convolutional layer and a 3×3 convolutional layer for feature extraction to obtain the feature U z , U z ∈R H×W×C , z = 1, 2, …, G, H, W, and C are the three dimensions of U z respectively;

[0044] Calculate the sum of the features extracted by each group of sub-cardinality within the k-th group of cardinality units as

[0045] Use average pooling along the dimension of the input channel to obtain the average feature s k , wherein, the c-th component of the feature s k is

[0046]

[0047] Calculate the weight of the c-th component of the i-th group of sub-cardinality within the k-th group of cardinality units as,

[0048]

[0049] Wherein, represents according to the feature s of the k-th group of cardinality units kThe weight of the i-th group of the structure for splitting the c-th component

[0050] Perform a weighted sum of the features extracted by each group of sub-bases within each base unit to obtain the features of the base unit. Then, the c-th component of the features of the k-th base unit is

[0051]

[0052] The features of each base unit are concatenated and added to the input X to obtain the features Y extracted by the ResNeSt convolutional network as

[0053] V = Concat{V 1 , V 2 , …, V K}

[0054]

[0055] In the formula, Concat{·} represents the concatenation operation, represents the transformation of the convolutional layer and pooling layer that makes the number of channels of the input X consistent with V.

[0056] Preferably, the method further includes:

[0057] Construct the loss function of the insulator target detection model using the annotation file of the distribution network inspection line image as

[0058] FL(p t ) = -(1 - p t ) γ log(p t )

[0059]

[0060] In the formula, a is the category label of the detection target, a = 1 indicates that the detected object is an insulator, a = 0 indicates a non-insulator, p is the probability that the detected object is an insulator, and γ is the given focusing parameter.

[0061] From the above technical solutions, it can be seen that the present invention has the following advantages:

[0062] The present invention preprocesses an insulator detection dataset by constructing the insulator detection dataset. Using a single-stage object detector to perform transfer learning training on the preprocessed insulator detection dataset to obtain an initial insulator detection model, the insulator detection dataset is divided into multiple datasets according to the background category, and the corresponding background category is labeled in each dataset. Using the initial insulator detection model to perform transfer learning training on each dataset respectively to obtain an insulator target detection model under different background categories, so as to achieve accurate detection of insulators under various backgrounds. Calculate the accuracy of the insulator target detection model under different background categories in detecting the corresponding dataset. If the accuracy is lower than the preset accuracy threshold, update the network parameters of the insulator target detection model under the corresponding background category, and perform transfer learning until the accuracy of insulator detection under each background reaches the set value, thereby improving the accuracy of insulator detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 FIG. is a flowchart of an insulator detection method based on background classification and transfer learning provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0064] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0065] For ease of understanding, please refer to Figure 1 , an insulator detection method based on background classification and transfer learning provided by the present invention includes the following steps:

[0066] Step 1: Obtain a plurality of distribution network inspection line images, use the labelme software to mark the position of the insulator with a rectangular box on each distribution network inspection line image, and generate a marking file to form an insulator detection dataset. The marking file includes target position coordinates and category information.

[0067] Step 2: Preprocess the insulator detection dataset. The preprocessing methods include histogram equalization algorithm, image filtering algorithm, and image sharpening.

[0068] Step 3: Use a single-stage object detector to perform transfer learning training on the preprocessed insulator detection dataset to obtain an initial insulator detection model. The single-stage object detector uses the ResNeSt convolutional network as the backbone network and the Feature Pyramid BiFPN as the feature extraction network. The ResNeSt convolutional network is pre-trained in the ImageNet deep learning network.

[0069] It can be understood that by using the Feature Pyramid BiFPN as the feature extraction network, multi-scale features of the insulator can be extracted, thereby improving the accuracy of insulator detection.

[0070] Step 4: Divide the insulator detection dataset into multiple datasets according to the background category, and label the corresponding background category in each dataset.

[0071] Among them, the background category can be divided according to the brightness level to simulate multiple scenarios, such as night, rainy day, and daytime, etc.

[0072] Step 5: Use the initial insulator detection model to perform transfer learning training on each dataset respectively to obtain insulator object detection models under different background categories.

[0073] In this embodiment, by performing transfer learning training on each dataset, insulator object detection models for different background categories can be obtained to improve the object detection accuracy under the corresponding background categories.

[0074] Step 6: Calculate the accuracy of the insulator object detection models under different background categories for detecting the corresponding datasets. If the accuracy is lower than the preset accuracy threshold, update the network parameters of the insulator object detection models under the corresponding background categories, and execute Step 5 until the accuracy of the insulator object detection models all reaches the preset accuracy threshold, and output the best insulator object detection model.

