CN104978580B - A kind of insulator recognition methods for unmanned plane inspection transmission line of electricity - Google Patents

Abstract

The invention discloses an efficient identification method for transmission line insulators used in unmanned aerial vehicle inspections, including image acquisition and processing: extracting sub-images for training from transmission line insulator images and performing preliminary processing to form a training data set; Pack the sub-images extracted for training, and add the corresponding labels of the images; use the convolutional neural network (CNN) algorithm in deep learning to train the data, and obtain a detection model for insulators; insulator target area detection: Detect the transmission line image to obtain the candidate frame of the insulator target; perform non-maximum suppression on the candidate frame to obtain the final insulator candidate frame; perform a straight line fitting operation on the obtained final insulator candidate frame to obtain the center point, candidate The angle and size information of the box are finally marked on the image of the transmission line insulator. The application screens the images obtained by the patrol inspection to reduce the burden of manual screening, and has broad application prospects.

Description

An insulator identification method for UAV inspection of transmission lines

technical field

The invention relates to digital image processing and pattern recognition technology in the technical field of transmission line equipment detection, in particular to an efficient identification method for unmanned aerial vehicle inspection of transmission line insulators.

Background technique

Insulators are an important part of overhead transmission lines, used to support and fix busbars and live conductors, and to provide sufficient distance and insulation between live conductors or between conductors and the earth. Due to the long-term exposure of overhead transmission lines to the natural environment, affected by natural or human factors, there are problems such as line aging and damage. If these problems are not regularly inspected and repaired, major safety accidents may occur.

Manual line inspection is inefficient and dangerous. With the development of UAV technology, the use of UAV aerial photography technology to collect high-voltage line images and process these image information can reduce personnel costs and ensure the safety of construction personnel, while improving work efficiency.

Due to the complex geographical environment of the location of the transmission line towers, the acquired image background is also relatively complex, which makes it difficult to identify and locate subsequent targets. Moreover, the state detection technology of transmission lines based on aerial images is still in its infancy. References and Research results are few.

In the prior art, one type of method utilizes color information to segment a color image by using a maximum threshold method and a maximum inter-class variance method. Disadvantages: This method is seriously affected by light, and the natural environment of the transmission line is complex, with complex backgrounds such as trees, rivers, roads, etc., which makes the detection accuracy of this method not high.

Another method is to use the ellipse information of the sheet structure of the insulator and use the Hough transform to detect the ellipse. Disadvantages: Due to the problem of the shooting angle, there is occlusion between the slices, resulting in detection errors.

Contents of the invention

In order to solve the deficiencies in the existing technology, the present invention discloses an efficient identification method for transmission line insulators used in drone inspections. The straight line fitting method calculates the angle information of the insulator, which can meet the technical requirements for the identification and positioning of the insulator in the aerial line inspection image of the UAV, and improve the accuracy and robustness of the insulator detection.

To achieve the above object, the specific scheme of the present invention is as follows:

An insulator identification method for unmanned aerial vehicle inspection of transmission lines, comprising the following steps:

Step 1: Image acquisition and processing: Extract sub-images for training from the transmission line insulator images and perform preliminary processing to form a training data set;

Step 2: Pack the sub-images extracted for training, and add labels corresponding to the images;

Step 3: Use the convolutional neural network (CNN) algorithm in deep learning to train the data in the step 1 data set to obtain a detection model for insulators;

Step 4: Insulator target area detection: detect the transmission line image, and obtain the candidate frame of the insulator target;

Step 5: Perform non-maximum suppression on the candidate frame to obtain the final insulator candidate frame;

Step 6: Perform a straight line fitting operation on the obtained final insulator candidate frame, obtain the center point, angle and size information of the candidate frame, and finally mark on the transmission line insulator image.

In the first step, the image acquisition and processing process is: extracting the regional images of insulator components, poles and backgrounds from the transmission line insulator image, performing scaling processing on the extracted image, and performing angle rotation operation on the extracted insulator.

In the third step, when using the convolutional neural network algorithm for model training:

First, set the template parameters used for the initial training, including the number of layers of the convolutional neural network, the size of the convolution kernel, the initial weight of each node, whether to use downsampling, the number of data input and output for each layer, the activation function, and each Learning Efficiency of Gradient Descent for Multilayer Convolutional Neural Networks.

In said step three, the specific insulator detection model training process based on convolutional neural network includes the following steps:

3-1) Feedforward Pass: The input color image is divided into RGB three channels to extract pixel information, which is used as the input information of the convolutional neural network;

3-2) BackPropagation Pass: optimize the loss function between the category label obtained by the forward pass and the actual label of the sample.

