CN104978580A - Insulator identification method for unmanned aerial vehicle polling electric transmission line - Google Patents

Abstract

The invention discloses an insulator identification method for an unmanned aerial vehicle polling an electric transmission line. The method comprises the steps: image collecting and processing: extracting sub-images for training from insulator images of the electric transmission line and initially processing the sub-images to form a training data set; packaging the extracted sub-images for training and adding labels corresponding to the sub-images; training the data by virtue of a convolutional neural network (CNN) algorithm in deep learning to obtain a detection model for insulators; insulator target region detection: detecting an electric transmission line image to obtain a candidate frame of the insulator target; carrying out non-maxima suppression on the candidate frame to obtain a final insulator candidate frame; and carrying out straight line fitting operation on the obtained final insulator candidate frame to obtain a central point and the angle and size information of the candidate frame and finally labeling on the insulator images of the electric transmission line. According to the method disclosed by the invention, images obtained by polling are screened, so that the burden of manual screening is alleviated and the method has a broad application prospect.

Description

A kind of insulator recognition methods of patrolling and examining transmission line of electricity for unmanned plane

Technical field

The present invention relates to Digital Image Processing and mode identification technology in transmission line equipment detection technique field, particularly relate to a kind of recognition methods of patrolling and examining electric transmission line isolator efficiently for unmanned plane.

Background technology

Insulator is ingredient important in overhead transmission line, is used for supporting and fixing bus and energized conductor make between energized conductor or have enough Distance geometry to insulate between conductor and the earth.Because overhead transmission line is chronically exposed in physical environment, be subject to the impact of nature or human factor, there is the problem such as aging circuit and destruction, serious accident may be caused to occur if do not make regular check on these problems and overhaul.

Artificial line walking detection efficiency is low, and dangerous high.Along with the development of unmanned air vehicle technique, gather high-tension line image these image informations are processed by unmanned plane technology of taking photo by plane, can personnel cost be reduced and ensure the safety of workmen, can increase work efficiency simultaneously.

Because electric power line pole tower position geographical environment is complicated, cause the image background also relative complex obtained, this causes difficulty to the identification of succeeding target and location, and be still in the starting stage based on the transmission line status detection technique of Aerial Images, can list of references and achievement in research less.

In prior art, class methods utilize colouring information, and use max-thresholds method, maximum variance between clusters are split coloured image.The shortcoming existed: the impact of this class methods light photograph is relatively more serious, and physical environment residing for transmission line of electricity is complicated, has trees, the complex background such as river, road, makes this class methods Detection accuracy not high.

Another kind of method is the oval information utilizing the schistose texture of insulator, detects ellipse by Hough transform.The shortcoming existed: due to the problem of shooting angle, there is the situation of blocking between sheet and sheet, cause metrical error.

Summary of the invention

For solving the deficiency that prior art exists, the invention discloses one and patrol and examine electric transmission line isolator recognition methods for unmanned plane efficiently, this technology utilizes degree of depth learning method by learning the transmission line of electricity image document of taking photo by plane, then line fitting method is utilized to calculate the angle information of insulator, the identification location technology requirement of insulator in line walking image that unmanned plane taken photo by plane can be completed, improve the accuracy to isolator detecting and robustness.

For achieving the above object, concrete scheme of the present invention is as follows:

Patrol and examine an insulator recognition methods for transmission line of electricity for unmanned plane, comprise the following steps:

Step one: image variants: extract the subimage for training and carry out rough handling from electric transmission line isolator image, forms training dataset;

Step 2: carry out packing process to what extract for the subimage of training, add the label that image is corresponding;

Step 3: the data of convolutional neural networks (CNN) algorithm to step one data centralization in utilizing the degree of depth to learn are trained, and obtain the detection model for insulator;

Step 4: insulator target area is detected: detect transmission line of electricity image, obtains the candidate frame of insulator target;

Step 5: carry out non-maxima suppression to candidate frame, obtains final insulator candidate frame;

Step 6: carry out fitting a straight line operation to the final insulator candidate frame obtained, obtain central point, the angle of candidate frame and size information, finally mark on electric transmission line isolator image.

In described step one, image variants process is: the area image extracting insulator part, shaft tower and background in electric transmission line isolator image, the image extracted is carried out convergent-divergent process, the insulator extracted is carried out to the rotation process of angle.