[0075] In a specific embodiment, Step 1 specifically includes:

[0076] Obtain a number of distribution network inspection line images based on the drone;

[0077] Use the labelme software to set and load the distribution network inspection line images, use a rectangular box to frame and label the insulators in the distribution network inspection line images, generate annotation files, and form an insulator detection dataset.

[0078] In a specific embodiment, the process of preprocessing the insulator detection dataset using the histogram equalization algorithm is specifically as follows:

[0079] According to the number of pixels and the gray level depth of the power distribution inspection line image before histogram equalization, calculate the frequency density of the power distribution inspection line image after histogram equalization as,

[0080]

[0081] where H B (D) represents the frequency density of the power distribution inspection line image after histogram equalization, A0 represents the number of pixels of the power distribution inspection line image before histogram equalization, and L represents the gray level depth of the power distribution inspection line image before histogram equalization;

[0082] Let the gray value D of the power distribution inspection line image A before histogram equalization A be mapped to the gray value D in the power distribution inspection line image after histogram equalization through the gray value transformation function B , that is, D B = f(D A ), f(D A ) represents the gray value transformation function, then the number of pixels of gray value D A within [0, D A is equal to the number of pixels of gray value D B within [0, D B , that is

[0083]

[0084] Solving the above equation gives the gray value transformation function f(D A ) as

[0085]

[0086] Discretize the gray value transformation function f(D A ) to obtain the discretized gray value transformation function f′(D A ). Input each pixel in the power distribution inspection line image A before histogram equalization into the discretized gray value transformation function f′(D A ) to obtain the gray value D B of the power distribution inspection line image B after histogram equalization, so as to realize histogram equalization for each power distribution inspection line image in the insulator detection dataset. Among them, the gray value D B is

[0087]

[0088] It can be understood that the gray value transformation function f(D A ) is discretized, so as to perform rounding processing on the gray value transformation function to facilitate the processing of each pixel in the power distribution inspection line image A.

[0089] In a specific embodiment, the process of preprocessing the insulator detection data set using an image filtering algorithm is specifically as follows:

[0090] Filter the power distribution inspection line image in the insulator detection data set based on a median filter. Let the pixel value at the i-th row and j-th column in the power distribution inspection line image be p(i, j). Then, the pixel value at this position after being filtered by the median filter is

[0091] med({p(i + x, j + y)|x ∈ {-1, 0, 1}, y ∈ {-1, 0, 1}})

[0092] In the formula, med(·) is a function for taking the median of the set elements, and p(i + x, j + y) is the pixel value after being filtered by the median filter;

[0093] Use a Gaussian filter to filter and denoise the power distribution inspection line image after being filtered by the median filter. The pixel value after being filtered by the Gaussian filter is

[0094]

[0095] In the formula, σ is the given standard deviation of the pixel value.

[0096] In a specific embodiment, the process of preprocessing the insulator detection data set using image sharpening is specifically as follows:

[0097] Calculate the second-order partial derivative of the pixel value of the power distribution inspection line image in the insulator detection data set as

[0098]

[0099]

[0100] In the formula, p0(i, j) is the gray value of the pixel at the i-th row and j-th column of the image before the Laplace operator acts on the pixel of the power distribution inspection line image;

[0101] Calculate the gray value after the Laplace operator acts on the pixel of the power distribution inspection line image as

[0102]

[0103] Calculate the pixel value of the power distribution inspection line image after image sharpening as

[0104]

[0105] In the formula, k is the coefficient of the diffusion effect.

[0106] In a specific embodiment, the method further includes:

[0107] The input X of the ResNeSt convolutional network is divided into K groups of cardinality units along the dimension of the input channels. Each group of cardinality units is divided into R groups of sub-cardinalities, and the input channels are divided into G groups of sub-channels, where G = KR;

[0108] Based on the ResNeSt convolutional network, each group of sub-channels is successively passed through a 1×1 convolutional layer and a 3×3 convolutional layer for feature extraction to obtain a feature U z , U z ∈R H×W×C , z = 1, 2, …, G, where H, W, and C are the three dimensions of U z respectively;

[0109] Calculate the sum of the features extracted by each group of sub-cardinalities within the k-th group of cardinality units as

[0110] Use average pooling along the dimension of the input channels to obtain an average feature s k , where the c-th component of the feature s k is

[0111]

[0112] Calculate the weight of the c-th component of the i-th group of sub-cardinalities within the k-th group of cardinality units as