In the step 3-1), the structure of the convolutional neural network uses a six-layer convolutional neural network structure training template;

The first layer is a convolutional layer, the third layer is a convolutional layer, and the fifth layer is a convolutional layer. The convolutional layer uses a convolution kernel with a set pixel size and input information to perform a convolution operation to obtain a feature vector;

The second layer is the downsampling layer, and the fourth layer is the downsampling layer. The downsampling layer performs the downsampling operation according to the definition in the parameter template, and uses the maximum value in the image block of the set pixel size as the output. Through the down-sampling process, the amount of data processing can be effectively reduced on the basis of preserving useful information;

The sixth layer is the fully connected layer: the feature vector obtained in the fifth layer is integrated to form a long vector, which is passed to the activation function to obtain the judgment of the input sample category, and the largest output value is selected as the label of the input sample.

In the step 3-2), it is assumed that the form of the training set is {(x ⁽¹⁾ ,y ⁽¹⁾ ),(x ⁽²⁾ ,y ⁽²⁾ ),...,(x ⁽ⁿ⁾ ,y ⁽ⁿ⁾ )}, x ⁽ⁱ⁾ represents the i-th training data, y ⁽ⁱ⁾ represents the data label corresponding to the data x ⁽ⁱ⁾ , and the training data set contains n samples.

For a single sample (x,y), the learned result is h _w,b (x), and its loss function is:

So the overall loss function J(W,b) is:

In formula (2), s _l , s _l+1 represent the number of neuron nodes in layer l and layer l+1, n _l represents the number of layers of the neural network, Represents the weight coefficient connecting the i-th node of the l-th layer and the j-th node of the l+1-th layer in the l-th group of weight parameters, is the bias of the i-th node in layer l, the first item J(W,b; x ⁽ⁱ⁾ , y ⁽ⁱ⁾ ) is a mean square error item, and the second item is a regularization item, which is used to reduce the weight The magnitude of , to prevent over-fitting, γ is the control coefficient.

The parameters W and b are updated using the gradient descent method:

Among them, α in formula (3) is the learning rate, which is used to control the speed of gradient descent.

In the fourth step, the convolutional neural network is used to detect the insulator, and the detailed method is as follows:

4-1) Extract the trained convolutional neural network template, including weights, biases and network structures used for training; then initialize the test program according to these parameters;

4-2) Load the image. Since the image collected by the UAV is large, the image is scaled to speed up the subsequent calculation. In order to accurately locate the position of the insulator, a multi-scale method is added to slide the frame on multiple scales. Operation, to obtain a specific target image block;

4-3) The target image block is used as input, and a forward conduction operation is performed to obtain the category to which the target block belongs;

4-4) Save the target block information of the insulation subcategory, including the starting position and length and width information of the target block.

In the sixth step, use the straight line fitting method to mark the insulator, and the specific process is as follows:

6-1) Obtain the m _i , i={1,...,t} candidate frame information of the t categories obtained in step 5, calculate the center point position of each candidate frame, and save it;

6-2) The position of the center point is (P, Q). The center position of each category can be fitted by linear fitting, and the position of each insulator can be accurately positioned. Here, one-element linear fitting can be used to solve the problem. this problem:

Y=kX+b' (4)

6-3) Calculate a straight line that can best reflect the relationship between X and Y by using the least squares fitting method.

In the step 6-3), the loss function is defined as:

Among them, (p _j , q _j ) is the center point of the jth rectangular frame belonging to class i.

The optimal solution of b' and k is obtained by taking partial derivatives of b' and k:

The final solution to get k and b' is

Among them, (p _j , q _j ) is the center point of the jth rectangular frame belonging to the i class, and m _i represents the number of the i class rectangular frame.

Beneficial effects of the present invention:

This application uses convolution layer, downsampling layer and full connection layer: after three layers of convolution and two layers of downsampling operations, all feature blocks are fully connected, so that the final feature description of the image block is obtained .

The effective identification technology of transmission line insulators can effectively realize the positioning of insulator targets in inspection images, and provide a basis for subsequent defect diagnosis. At the same time, this technology can also screen the images obtained by inspections, reducing the burden of manual screening, and has broad application prospects.

Description of drawings

Figure 1 CNN structure diagram;

Figure 2 Convolution and downsampling process;

Figure 3 UAV aerial photography transmission line image;

Figure 4 The transmission line image after the candidate frame is marked;

Figure 5 The image of the final insulator inspection results;

Figure 6 training flow chart;

Fig. 7 Target detection and localization flow chart.