In described step 3, when carrying out model training with convolutional neural networks algorithm:

First the template parameter that initial training uses is set, comprising the number of plies of convolutional neural networks, the size of convolution kernel, whether the initial weight of each node, use down-sampling process, the number of every layer data input and output, activation function, the learning efficiency of the Gradient Descent of every layer of convolutional neural networks.

In described step 3, the concrete isolator detecting model training process based on convolutional neural networks comprises the following steps:

3-1) forward conduction (Feedforward Pass): divided by the coloured image of input RGB triple channel to extract Pixel Information, as the input information of convolutional neural networks;

3-2) backward conduction (BackPropagation Pass): the loss function between the class label that optimization forward conduction obtains and sample physical tags.

Described step 3-1) in, the structure of convolutional neural networks, uses the convolutional neural networks structured training template of six layers;

Ground floor is convolutional layer, and third layer is convolutional layer, and layer 5 is convolutional layer, and the convolution kernel of convolutional layer setting pixel size and input information are carried out convolution operation and obtained proper vector;

The second layer is down-sampled layer, 4th layer is down-sampled layer, down-sampled layer is according to the definition in parameterized template, carry out down-sampled operation, what adopt is get maximal value as output in the image block of the pixel size of setting, effectively on the basis of preserving useful information, reduces data processing amount by down-sampled process;

Layer 6 is full articulamentum: the proper vector obtained by layer 5 is integrated, and forms a long vector, is passed to the judgement of activation function acquisition to input amendment classification, selects maximum output valve as the label of input amendment.

Described step 3-2) in, suppose that being used for training set form is { (x ⁽¹⁾, y ⁽¹⁾), (x ⁽²⁾, y ⁽²⁾) ..., (x ⁽ⁿ⁾, y ⁽ⁿ⁾), x ⁽ⁱ⁾represent i-th training data, y ⁽ⁱ⁾represent data x ⁽ⁱ⁾corresponding data label, training dataset comprises n sample.

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

$J(W,b;x,y)=\frac{1}{2}\left|\right|h_{w,b}\left(x\right)-y|{|}^2---\left(1\right)$

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

$\begin{array}{c}J(W,b)=\[\frac{1}{n}\underset{i=1}{\overset{n}{\Σ}}J(W,b;x^{\left(i\right)},y^{\left(i\right)})\]+\frac{\mathrm{\γ}}{2}\underset{l=1}{\overset{n_l-1}{\Σ}}\underset{i=1}{\overset{s_l}{\Σ}}\underset{j=1}{\overset{s_{l+1}}{\Σ}}{\left(W_{ji}^{\left(l\right)}\right)}^2\\ =\[\frac{1}{n}\underset{i=1}{\overset{n}{\Σ}}\left(\frac{1}{2}{\left|\right|h_{w,b}(x^{\left(i\right)}-y^{\left(i\right)})\left|\right|}^2\right)\]+\frac{\mathrm{\γ}}{2}\underset{l=1}{\overset{n_l-1}{\Σ}}\underset{i=1}{\overset{s_l}{\Σ}}\underset{j=1}{\overset{s_{l+1}}{\Σ}}{\left(W_{ji}^{\left(l\right)}\right)}^2\end{array}---\left(2\right)$

In formula (2), s _l, s _l+1represent the number of l layer and l+1 layer neuron node, n _lrepresent the number of plies of neural network, represent in l group weight parameter the weight coefficient connecting l layer i-th node and a l+1 layer jth node, be the biased of l layer i-th node, Section 1 J (W, b; x ⁽ⁱ⁾, y ⁽ⁱ⁾) be a mean square deviation item, Section 2 is regularization term, is used for reducing the amplitude of weight, and prevent over-fitting, γ is control coefrficient.

With gradient descent method, parameter W and b is upgraded:

$\begin{array}{c}{W}_{ji}^{\left(l\right)}=W_{ji}^{\left(l\right)}-\mathrm{\α}\frac{\∂}{\∂W_{ji}^{\left(l\right)}}J(W,b)\\ b_i^{\left(l\right)}=b_i^{\left(l\right)}-\mathrm{\α}-\frac{\∂}{\∂b_i^{\left(l\right)}}J(W,b)\end{array}---\left(3\right)$

Wherein, the α in formula (3) is learning rate, for controlling the speed of Gradient Descent.

