CN109344766A - Identification method of slider-type circuit breaker based on inspection robot - Google Patents

Abstract

The slide block type breaker recognition methods based on crusing robot that the invention proposes a kind of.The present invention is broadly divided into 5 steps: (1) utilizing picture number collection training SVM multi-categorizer；(2) crusing robot reaches specified inspection point and obtains picture to be detected；(3) coarse positioning and accurate positioning are carried out to target area, screening object candidate area obtains slide block type breaker；(4) to slide block type breaker progress image preprocessing is got, connection maximum two regions of area are extracted；(5) the HOG feature of two connected regions is extracted respectively, and is sent to SVM multi-categorizer and is obtained final recognition result.The present invention utilizes machine learning, and slide block type breaker detection identification mission can be efficiently accomplished under the conditions of different illumination, posture, the gentle accuracy rate of Automated water of image recognition under complex environment is improved, reduces missing inspection, erroneous detection problem to greatest extent.

Description

Identification method of slider-type circuit breaker based on inspection robot

technical field

The invention relates to a target detection technology, in particular to a method for identifying a slider-type circuit breaker based on an inspection robot.

Background technique

The power industry is closely related to people's lives. The slider-type circuit breaker in the substation is the most basic device in the power industry, and it is very important to the power supply. In recent years, the lack of detection and identification of the slider-type circuit breaker has resulted in the failure of normal transmission of electricity, which has caused huge economic losses to people's lives and industrial production.

At present, there are two main types of circuit breaker detection methods. The first is the manual inspection method. However, because most of the circuit breakers in substations exist in the wild, the staff are generally far away, and when problems occur, they cannot be solved in time, resulting in the failure of the power supply system to respond in time. In addition, manual inspection and inspection often consume a lot of manpower and time, and are prone to errors in a long-term, high-intensity working environment. Therefore, the manual inspection detection method has the disadvantages of high labor intensity, low efficiency, insufficient inspection inspection, poor reliability, and high risk. In recent years, with the promotion of inspection robots, the detection work of slider-type circuit breakers has gradually developed towards an intelligent direction. Using electric inspection robots to replace manual inspection has the advantages of high efficiency and high reliability. However, most of the current methods use traditional image processing methods for detection and recognition. In the case of changing lighting conditions, the detection effect is not good. Generally, a set of parameters is required for one lighting condition, which requires a more general method. Detection and recognition methods to deal with detection tasks under different lighting and attitude conditions.

SUMMARY OF THE INVENTION

The purpose of the present invention is to propose a method for identifying a slider-type circuit breaker based on an inspection robot, so as to solve the problems in the existing slider-type circuit breaker detection and identification technology that the robot's position is not timed, the target scale and angle change greatly, and the target It is greatly affected by light, which leads to the problem of inaccurate detection and recognition.

The technical solution for realizing the present invention is: a method for identifying a sliding block circuit breaker based on an inspection robot, and the specific steps are:

Step 1. For each inspection point, select an image in the center of the circuit breaker taken at the inspection point as a template image, and use the slider-type circuit breaker image number set collected in advance to train the SVM multi-classifier;

Step 2. The inspection robot reaches the designated inspection point through positioning and navigation, obtains the image of the on-site slider-type circuit breaker and reads it in the form of a grayscale image, which is used for the detection and identification of the slider-type circuit breaker;

Step 3. Perform coarse positioning and precise positioning on the target area to be detected, and filter the target candidate area to obtain the slider-type circuit breaker image;

Step 4. Preprocess the acquired image of the on-site slider-type circuit breaker, extract the two areas with the largest connected area, and divide them into left and right parts according to their positions;

Step 5. Perform pixel adjustment on the two separated areas respectively, slide on the image with a sliding window with a length of m pixels and a width of n pixels, extract the HOG feature from the window, and send the calculated HOG feature operator to the SVM multi-classification get the final recognition result.

Preferably, the specific method for training the SVM multi-classifier in step 1 is:

Step 1-1. Collect the slider-type circuit breaker image set in advance as a set of positive and negative training samples;

Step 1-2, extract the HOG features of the positive and negative training sample sets;

Steps 1-3: Assign sample labels to all positive and negative training sample sets, and send the HOG features and sample labels of the training sample set to the SVM for training.

Preferably, the specific method for preparing the training sample set in step 1-1 is:

(1) Collect the pictures of the slider-type circuit breaker, take the pictures in the energy storage state as the positive sample set, and take the pictures in the non-energy storage state as the negative sample set;

(2) Crop the picture and delete the redundant information outside the slider display window area on the slider-type circuit breaker;

(3) Scale the image to be m pixels long and n pixels wide.