Detailed ways:

The present invention is described in detail below in conjunction with accompanying drawing:

The training flow chart of the experimental process is shown in Figure 6. The data set is extracted, the data set is packaged, and the training process is started. Before training, the training hierarchy and initialization parameters are performed. During training, forward training, reverse feedback, and judgment of the maximum number of iterations /Accuracy requirement, if yes, output template, otherwise, continue training.

The flow chart of target recognition and positioning is shown in Figure 7. Input the image, initialize the detection model, and extract the subgraph through the sliding window to determine whether it is an insulator. If not, return to the process of extracting the subgraph through the sliding window. If yes, store the candidate frame and fit The candidate box is output at the position marked in the original image.

An efficient method for identifying insulators of transmission lines by UAV inspections, the specific steps include:

1) Training set: Extract sub-images for training from the transmission line image to form a training data set.

2) pack the extracted insulator image, and add the label of the corresponding image;

3) Use the convolutional neural network (CNN) algorithm in deep learning to train the data to obtain a detection model for insulators;

4) Insulator target area detection: detect the transmission line image, and obtain the candidate frame of the insulator target;

5) Perform non-maximum suppression on the candidate frame to obtain the final insulator candidate frame.

6) Perform a straight line fitting operation on the obtained candidate frame, obtain the center point, the angle and size information of the frame, and finally mark it on the original image.

The image acquisition process of the step 1) is: extract the regional images of the insulator parts, towers and background in the original image, in order to meet the requirements of the training, the extracted images are scaled, and the size is scaled to 64*64 pixels. In actual situations, the image of the tower may be tilted due to the shooting angle. Therefore, the extracted insulators are rotated at a small angle to increase the diversity of the data set and improve the robustness of the training model.

Described step 3) in step, the concrete method that carries out model training with CNN algorithm is as follows:

(1) First, set the template parameters used for the initial training, including the number of layers of the convolutional neural network, the size of the convolution kernel, the initial weight of each node, whether to use downsampling processing, and the number of data inputs and outputs for each layer , activation function, etc., as well as the setting of the learning efficiency of the gradient descent of the corresponding template.

(2) Network structure, here we use a six-layer convolutional neural network structure training template. The form is shown in the CNN structure diagram in Figure 1.

The specific CNN-based insulator detection model training process mainly has the following two steps:

A. Forward conduction: Divide the input color image into RGB three channels to extract pixel information as the input information of the convolutional network. The first layer is the convolutional layer (Convolution), which uses a 5*5 pixel-sized convolution kernel and input The information is convoluted. Through the convolution operation, the original signal can be enhanced and the influence of noise can be reduced. The second layer is the downsampling layer. According to the definition in the parameter template, the downsampling operation (SubSampling/Pooling) is performed. Here, the maximum value in the 2*2 image block is used as the output. The downsampling process can effectively reduce the amount of data processing on the basis of preserving useful information; the third convolutional layer, the fourth downsampling layer, and the fifth convolutional layer perform similar operations; the sixth layer is fully connected Layer: Integrate the feature vectors obtained in the fifth layer to form a long vector, which is passed to the activation function to obtain the judgment of the input sample category. as shown in picture 2.

The convolutional layer, downsampling layer and full connection layer are further described below:

Input a color image, use a trainable filter fx to perform a convolution operation with the image, and then add a paranoid bx, which extracts an image feature and obtains the Cx of the convolutional layer; in order to reduce the amount of data, the four-neighborhood local The maximum value of the pixel is changed as the output, and then the nonlinear transformation of the data is completed through the ReLu activation function. The nonlinear transformation can reduce the linear relationship between features and enhance the expressive ability of features. Moreover, after the processing of such a downsampling layer, the amount of data obtained is only a quarter of that of the convolutional layer.

After three layers of convolution and two layers of downsampling operations, all feature blocks are fully connected, so that the final feature description of the image block is obtained. In order to predict the category of the image, the obtained feature descriptor is subjected to activation function operation processing, and the largest output value is selected as the label of the image block.