In described step 4, use convolutional neural networks to carry out the detection of insulator, method detailed is as follows:

4-1) extract the convolutional neural networks template trained, comprise weight, be biased and train the network structure used; Then according to these parameter initialization test procedures;

4-2) be loaded into image, the image collected due to unmanned plane is comparatively large, carries out convergent-divergent to accelerate follow-up computing to image, in order to locate the position of insulator accurately, add multi-scale method, in the enterprising line slip frame operation of multiple yardstick, obtain concrete target image block;

4-3) using target image block as input, carry out forward conduction operation, obtain the generic of object block;

4-4) preserve other object block information of insulator class, comprising start position and the length and width information of object block.

In described step 6, use line fitting method to mark insulator, detailed process is as follows:

The m of t classification 6-1) obtained in obtaining step five _i, i={1 ..., t} candidate frame information, calculates the center position of each candidate frame, and preserves;

6-2) center position is (P, Q), can carry out matching by the mode of linear fit to the center of each classification, the position of each insulator in accurate location, can solve this problem here with unitary linear fit:

Y＝kX+b' (4)

6-3) with least square fitting method, calculate the straight line that can reflect X and Y relation.

Described step 6-3) in, definition loss function is:

$L=\underset{j=1}{\overset{m_i}{\Σ}}(q_j-(b^{\′}+{\mathrm{kp}}_j){)}^2---\left(5\right)$

Wherein, (p _j, q _j) for belonging to the central point of a jth rectangle frame of i class.

By asking local derviation to b' and k, obtain the optimum solution of b' and k:

$\begin{array}{c}\frac{\∂L}{\∂b^{\′}}=-2\underset{j=1}{\overset{m_i}{\Σ}}(q_j-(b^{\′}+{\mathrm{kp}}_j))\\ \frac{\∂L}{\∂k}=-2\underset{j=1}{\overset{m_i}{\Σ}}(q_j-(b^{\′}+{\mathrm{kp}}_j))p_j\end{array}---\left(6\right)$

The last solution obtaining k and b' is

$\begin{array}{c}k=\frac{m_i*\underset{j=1}{\overset{m_i}{\Σ}}{p}_j{q}_j-\underset{j=1}{\overset{m_i}{\Σ}}{p}_j\underset{j=1}{\overset{m_i}{\Σ}}{q}_j}{m_i*\underset{j=1}{\overset{m_i}{\Σ}}{p}_j^2-\underset{j=1}{\overset{m_i}{\Σ}}{q}_j^2}\\ b^{\′}=\frac{m_i*\underset{j=1}{\overset{m_i}{\Σ}}{p}_j^2{q}_j-\underset{j=1}{\overset{m_i}{\Σ}}{p}_j\underset{j=1}{\overset{m_i}{\Σ}}{q}_j}{m_i*\underset{j=1}{\overset{m_i}{\Σ}}{p}_j^2-\underset{j=1}{\overset{m_i}{\Σ}}{p}_j^2}\end{array}---\left(7\right).$

Wherein, (p _j, q _j) for belonging to the central point of a jth rectangle frame of i class, m _irepresent the number of the rectangle frame of the i-th class.

Beneficial effect of the present invention:

The application adopts convolutional layer, down-sampled layer and full linking layer: after the two-layer down-sampled operation of three-layer coil sum, all characteristic blocks are carried out full attended operation, so just obtain the final feature interpretation to image block.

Effective recognition technology of electric transmission line isolator effectively can realize the location to the insulator target of patrolling and examining in image, for subsequent defective diagnosis provides basis.Meanwhile, this technology also can be screened the image of patrolling and examining acquisition, alleviates the burden of artificial examination, has broad application prospects.

Accompanying drawing explanation

Fig. 1 CNN structural drawing;

The down-sampled process of Fig. 2 Convolution sums;

Fig. 3 unmanned plane is taken photo by plane transmission line of electricity image;

Transmission line of electricity image after Fig. 4 candidate frame mark;

The final isolator detecting result images of Fig. 5;

Fig. 6 trains process flow diagram;

Fig. 7 target detection and positioning flow figure.