[0113]

[0114] where represents the weight of the c-th component of the i-th group of partitions constructed based on the feature s k of the k-th group of cardinality units;

[0115] Perform weighted summation on the features extracted by each group of sub-cardinalities within each cardinality unit to obtain the feature of the cardinality unit. Then, the c-th component of the feature of the k-th cardinality unit is

[0116]

[0117] Concatenate the features of each cardinality unit and add them to the input X to obtain the feature Y extracted by the ResNeSt convolutional network as

[0118] V = Concat{V 1 , V 2 , …, V K}

[0119]

[0120] In the formula, Concat{·} represents a concatenation operation, represents the transformation of the convolutional layer and pooling layer that makes the number of channels of the input X consistent with that of V.

[0121] In a specific embodiment, the method further includes:

[0122] Constructing a loss function of the insulator target detection model by using the annotation file of the distribution network inspection line image,

[0123] FL(P t ) = -(1 - P t ) γ log(P t )

[0124]

[0125] In the formula, a is the category label of the detection target, a = 1 indicates that the detected object is an insulator, a = 0 indicates a non-insulator, p is the probability that the detected object is an insulator, and γ is a given focusing parameter.

[0126] Among them, the Adam algorithm is used for optimization, and multi-scale features are input into a fully connected neural network. The output layer of this neural network has only one unit, and the value of this unit is the probability that the detected object corresponding to the image is an insulator.

[0127] An insulator detection method based on background classification and transfer learning provided by the present invention preprocesses the insulator detection data set by constructing an insulator detection data set. Using a single-stage object detector to perform transfer learning training on the preprocessed insulator detection data set to obtain an initial insulator detection model, dividing the insulator detection data set into multiple data sets according to background categories, annotating the corresponding background categories in each data set, and using the initial insulator detection model to perform transfer learning training on each data set respectively to obtain insulator target detection models under different background categories, realizing accurate detection of insulators in each background, calculating the accuracy of the insulator target detection models under different background categories for detecting the corresponding data sets, if the accuracy is lower than the preset accuracy threshold, updating the network parameters of the insulator target detection model under the corresponding background category, and performing transfer learning until the accuracy of insulator detection in each background reaches the set value, thereby improving the accuracy of insulator detection.

[0128] Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working processes of the above-described system, device, and unit can refer to the corresponding processes in the foregoing method embodiments and will not be repeated here.

[0129] In several embodiments provided by the present invention, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces. The indirect couplings or communication connections of devices or units can be in electrical, mechanical or other forms.

[0130] The units described as separate components may or may not be physically separated. The components shown as units may or may not be physical units, that is, they can be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0131] In addition, in each embodiment of the present invention, the functional units can be integrated in a processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit. The above-mentioned integrated units can be implemented in the form of hardware or in the form of software functional units.

[0132] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of each embodiment of the present invention.