B. Backward conduction: Calculate the loss function through the category label obtained by Feedforward Pass and the actual label of the sample. Assume that the form we use for the training set is {(x ⁽¹⁾ ,y ⁽¹⁾ ),(x ⁽²⁾ ,y ⁽²⁾ ),...,(x ⁽ⁿ⁾ ,y ⁽ⁿ⁾ )}, The training data set contains n samples. For a single sample (x, y), the learned result is h _{w, b} (x), and its loss function is:

So the overall loss function is:

Formula (2) s _l represents the number of neurons in layer l, Indicates the weight coefficient connecting the i-th node of layer l and the j-th node of layer l+1 in the l-th group of weight parameters. The first item J(W,b; x ⁽ⁱ⁾ , y ⁽ⁱ⁾ ) is a mean square error item, the second item is a regularization item, which is used to reduce the magnitude of the weight and prevent overfitting, and γ is the control coefficient . The parameters W and b are updated with each iteration update of the gradient descent method:

α in formula (3) is the learning rate, which is used to control the speed of gradient descent.

In the following pseudocode, ΔW ^(l) is a matrix with the same dimensions as matrix W ^(l) , and Δb ^(l) is a vector with the same dimensions as b ^(l) . Below, an iterative process in implementing the gradient descent method is given:

1. For all layers l, let ΔW ^(l) :=0, Δb ^(l) :=0;

2. For i=1 to n:

a. Calculated using the backpropagation algorithm

b. Calculate

c. Calculate

3. Update weight parameters:

In this way, the gradient descent method can be used repeatedly to iteratively calculate the value to reduce the loss function J(W,b), and then solve the entire neural network.

In described step 4), use CNN to carry out the detection of insulator, detailed method is as follows:

(1) Extract the trained CNN template, including weights, biases, and the network structure used for training; then initialize the test program framework according to these parameters;

(2) Load the image (as shown in Figure 3). Since the image collected by the drone is large (5184*3456), the image is scaled to speed up subsequent calculations. In order to accurately locate the position of the insulator, a multi-scale method is added to perform sliding frame operations on multiple scales to obtain specific target image blocks.

(3) The target image block is used as input, and the CNN Feedforward Pass operation is performed to obtain the category of the target block. (4) Save the target block information of the insulator category, including the starting position and length and width information of the target block, and Fig. 4 is the image of the insulator through marking;

In the step 6), the fitting method is used to mark the insulator, and the specific process is as follows:

(1) Obtain the m _i , i={1,...,t} candidate frame information of t categories obtained in step (5), and calculate the center point of each frame

set, and save.

(2) The position of the center point is (P, Q), and the center position of each category can be fitted by a linear fitting method to accurately locate the position of each insulator. Here a unary linear fit can be used to solve this problem:

Y=kX+b' (4)

(3) Calculate a straight line that can best reflect the relationship between X and Y by using the least squares fitting method:

Define the loss function as:

The optimal solution of b' and k is obtained by taking partial derivatives of b' and k:

The final solution to get k and b' is

In this way, the angle information of the insulator is determined, and the specific position of the insulator is marked on the original map according to the angle information. Figure 5 is the final insulator marking image.

Although the specific implementation of the present invention has been described above in conjunction with the accompanying drawings, it does not limit the protection scope of the present invention. Those skilled in the art should understand that on the basis of the technical solution of the present invention, those skilled in the art do not need to pay creative work Various modifications or variations that can be made are still within the protection scope of the present invention.

Claims (10)