Embodiment:

Below in conjunction with accompanying drawing, the present invention is described in detail:

The training process flow diagram of experimentation as shown in Figure 6, extracts data set, and data set is packed, start training process, carry out training hierarchical structure and initiation parameter before training, during training, train forward, reverse feedback, judge maximum iteration time/accuracy requirement, if so, then output template, otherwise, continue training.

Target identification and positioning flow figure as shown in Figure 7, input picture, initialization detection model, sliding window extract subgraph, judge whether it is insulator, if not, return sliding window and extract subgraph process, if, store candidate frame, matching candidate frame, in former figure mark position, export.

One patrols and examines electric transmission line isolator recognition methods for unmanned plane efficiently, and concrete steps comprise:

1) training set: extract the subimage for training from transmission line of electricity image, forms training dataset.

2) packing process is carried out to the insulator image extracted, add the label of correspondence image;

3) convolutional neural networks (CNN) algorithm in utilizing the degree of depth to learn is trained data, obtains the detection model for insulator;

4) insulator target area is detected: detect transmission line of electricity image, obtains the candidate frame of insulator target;

5) non-maxima suppression is carried out to candidate frame, obtain final insulator candidate frame.

6) carry out fitting a straight line operation to the candidate frame obtained, obtain central point, the angle of frame and size information, finally mark on former figure.

Described step 1) image acquisition process be: the area image extracting insulator part, shaft tower and background in former figure, in order to the requirement of satisfied training, the image extracted is carried out convergent-divergent process, and size zooms to 64*64 pixel.In a practical situation, because shooting angle reason obtains the situation that shaft tower image may exist inclination, for this carries out low-angle rotation process to the insulator extracted, increase the diversity of data set, improve the robustness of training pattern.

Described step 3) in step, the concrete grammar carrying out model training with CNN algorithm is as follows:

(1) first the template parameter that initial training uses will be set, comprising the number of plies of convolutional neural networks, the size of convolution kernel, the initial weight of each node, whether use down-sampling process, the number of every layer data input and output, activation function etc., also have the setting of the learning efficiency of the Gradient Descent of corresponding templates.

(2) network structure, we use the convolutional neural networks structured training template of six layers here.Form is as shown in CNN structural drawing in Fig. 1.

The concrete isolator detecting model training process based on CNN mainly contains following two steps:

A. forward conduction: divided by the coloured image of input RGB triple channel to extract Pixel Information, as the input information of convolutional network, ground floor is convolutional layer (Convolution), carries out convolution operation by the convolution kernel of 5*5 pixel size and input information.Original signal can be made to strengthen by convolution operation, and reduce the impact of noise.The second layer is down-sampled layer, according to the definition in parameterized template, carries out down-sampled operation (SubSampling/Pooling), and what adopt here is get maximal value as output in the image block of 2*2.Effectively data processing amount can be reduced on the basis of preserving useful information by down-sampled process; Third layer convolutional layer, the 4th layer of down-sampled layer, layer 5 convolutional layer carry out similar operation; Layer 6 is full articulamentum: the proper vector obtained by layer 5 is integrated, and forms a long vector, is passed to the judgement of activation function acquisition to input amendment classification.As shown in Figure 2.

Below convolutional layer, down-sampled layer and full linking layer are described further:

Input color image, carry out convolution algorithm with a trainable wave filter fx and image, then add a bigoted bx, this has just extracted a characteristics of image, obtains the Cx of convolutional layer; In order to reduce data volume, in four neighborhood local, pixel gets maximal value change as exporting, and then by the nonlinear transformation of the complete paired data of ReLu activation function, nonlinear transformation can reduce the linear relationship between feature and feature, the expressive faculty of Enhanced feature.And after the process by down-sampled like this layer, the data volume obtained is only 1/4th of convolutional layer.

After the two-layer down-sampled operation of three-layer coil sum, all characteristic blocks are carried out full attended operation, so just obtain the final feature interpretation to image block.In order to the classification of predicted picture, the Feature Descriptor obtained is carried out activation function calculation process, select maximum output valve as the label of image block.