Claims

1. An insulator detection method based on background classification and transfer learning, characterized in that, It includes the following steps: Step 1: Obtain several images of the distribution network inspection line. Use the labelme software to mark the positions of the insulators on each image of the distribution network inspection line with rectangular frames and generate annotation files to form an insulator detection dataset. The annotation files include target position coordinates and category information. Step 2: Preprocess the insulator detection dataset. The preprocessing methods include histogram equalization algorithm, image filtering algorithm, and image sharpening. Among them, the process of preprocessing the insulator detection dataset using the histogram equalization algorithm is specifically as follows: According to the number of pixels and gray-level depth of the distribution network inspection line image before histogram equalization, calculate the frequency density of the distribution network inspection line image after histogram equalization as Where H B (D) represents the frequency density of the distribution network inspection line image after histogram equalization, A0 represents the number of pixels of the distribution network inspection line image before histogram equalization, and L represents the gray-level depth of the distribution network inspection line image before histogram equalization; Let the gray value of the distribution network inspection line image A before histogram equalization be D A is mapped to the gray value D in the distribution network inspection line image after histogram equalization through the gray value transformation function B , that is, D B = f(D A ), where f(D A ) represents the gray value transformation function. Then the number of pixels of gray value D A within [0, D A is equal to the number of pixels of gray value D B within [0, D B , that is Solving the above equation gives the gray value transformation function f(D A ), which is Discretize the gray value transformation function f(D A ) to obtain the discretized gray value transformation function f'(D A ). Input each pixel in the distribution network inspection line image A before histogram equalization into the discretized gray value transformation function f'(D A ) to obtain the gray value D B of the distribution network inspection line image B after histogram equalization, thereby realizing histogram equalization for each distribution network inspection line image in the insulator detection dataset. Among them, the gray value D B is Step 3: Use a single-stage object detector to perform transfer learning training on the preprocessed insulator detection dataset to obtain an initial insulator detection model. The single-stage object detector uses a ResNeSt convolutional network as the backbone network and a Feature Pyramid BiFPN as the feature extraction network. The ResNeSt convolutional network is pre-trained in the ImageNet deep learning network. Step 4: Divide the insulator detection dataset into multiple datasets according to the background category and label the corresponding background category in each dataset. Step 5: Use the initial insulator detection model to perform transfer learning training on each dataset respectively to obtain insulator object detection models under different background categories. Step 6: Calculate the accuracy of the insulator object detection models under different background categories in detecting the corresponding datasets. If the accuracy is lower than the preset accuracy threshold, update the network parameters of the insulator object detection model under the corresponding background category and execute Step 5 until the accuracies of the insulator object detection models all reach the preset accuracy threshold, and output the best insulator object detection model. This method also includes: Divide the input X of the ResNeSt convolutional network along the dimension of the input channel into K groups of cardinality units, divide each group of cardinality units into R groups of sub-cardinalities, and divide the input channel into G groups of sub-channels, where G = KR. Based on the ResNeSt convolutional network, each group of sub-channels is successively passed through a 1×1 convolutional layer and a 3×3 convolutional layer for feature extraction to obtain the feature U z , U z ∈R H×W×C , z = 1, 2, …, G, where H, W, and C are the three dimensions of U z respectively; Calculate the sum of the features extracted from each sub-cardinality group within the k-th group of cardinality units as Calculate the weight of the c-th component of the i-th sub-radix in the k-th group of radix units is In the formula, represents the feature s of the k-th group of base units k to construct the weight of the c-th component of the i-th group of partitions; Perform weighted summation on the features extracted by each group of sub-cardinalities within each cardinality unit to obtain the features of the cardinality unit. Then the c-th component of the features of the k-th cardinality unit is The features of each base unit are concatenated and added to the input X, and the feature Y extracted by the ResNeSt convolutional network is V = Concat{V 1 , V 2 , …, V K In the formula, Concat{·} represents the concatenation operation, represents the transformation of the convolutional layer and pooling layer that makes the number of channels of the input X consistent with V.

2. The insulator detection method based on background classification and transfer learning according to claim 1, wherein Step 1 specifically includes: Obtain several images of the distribution network inspection line based on a drone. Use the labelme software to set and load the images of the distribution network inspection line, use a rectangular frame to select and mark the insulators in the images of the distribution network inspection line, generate annotation files, and form an insulator detection dataset.

3. The insulator detection method based on background classification and transfer learning according to claim 1, characterized in that, The process of preprocessing the insulator detection dataset using the image filtering algorithm is specifically as follows: Filter the images of the distribution network inspection line in the insulator detection dataset based on a median filter. Assume that the pixel value at the i-th row and j-th column of the image of the distribution network inspection line is p(i, j). Then the pixel value at this position after filtering by the median filter is med({p(i + x, j + y)|x ∈ {-1, 0, 1}, y ∈ {-1, 0, 1}}) Where med(·) is a function to take the median of the set elements, and p(i + x, j + y) is the pixel value after filtering by the median filter; Use a Gaussian filter to filter and denoise the image of the distribution network inspection line after filtering by the median filter. The pixel value after Gaussian filtering is Where σ is the given standard deviation of the pixel value.

4. The insulator detection method based on background classification and transfer learning according to claim 1, wherein The process of preprocessing the insulator detection data set by image sharpening is specifically as follows: Calculate the second-order partial derivative of the pixel value of the distribution network inspection line image in the insulator detection data set as Where p0(i, j) is the gray value of the pixel at the i-th row and j-th column of the image before the Laplacian operator acts on the pixel of the distribution network inspection line image; Calculate the gray value after the Laplacian operator acts on the pixel of the distribution network inspection line image as Calculate the pixel value of the distribution network inspection line image after image sharpening as Where k is the coefficient of the diffusion effect.

5. The insulator detection method based on background classification and transfer learning according to claim 1, characterized in that It also includes: Use the annotation file of the distribution network inspection line image to construct the loss function of the insulator target detection model as FL(p t ) = -(1 - p t ) γ log(p t ) Where a is the class label of the detection target, a = 1 indicates that the detected object is an insulator, a = 0 indicates a non-insulator, p is the probability that the detected object is an insulator, and γ is the given focusing parameter.

Images