1. An insulator identification method for unmanned aerial vehicle inspection transmission line, it is characterized in that, comprises the following steps: Step 1: Image acquisition and processing: Extract sub-images for training from the transmission line insulator images and perform preliminary processing to form a training data set; Step 2: Pack the sub-images extracted for training, and add the labels of the corresponding images; Step 3: Use the convolutional neural network algorithm in deep learning to train the data in the step 1 dataset to obtain a detection model for insulators; Step 4: Insulator target area detection: detect the transmission line image, and obtain the candidate frame of the insulator target; Step 5: Perform non-maximum suppression on the candidate frame to obtain the final insulator candidate frame; Step 6: Perform a straight line fitting operation on the obtained final insulator candidate frame, obtain the center point, angle and size information of the candidate frame, and finally mark on the transmission line insulator image. 2. A kind of insulator identification method for unmanned aerial vehicle inspection transmission line as claimed in claim 1, it is characterized in that, in described step 1, image acquisition and processing process are: extract insulator in transmission line insulator image For the regional images of components, towers and backgrounds, the extracted images are scaled, and the angles of the extracted insulators are rotated. 3. A kind of insulator identification method for unmanned aerial vehicle inspection transmission line as claimed in claim 1, it is characterized in that, in described step 3, when carrying out model training with convolutional neural network algorithm: First set the template parameters used for initial training, including the number of layers of the convolutional neural network, the size of the convolution kernel, the initial weight of each node, whether to use downsampling processing, the number of data inputs and outputs for each layer, and the activation function. Learning Efficiency of Gradient Descent in Per-Layer Convolutional Neural Networks. 4. A kind of insulator identification method for unmanned aerial vehicle inspection transmission line as claimed in claim 1 or 3, it is characterized in that, in described step 3, concrete insulator detection model training process based on convolutional neural network Include the following steps: 3-1) Forward conduction: the input color image is divided into RGB three channels to extract pixel information, which is used as the input information of the convolutional neural network; 3-2) Backward conduction: Calculate the loss function through the category label obtained by the forward conduction and the actual label of the sample. 5. A kind of insulator identification method for unmanned aerial vehicle inspection transmission line as claimed in claim 4, it is characterized in that, in described step 3-1), the structure of convolutional neural network, uses the volume of six layers Convolutional neural network structure training template, the first layer is a convolutional layer, the third layer is a convolutional layer, and the fifth layer is a convolutional layer. The convolutional layer uses a convolution kernel with a set pixel size to perform convolution operations with input information. Get the feature vector; The second layer is the downsampling layer, and the fourth layer is the downsampling layer. The downsampling layer performs the downsampling operation according to the definition in the parameter template, and uses the maximum value in the image block of the set pixel size as the output. Through the down-sampling process, the amount of data processing can be effectively reduced on the basis of preserving useful information; The sixth layer is the fully connected layer: the feature vector obtained in the fifth layer is integrated to form a long vector, which is passed to the activation function to obtain the judgment of the input sample category, and the largest output value is selected as the label of the image block. 