B. to conduction after: the class label obtained by Feedforward Pass and sample physical tags carry out counting loss function.Suppose that we are { (x for training set form ⁽¹⁾, y ⁽¹⁾), (x ⁽²⁾, y ⁽²⁾) ..., (x ⁽ⁿ⁾, y ⁽ⁿ⁾), training dataset comprises n sample, and for single sample (x, y), the result learnt is h _w,b(x), its loss function is:

$J(W,b;x,y)=\frac{1}{2}\left|\right|h_{w,b}\left(x\right)-y|{|}^2---\left(1\right)$

So overall loss function is:

$\begin{array}{c}J(W,b)=\[\frac{1}{n}\underset{i=1}{\overset{n}{\Σ}}J(W,b;x^{\left(i\right)},y^{\left(i\right)})\]+\frac{\mathrm{\γ}}{2}\underset{l=1}{\overset{n_l-1}{\Σ}}\underset{i=1}{\overset{s_l}{\Σ}}\underset{j=1}{\overset{s_{l+1}}{\Σ}}{\left(W_{ji}^{\left(l\right)}\right)}^2\\ =\[\frac{1}{n}\underset{i=1}{\overset{n}{\Σ}}\left(\frac{1}{2}{\left|\right|h_{w,b}(x^{\left(i\right)}-y^{\left(i\right)})\left|\right|}^2\right)\]+\frac{\mathrm{\γ}}{2}\underset{l=1}{\overset{n_l-1}{\Σ}}\underset{i=1}{\overset{s_l}{\Σ}}\underset{j=1}{\overset{s_{l+1}}{\Σ}}{\left(W_{ji}^{\left(l\right)}\right)}^2\end{array}---\left(2\right)$

Formula (2) s _lrepresent neuronic number in l layer, represent in l group weight parameter the weight coefficient connecting l layer i-th node and a l+1 layer jth node.Section 1 J (W, b; x ⁽ⁱ⁾, y ⁽ⁱ⁾) be a mean square deviation item, Section 2 is regularization term, is used for reducing the amplitude of weight, and prevent over-fitting, γ is control coefrficient.Upgrade by each iteration of gradient descent method and parameter W and b upgraded:

$\begin{array}{c}{W}_{ji}^{\left(l\right)}=W_{ji}^{\left(l\right)}-\mathrm{\α}\frac{\∂}{\∂W_{ji}^{\left(l\right)}}J(W,b)\\ b_i^{\left(l\right)}=b_i^{\left(l\right)}-\mathrm{\α}\frac{\∂}{\∂b_i^{\left(l\right)}}J(W,b)\end{array}---\left(3\right)$

α in formula (3) is learning rate, for controlling the speed of Gradient Descent.

In false code below, Δ W ^(l)one and matrix W ^(l)the matrix that dimension is identical, Δ b ^(l)one and b ^(l)the vector that dimension is identical.Below, the iterative process realized in gradient descent method is provided:

1. couple all layer l, make Δ W ^(l):=0, Δ b ^(l):=0;

2. couple i=1 to n:

A. back-propagation algorithm is used to calculate

B. calculate ${\mathrm{\ΔW}}^{\left(l\right)}:={\mathrm{\ΔW}}^{\left(l\right)}+{\▿}_{w^{\left(l\right)}}J(W,b;x,y);$

C. calculate ${\mathrm{\Δb}}^{\left(l\right)}:={\mathrm{\Δb}}^{\left(l\right)}+{\▿}_{b^{\left(i\right)}}J(W,b;x,y).$

3. upgrade weight parameter:

$W^{\left(l\right)}=W^{\left(l\right)}-\mathrm{\α}\[\frac{1}{n}{\mathrm{\ΔW}}^{\left(l\right)}\]+{\mathrm{\γW}}^{\left(l\right)}$

$b^{\left(l\right)}=b^{\left(l\right)}-\mathrm{\α}\[\frac{1}{n}{\mathrm{\Δb}}^{\left(l\right)}\]$

So, just recursive use gradient descent method carrys out the value that iterative computation reaches reduction loss function J (W, b), and then solves whole neural network.