6. a kind of insulator identification method that is used for unmanned aerial vehicle inspection transmission line as claimed in claim 4, is characterized in that, in described step 3-2), it is assumed that the training set form is {(x ^{(1 )} ,y ⁽¹⁾ ),(x ⁽²⁾ ,y ⁽²⁾ ),...,(x ⁽ⁿ⁾ ,y ⁽ⁿ⁾ )}, x ⁽ⁱ⁾ represents the i-th training data, y ^{( i)} represents the data label corresponding to the data x ⁽ⁱ⁾ , and the training data set contains n samples; For a single sample (x,y), the learned result is h _w,b (x), and its loss function is: <mrow><mi>J</mi><mrow><mo>(</mo><mi>W</mi><mo>,</mo><mi>b</mi><mo>;</mo><mi>x</mi><mo>,</mo><mi>y</mi><mo>)</mo></mrow><mo>=</mo><mfrac><mn>1</mn><mn>2</mn></mfrac><mo>|</mo><mo>|</mo><msub><mi>h</mi><mrow><mi>w</mi><mo>,</mo><mi>b</mi></mrow></msub><mrow><mo>(</mo><mi>x</mi><mo>)</mo></mrow><mo>-</mo><mi>y</mi><mo>|</mo><msup><mo>|</mo><mn>2</mn></msup><mo>-</mo><mo>-</mo><mo>-</mo><mrow><mo>(</mo><mn>1</mn><mo>)</mo></mrow></mrow> So the overall loss function is: <mrow><mtable><mtr><mtd><mrow><mi>J</mi><mrow><mo>(</mo><mi>W</mi><mo>,</mo><mi>b</mi><mo>)</mo></mrow><mo>=</mo><mo>&amp;lsqb;</mo><mfrac><mn>1</mn><mi>n</mi></mfrac><munderover><mi>&amp;Sigma;</mi><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><mi>n</mi></munderover><mi>J</mi><mrow><mo>(</mo><mi>W</mi><mo>,</mo><mi>b</mi><mo>;</mo><msup><mi>x</mi><mrow><mo>(</mo><mi>i</mi><mo>)</mo></mrow></msup><mo>,</mo><msup><mi>y</mi><mrow><mo>(</mo><mi>i</mi><mo>)</mo></mrow></msup><mo>)</mo></mrow><mo>&amp;rsqb;</mo><mo>+</mo><mfrac><mi>&amp;gamma;</mi><mn>2</mn></mfrac><munderover><mi>&amp;Sigma;</mi><mrow><mi>l</mi>mi><mo>=</mo><mn>1</mn></mrow><mrow><msub><mi>n</mi><mi>l</mi></msub><mo>-</mo><mn>1</mn></mrow></munderover><munderover><mi>&amp;Sigma;</mi><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><msub><mi>s</mi><mi>l</mi></msub></munderover><munderover><mi>&amp;Sigma;</mi><mrow><mi>j</mi><mo>=</mo><mn>1</mn></mrow><msub><mi>s</mi><mrow><mi>l</mi><mo>+</mo><mn>1</mn></mrow></msub></munderover><msup><mrow><mo>(</mo><msubsup><mi>W</mi><mrow><mi>j</mi><mi>i</mi></mrow><mrow><mo>(</mo><mi>l</mi><mo>)</mo></mrow></msubsup><mo>)</mo></mrow><mn>2</mn></msup></mrow></mtd></mtr><mtr><mtd><mrow><mo>=</mo><mo>&amp;lsqb;</mo><mfrac><mn>1</mn><mi>n</mi></mfrac><munderover><mi>&amp;Sigma;</mi><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><mi>n</mi></munderover><mrow><mo>(</mo><mfrac><mn>1</mn><mn>2</mn></mfrac><mo>|</mo><mo>|</mo><msub><mi>h</mi><mrow><mi>w</mi><mo>,</mo><mi>b</mi></mrow></msub><mo>(</mo><msup><mi>x</mi><mrow><mo>(</mo><mi>i</mi><mo>)</mo></mrow></msup><mo>)</mo><mo>-</mo><msup><mi>y</mi><mrow><mo>(</mo><mi>i</mi><mo>)</mo></mrow></msup><mo>|</mo><msup><mo>|</mo><mn>2</mn></msup><mo>)</mo></mrow><mo>&amp;rsqb;</mo><mo>+</mo><mo>+</mo><mfrac><mi>&amp;gamma;</mi><mn>2</mn></mfrac><munderover><mi>&amp;Sigma;</mi><mrow><mi>l</mi><mo>=</mo><mn>1</mn></mrow><mrow><msub><mi>n</mi><mi>l</mi></msub><mo>-</mo><mn>1</mn></mrow></munderover><munderover><mi>&amp;Sigma;</mi><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><msub><mi>s</mi><mi>l</mi></msub></munderover><munderover><mi>&amp;Sigma;</mi><mrow><mi>j</mi><mo>=</mo><mn>1</mn></mrow><msub><mi>s</mi><mrow><mi>l</mi><mo>+</mo><mn>1</mn></mrow></msub></munderover><msup><mrow><mo>(</mo><msubsup><mi>W</mi><mrow><mi>j</mi><mi>i</mi></mrow><mrow><mo>(</mo><mi>l</mi><mo>)</mo></mrow></msubsup><mo>)</mo></mrow><mn>2</mn></msup></mrow></mtd></mtr></mtable><mo>-</mo><mo>-</mo><mo>-</mo><mrow><mo>(</mo><mn>2</mn><mo>)</mo></mrow></mrow> Formula (2) s _l represents the number of neurons in layer l, Represents the weight coefficient connecting the i-th node of the l-th layer and the j-th node of the l+1-th layer in the l-th group of weight parameters, is the bias of the i-th node in layer l, the first item J(W,b; x ⁽ⁱ⁾ , y ⁽ⁱ⁾ ) is a mean square error item, and the second item is a regularization item, which is used to reduce the weight The magnitude of , to prevent over-fitting, γ is the control coefficient. 7. A kind of insulator identification method for unmanned aerial vehicle inspection transmission line as claimed in claim 6, it is characterized in that, update parameter W and b with gradient descent method each iterative update: <mrow><mtable><mtr><mtd><mrow><msubsup><mi>W</mi><mrow><mi>j</mi><mi>i</mi></mrow><mrow><mo>(</mo><mi>l</mi><mo>)</mo></mrow></msubsup><mo>=</mo><msubsup><mi>W</mi><mrow><mi>j</mi><mi>i</mi></mrow><mrow><mo>(</mo><mi>l</mi><mo>)</mo></mrow></msubsup><mo>-</mo><mi>&amp;alpha;</mi><mfrac><mo>&amp;part;</mo><mrow><mo>&amp;part;</mo><msubsup><mi>W</mi><mrow><mi>j</mi><mi>i</mi></mrow><mrow><mo>(</mo><mi>l</mi><mo>)</mo></mrow></msubsup></mrow></mfrac><mi>J</mi><mrow><mo>(</mo><mi>W</mi><mo>,</mo><mi>b</mi><mo>)</mo></mrow></mrow></mtd></mo>mtr><mtr><mtd><mrow><msubsup><mi>b</mi><mi>i</mi><mrow><mo>(</mo><mi>l</mi><mo>)</mo></mrow></msubsup><mo>=</mo><msubsup><mi>b</mi><mi>i</mi><mrow><mo>(</mo><mi>l</mi><mo>)</mo></mrow></msubsup><mo>-</mo><mi>&amp;alpha;</mi><mfrac><mo>&amp;part;</mo><mrow><mo>&amp;part;</mo><msubsup><mi>b</mi><mi>i</mi><mrow><mo>(</mo><mi>l</mi><mo>)</mo></mrow></msubsup></mrow></mfrac><mi>J</mi><mrow><mo>(</mo><mi>W</mi><mo>,</mo><mi>b</mi><mo>)</mo></mrow></mrow></mtd></mtr></mtable><mo>-</mo><mo>-</mo><mo>-</mo><mrow><mo>(</mo><mn>3</mn><mo>)</mo></mrow></mrow> Among them, α in formula (3) is the learning rate, which is used to control the speed of gradient descent. 8. A kind of insulator identification method for unmanned aerial vehicle inspection transmission line as claimed in claim 1, it is characterized in that, in described step 4, use convolutional neural network to carry out the detection of insulator, detailed method is as follows: 4-1) Extract the trained convolutional neural network template, including weights, biases, and the network structure used for training; then initialize the test program according to these parameters; 4-2) Load the image. Since the image collected by the drone is large in scale, the image is scaled to speed up subsequent calculations. In order to accurately locate the position of the insulator, a multi-scale method is added to slide on multiple scales Frame operation to obtain specific target image blocks; 4-3) The target image block is used as input, and a forward conduction operation is performed to obtain the category to which the target block belongs; 4-4) Save the target block information of the insulation subcategory, including the starting position and length and width information of the target block. 9. A kind of insulator identification method for UAV inspection transmission line as claimed in claim 1, it is characterized in that, in described step 6, use straight line fitting method to mark insulator, specific process is as follows: 6-1) Obtain the m _i , i={1,...,t} candidate frame information of t categories obtained in step 5, calculate the position of the center point of each candidate frame, and save it, m _i represents the The number of rectangular boxes of class i; 6-2) The position of the center point is (P, Q), and the center position of each category is fitted by linear fitting, and the position of each insulator is accurately positioned, and this problem is solved by unary linear fitting: Y=kX+b' (4) 6-3) Calculate a straight line that can best reflect the relationship between X and Y by using the least squares fitting method. 