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

(1) extract the CNN template trained, comprise weight, be biased and train the network structure used; Then according to these parameter initialization test procedure frameworks;

(2) be loaded into image (as shown in Figure 3), due to the image comparatively large (5184*3456) that unmanned plane collects, convergent-divergent carried out to accelerate follow-up computing to image.In order to locate the position of insulator accurately, adding multi-scale method, in the enterprising line slip frame operation of multiple yardstick, obtaining concrete target image block.

(3) using target image block as input, carry out CNN Feedforward Pass and operate, obtain the generic of object block.(4) preserve other object block information of insulator class, comprising start position and the length and width information of object block, Fig. 4 is the image through marking insulator;

Described step 6) in, use approximating method to mark insulator, detailed process is as follows:

(1) m of t classification that obtains of obtaining step (5) _i, i={1 ..., t} candidate frame information, calculates the central point position of each frame

Put, and preserve.

(2) center position is (P, Q), can carry out matching by the mode of linear fit to the center of each classification, the position of each insulator in accurate location.Here this problem can be solved with unitary linear fit:

Y＝kX+b' (4)

(3) with least square fitting method, the straight line that can reflect X and Y relation is calculated:

Definition loss function is:

$L=\underset{j=1}{\overset{m_i}{\Σ}}(q_j-(b^{\′}+{\mathrm{kp}}_j){)}^2---\left(5\right)$

By asking local derviation to b' and k, obtain the optimum solution of b' and k:

$\begin{array}{c}\frac{\∂L}{\∂b^{\′}}=-2\underset{j=1}{\overset{m_i}{\Σ}}(q_j-(b^{\′}+{\mathrm{kp}}_j))\\ \frac{\∂L}{\∂k}=-2\underset{j=1}{\overset{m_i}{\Σ}}(q_j-(b^{\′}+{\mathrm{kp}}_j))p_j\end{array}---\left(6\right)$

The last solution obtaining k and b' is

$\begin{array}{c}k=\frac{m_i*\underset{j=1}{\overset{m_i}{\Σ}}{p}_j{q}_j-\underset{j=1}{\overset{m_i}{\Σ}}{p}_j\underset{j=1}{\overset{m_i}{\Σ}}{q}_j}{m_i*\underset{j=1}{\overset{m_i}{\Σ}}{p}_j^2-\underset{j=1}{\overset{m_i}{\Σ}}{q}_j^2}\\ b^{\′}=\frac{m_i*\underset{j=1}{\overset{m_i}{\Σ}}{p}_j^2{q}_j-\underset{j=1}{\overset{m_i}{\Σ}}{p}_j\underset{j=1}{\overset{m_i}{\Σ}}{q}_j}{m_i*\underset{j=1}{\overset{m_i}{\Σ}}{p}_j^2-\underset{j=1}{\overset{m_i}{\Σ}}{p}_j^2}\end{array}---\left(7\right).$

So just determine the angle information of insulator, on former figure, mark the particular location of insulator according to angle information.As the insulator marking image that Fig. 5 is final.

By reference to the accompanying drawings the specific embodiment of the present invention is described although above-mentioned; but not limiting the scope of the invention; one of ordinary skill in the art should be understood that; on the basis of technical scheme of the present invention, those skilled in the art do not need to pay various amendment or distortion that creative work can make still within protection scope of the present invention.

Claims (10)

1. patrol and examine an insulator recognition methods for transmission line of electricity for unmanned plane, it is characterized in that, comprise the following steps:

Step one: image variants: extract the subimage for training and carry out rough handling from electric transmission line isolator image, forms training dataset;

Step 2: carry out packing process to what extract for the subimage of training, add the label of correspondence image;

Step 3: the data of convolutional neural networks algorithm to step one data centralization in utilizing the degree of depth to learn are trained, and obtain the detection model for insulator;

Step 4: insulator target area is detected: detect transmission line of electricity image, obtains the candidate frame of insulator target;

Step 5: carry out non-maxima suppression to candidate frame, obtains final insulator candidate frame;

Step 6: carry out fitting a straight line operation to the final insulator candidate frame obtained, obtain central point, the angle of candidate frame and size information, finally mark on electric transmission line isolator image.

2. a kind of insulator recognition methods of patrolling and examining transmission line of electricity for unmanned plane as claimed in claim 1, it is characterized in that, in described step one, image variants process is: the area image extracting insulator part, shaft tower and background in electric transmission line isolator image, the image extracted is carried out convergent-divergent process, the insulator extracted is carried out to the rotation process of angle.