10. A kind of insulator identification method for unmanned aerial vehicle inspection transmission line as claimed in claim 9, it is characterized in that, in described step 6-3), define loss function as: <mrow><mi>L</mi><mo>=</mo><munderover><mo>&amp;Sigma;</mo><mrow><mi>j</mi><mo>=</mo><mn>1</mn></mrow><msub><mi>m</mi><mi>i</mi></msub></munderover><msup><mrow><mo>(</mo><msub><mi>q</mi><mi>j</mi></msub><mo>-</mo><mo>(</mo><mrow><msup><mi>b</mi><mo>&amp;prime;</mo></msup><mo>+</mo><msub><mi>kp</mi><mi>j</mi></msub></mrow><mo>)</mo><mo>)</mo></mrow><mn>2</mn></msup><mo>-</mo><mo>-</mo><mo>-</mo><mrow><mo>(</mo><mn>5</mn><mo>)</mo></mrow></mrow> The optimal solution of b' and k is obtained by taking partial derivatives of b' and k: <mrow><mtable><mtr><mtd><mrow><mfrac><mrow><mo>&amp;part;</mo><mi>L</mi></mrow><mrow><mo>&amp;part;</mo><msup><mi>b</mi><mo>&amp;prime;</mo></msup></mrow></mfrac><mo>=</mo><mn>2</mn><munderover><mo>&amp;Sigma;</mo><mrow><mi>j</mi><mo>=</mo><mn>1</mn></mrow><msub><mi>m</mi><mi>i</mi></msub></munderover><mrow><mo>(</mo><msub><mi>q</mi><mi>j</mi></msub><mo>-</mo><mo>(</mo><mrow><msup><mi>b</mi><mo>&amp;prime;</mo></msup><mo>+</mo><msub><mi>kp</mi><mi>j</mi></msub></mrow><mo>)</mo><mo>)</mo></mrow></mrow></mtd></mtr><mtr><mtd><mrow><mfrac><mrow><mo>&amp;part;</mo><mi>L</mi></mrow><mrow><mo>&amp;part;</mo><mi>k</mi></mrow></mfrac><mo>=</mo><mo>-</mo><mn>2</mn><munderover><mo>&amp;Sigma;</mo><mrow><mi>j</mi><mo>=</mo><mn>1</mn></mrow><msub><mi>m</mi><mi>i</mi></msub></munderover><mrow><mo>(</mo><msub><mi>q</mi><mi>j</mi></msub><mo>-</mo><mo>(</mo><mrow><msup><mi>b</mi><mo>&amp;prime;</mo></msup><mo>+</mo><msub><mi>kp</mi><mi>j</mi></msub></mrow><mo>)</mo><mo>)</mo></mrow><msub><mi>p</mi><mi>j</mi></msub></mrow></mtd></mtr></mtable><mo>-</mo><mo>-</mo><mo>-</mo><mrow><mo>(</mo><mn>5</mn><mo>)</mo></mrow></mrow> The final solution to get k and b' is <mrow><mtable><mtr><mtd><mrow><mi>k</mi><mo>=</mo><mfrac><mrow><msub><mi>m</mi><mi>i</mi></msub><mo>*</mo><munderover><mi>&amp;Sigma;</mi><mrow><mi>j</mi><mo>=</mo><mn>1</mn></mrow><msub><mi>m</mi><mi>i</mi></msub></munderover><msub><mi>p</mi><mi>j</mi></msub><msub><mi>q</mi><mi>j</mi></msub><mo>-</mo><munderover><mi>&amp;Sigma;</mi><mrow><mi>j</mi><mo>=</mo><mn>1</mn></mrow><msub><mi>m</mi><mi>i</mi></msub></munderover><msub><mi>p</mi><mi>j</mi></msub><munderover><mi>&amp;Sigma;</mi><mrow><mi>j</mi><mo>=</mo><mn>1</mn></mrow><msub><mi>m</mi><mi>i</mi></msub></munderover><msub><mi>q</mi><mi>j</mi></msub></mrow><mrow><msub><mi>m</mi><mi>i</mi></msub><mo>*</mo><munderover><mi>&amp;Sigma;</mi><mrow><mi>j</mi><mo>=</mo><mn>1</mn></mrow><msub><mi>m</mi><mi>i</mi></msub></munderover><msubsup><mi>p</mi><mi>j</mi><mn>2</mn></msubsup><mo>-</mo><munderover><mi>&amp;Sigma;</mi><mrow><mi>j</mi><mo>=</mo><mn>1</mn></mrow><msub><mi>m</mi><mi>i</mi></msub></munderover><msubsup><mi>q</mi><mi>j</mi><mn>2</mn></msubsup></mrow></mfrac></mrow></mtd></mtr><mtr><mtd><mrow><msup><mi>b</mi><mo>&amp;prime;</mo></msup><mo>=</mo><mfrac><mrow><msub><mi>m</mi><mi>i</mi></msub><mo>*</mo><munderover><mi>&amp;Sigma;</mi><mrow><mi>j</mi><mo>=</mo><mn>1</mn></mrow><msub><mi>m</mi><mi>i</mi></msub></munderover><msubsup><mi>p</mi><mi>j</mi><mn>2</mn></msubsup><msub><mi>q</mi><mi>j</mi></msub><mo>-</mo><munderover><mi>&amp;Sigma;</mi><mrow><mi>j</mi><mo>=</mo><mn>1</mn></mrow><msub><mi>m</mi><mi>i</mi></msub></munderover><msub><mi>p</mi><mi>j</mi></msub><munderover><mi>&amp;Sigma;</mi><mrow><mi>j</mi><mo>=</mo><mn>1</mn></mrow><msub><mi>m</mi><mi>i</mi></msub></munderover><msub><mi>q</mi><mi>j</mi></msub></mrow><mrow><msub><mi>m</mi><mi>i</mi></msub><mo>*</mo><munderover><mi>&amp;Sigma;</mi><mrow><mi>j</mi><mo>=</mo><mn>1</mn></mrow><msub><mi>m</mi><mi>i</mi></msub></munderover><msubsup><mi>p</mi><mi>j</mi><mn>2</mn></msubsup><mo>-</mo><munderover><mi>&amp;Sigma;</mi><mrow><mi>j</mi><mo>=</mo><mn>1</mn></mrow><msub><mi>m</mi><mi>i</mi></msub></munderover><msubsup><mi>q</mi><mi>j</mi><mn>2</mn></msubsup></mrow></mfrac></mrow></mtd></mtr></mtable><mo>-</mo><mo>-</mo><mo>-</mo><mrow><mo>(</mo><mn>7</mn><mo>)</mo></mrow><mo>;</mo></mrow> Among them, (p _j , q _j ) is the center point of the jth rectangular frame belonging to the i class, and m _i represents the number of the i class rectangular frame.