3. a kind of insulator recognition methods of patrolling and examining transmission line of electricity for unmanned plane as claimed in claim 1, is characterized in that, in described step 3, when carrying out model training with convolutional neural networks algorithm:

First the template parameter that initial training uses is set, comprising the number of plies of convolutional neural networks, the size of convolution kernel, the initial weight of each node, whether use down-sampling process, the number of every layer data input and output, activation function, the learning efficiency of the Gradient Descent of every layer of convolutional neural networks.

4. a kind of insulator recognition methods of patrolling and examining transmission line of electricity for unmanned plane as described in claim 1 or 3, is characterized in that, in described step 3, the concrete isolator detecting model training process based on convolutional neural networks comprises the following steps:

3-1) forward conduction: divided by the coloured image of input RGB triple channel to extract Pixel Information, as the input information of convolutional neural networks;

3-2) backward conduction: the class label obtained by forward conduction and sample physical tags carry out counting loss function.

5. a kind of insulator recognition methods of patrolling and examining transmission line of electricity for unmanned plane as claimed in claim 4, it is characterized in that, described step 3-1) in, the structure of convolutional neural networks, use the convolutional neural networks structured training template of six layers, ground floor is convolutional layer, and third layer is convolutional layer, layer 5 is convolutional layer, and the convolution kernel of convolutional layer setting pixel size and input information are carried out convolution operation and obtained proper vector;

The second layer is down-sampled layer, 4th layer is down-sampled layer, down-sampled layer is according to the definition in parameterized template, carry out down-sampled operation, what adopt is get maximal value as output in the image block of the pixel size of setting, effectively on the basis of preserving useful information, reduces data processing amount by down-sampled process;

Layer 6 is full articulamentum: the proper vector obtained by layer 5 is integrated, and forms a long vector, is passed to the judgement of activation function acquisition to input amendment classification, selects maximum output valve as the label of image block.

6. a kind of insulator recognition methods of patrolling and examining transmission line of electricity for unmanned plane as claimed in claim 4, is characterized in that, described step 3-2) in, suppose that being used for training set form is { (x ⁽¹⁾, y ⁽¹⁾), (x ⁽²⁾, y ⁽²⁾) ..., (x ⁽ⁿ⁾, y ⁽ⁿ⁾), x ⁽ⁱ⁾represent i-th training data, y ⁽ⁱ⁾represent data x ⁽ⁱ⁾corresponding data label, training dataset comprises n sample;

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

$J(W,b;x,y)=\frac{1}{2}\left|\right|h_{w,b}\left(x\right)-y|{|}^2---\left(1\right)$

So overall loss function is:

$\begin{array}{c}J(W,b)=\[\frac{1}{n}\underset{i=1}{\overset{n}{\Σ}}J(W,b;x^{\left(i\right)},y^{\left(i\right)})\]+\frac{\mathrm{\γ}}{2}\underset{l=1}{\overset{n_l-1}{\Σ}}\underset{i=1}{\overset{s_l}{\Σ}}\underset{j=1}{\overset{s_l+1}{\Σ}}{\left(W_{ji}^{\left(l\right)}\right)}^2\\ =\[\frac{1}{n}\underset{i=1}{\overset{n}{\Σ}}\left(\frac{1}{2}\right|\left|h_{w,b}\right(x^{\left(i\right)})-y^{\left(i\right)}\left|{|}^2\right)\]+\frac{\mathrm{\γ}}{2}\underset{l=1}{\overset{n_l-1}{\Σ}}\underset{i=1}{\overset{s_l}{\Σ}}\underset{j=1}{\overset{s_l+1}{\Σ}}{\left(W_{ji}^{\left(l\right)}\right)}^2\end{array}---\left(2\right)$

Formula (2) s _lrepresent neuronic number in l layer, represent in l group weight parameter the weight coefficient connecting l layer i-th node and a l+1 layer jth node, be the biased of l layer i-th node, Section 1 J (W, b; x ⁽ⁱ⁾, y ⁽ⁱ⁾) be a mean square deviation item, Section 2 is regularization term, is used for reducing the amplitude of weight, and prevent over-fitting, γ is control coefrficient.

7. a kind of insulator recognition methods of patrolling and examining transmission line of electricity for unmanned plane as claimed in claim 6, is characterized in that, upgrades upgrade parameter W and b by each iteration of gradient descent method:

$\begin{array}{c}{W}_{ji}^{\left(l\right)}=W_{ji}^{\left(l\right)}-\mathrm{\α}\frac{\∂}{\∂W_{ji}^{\left(l\right)}}J(W,b)\\ b_i^{\left(l\right)}=b_i^{\left(l\right)}-\mathrm{\α}\frac{\∂}{\∂b_i^{\left(l\right)}}J(W,b)\end{array}---\left(3\right)$

Wherein, the α in formula (3) is learning rate, for controlling the speed of Gradient Descent.

8. a kind of insulator recognition methods of patrolling and examining transmission line of electricity for unmanned plane as claimed in claim 1, is characterized in that, in described step 4, use convolutional neural networks to carry out the detection of insulator, method detailed is as follows:

4-1) extract the convolutional neural networks template trained, comprise weight, be biased and train the network structure used; Then according to these parameter initialization test procedures;

4-2) be loaded into image, the graphical rule collected due to unmanned plane is comparatively large, carries out convergent-divergent to accelerate follow-up computing to image, in order to locate the position of insulator accurately, add multi-scale method, in the enterprising line slip frame operation of multiple yardstick, obtain concrete target image block;

4-3) using target image block as input, carry out forward conduction operation, obtain the generic of object block;

4-4) preserve other object block information of insulator class, comprising start position and the length and width information of object block.

9. a kind of insulator recognition methods of patrolling and examining transmission line of electricity for unmanned plane as claimed in claim 1, is characterized in that, in described step 6, use line fitting method to mark insulator, detailed process is as follows:

The m of t classification 6-1) obtained in obtaining step five _i, i={1 ..., t} candidate frame information, calculates the center position of each candidate frame, and preserves;

6-2) center position is (P, Q), can carry out matching by the mode of linear fit to the center of each classification, the position of each insulator in accurate location, solves this problem with unitary linear fit:

Y＝kX+b' (4)

6-3) with least square fitting method, calculate the straight line that can reflect X and Y relation.

10. a kind of insulator recognition methods of patrolling and examining transmission line of electricity for unmanned plane as claimed in claim 9, is characterized in that, described step 6-3) in, definition loss function is:

$L=\underset{j=1}{\overset{m_i}{\Σ}}(q_j-(b^{\′}+{\mathrm{kp}}_j){)}^2---\left(5\right)$

By asking local derviation to b' and k, obtain the optimum solution of b' and k:

$\begin{array}{c}\frac{\∂L}{\∂b^{\′}}=-2\underset{j=1}{\overset{m_i}{\Σ}}(q_j-(b^{\′}+{\mathrm{kp}}_j))\\ \frac{\∂L}{\∂k}=-2\underset{j=1}{\overset{m_i}{\Σ}}(q_j-(b^{\′}+{\mathrm{kp}}_j))p_j\end{array}---\left(6\right)$

The last solution obtaining k and b' is

$\begin{array}{c}k\frac{m_i*\underset{j=1}{\overset{m_i}{\Σ}}{p}_j{q}_j-\underset{j=1}{\overset{m_i}{\Σ}}{p}_j\underset{j=1}{\overset{m_i}{\Σ}}{q}_j}{m_i*\underset{j=1}{\overset{m_i}{\Σ}}{p}_j^2-\underset{j=1}{\overset{m_i}{\Σ}}{q}_j^2}\\ b^{\′}=\frac{m_i*\underset{j=1}{\overset{m_i}{\Σ}}{p}_j^2{q}_j-\underset{j=1}{\overset{m_i}{\Σ}}{p}_j\underset{j=1}{\overset{m_i}{\Σ}}{q}_j}{m_i*\underset{j=1}{\overset{m_i}{\Σ}}{p}_j^2-\underset{j=1}{\overset{m_i}{\Σ}}{p}_j^2}\end{array}---\left(7\right);$

Wherein, (p _j, q _j) for belonging to the central point of a jth rectangle frame of i class, m _irepresent the number of the rectangle frame of the i-th class.