Preferably, the specific method for extracting the HOG feature of the positive and negative training sample set in step 1-2 is:

(1) Convert the color image to a grayscale image;

(2) Gamma correction is performed on the grayscale image to reduce local shadow and illumination changes in the image. The formula for Gamma correction is:

I(x,y)=I(x,y) ^gamma (1)

Among them, I(x,y) represents the pixel value of the xth row and the yth column of the image, and gamma takes a number between 0-1;

(3) Calculate the gradient of each pixel of the image according to the following formula:

G _x (x,y)=H(x+1,y)-H(x-1,y) (2)

G _y (x,y)=H(x,y+1)-H(x,y-1) (3)

Among them, G _x (x, y), G _y (x, y), H (x, y) represent the horizontal gradient, vertical gradient and pixel value at the pixel point (x, y) in the image, respectively. The gradient magnitude G(x,y) at (x,y) and the gradient direction α(x,y) can be obtained according to the following formula:

(4) Divide the image into square cells with a side length of a pixels, a is the greatest common factor of m and n, create a gradient direction histogram for each cell, and divide the gradient direction 360 degrees into k directions block, the direction range of the ith direction block is Count the gradient direction of each pixel in the cell. If the gradient direction belongs to a certain direction block, the count value of the corresponding direction block is added to the amplitude corresponding to this gradient;

(5) Combine the cells into blocks, and the normalized gradient histogram in the block rewrites the gradient histogram corresponding to each cell into a vector form, and concatenates all the gradient vectors in each block to form the gradient of the block Direction histogram vector; multiply the vector by the corresponding normalization factor, the calculation formula of the normalization factor is:

Among them, v represents a vector that has not been normalized, ||v|| ₂ represents the second-order norm of v, and e represents a constant;

(6) Concatenate the normalized vectors of all blocks in the image to obtain the HOG feature of the training sample set.

Preferably, in steps 1-3, the specific method for sending the HOG features and sample labels of the positive and negative training sample sets into the SVM for training is:

(1) The training goal of SVM is to find an optimal hyperplane that can classify positive and negative samples, and its mathematical form can be expressed as:

Among them, w represents the vector perpendicular to the hyperplane, ||w|| represents the norm of w, ξ _i represents the slack variable, which is a non-negative number, D is a parameter used to control the weight of the two terms in the objective function, x _i represents the HOG feature of the ith sample, _yi represents the sample label of the ith sample, and b represents a constant;

(2) Build the Lagrangian function:

where α _i represents the Lagrange multiplier, r _i =D-α _i , then let

The objective function is transformed into

Among them, d ^* represents the optimal value of the objective function;

(3) Minimize L with respect to w, b, ξ, namely:

Putting Equation (11) into Equation (8), the objective function is transformed into:

Among them, < _xi ,x _j > means to find the inner product of x _i , x _j ;

(4) Use the SMO algorithm to find the optimal value of the Lagrangian multiplier α _i , and use the heuristic algorithm to select a pair of Lagrangian multipliers α _i , α _j ; fix other parameters except α _i , α _j , determine the value of α _i under the condition that w takes the extreme value, and use α _i to represent α _j ; keep repeating until the objective function converges;

(5) Determine the optimal hyperplane according to the optimal value of the Lagrange multiplier:

in, Represents the optimal value of the Lagrange multiplier, w ^* , b ^* represent the direction of the optimal hyperplane and the offset from the origin, respectively;

(6) Obtain the classification decision function, that is, the trained SVM classifier:

.

Preferably, the specific steps for locating and screening the target area in step 3 are:

Step 3-1. Use Merlin Fourier transform and phase correlation technology to roughly locate the target circuit breaker area in the image to be detected;

Step 3-2, using the machine learning method to accurately locate the target circuit breaker area, send the image to be detected into the trained classifier, and obtain several target candidate areas;

Step 3-3, respectively intersect each target candidate area and the rough positioning target slider-type circuit breaker area and compare the parameter IOU; perceive each target candidate area image and the slider-type circuit breaker area image in the template image Hash calculation to obtain the perceptual hash index; calculate the mutual information index of each target candidate region image and the template image, and filter the target candidate region to obtain a slider circuit breaker.

Preferably, in step 3-3, the specific calculation methods of perceptual hash index, intersection ratio parameter IOU and mutual information index I(G _(X) , H _(Y) ) are respectively:

(1) Scale the target candidate area image and the template image to the same size, perform cosine transformation, select the low-frequency area in the upper left corner of the cosine-transformed image, remove the DC component of the coordinate (0,0) to obtain the feature vector, and calculate the target candidate area. The Hamming distance of the feature vector of the image and the template image, as a perceptual hash index;

(2) The specific calculation formula of the intersection ratio parameter IOU is:

In the formula, C is the rough positioning target circuit breaker area, and n _i is the target candidate area;

(3) The calculation formula of the mutual information index I(G _(X) ,H _(Y) ) is:

G _(X) and H _(Y) are the number of grayscale pixels in the template image and the candidate image, respectively, and W and H are the width and height of the candidate region image, respectively.

Preferably, in step 3-3, the specific method for screening the target candidate area to obtain the slider-type circuit breaker is as follows:

The confidence of each candidate area is calculated by weighting the intersection ratio IOU, mutual information and perceptual hash pHash of each candidate area, where D is a constant:

Confidence=1-(pHash+1/I(G _(X) ,H _(y) ))/(IOU+D)(20)

Sort all candidate regions in descending order of confidence, find the region with the highest confidence, and use this region as the candidate detection result. If the IOU of the alternative detection result is less than the set threshold thresholdIOU and (pHash+1/I(G _(X) ,H _(Y) )) is greater than the threshold thresholdA, the coarse positioning target circuit breaker area determined in step 3-2 As the final target, otherwise the candidate detection result is used as the final target. The threshold value thresholdIOU ranges from 0.1 to 0.4, and the threshold value thresholdA ranges from 10 to 50.

Preferably, the specific method of image preprocessing in step 4 is:

Step 4-1. Perform histogram equalization on the grayscale image to increase the overall contrast of the image and make the image clearer;

Step 4-2, perform Gaussian filtering on the image to eliminate Gaussian noise on the image;

Step 4-3, perform Otsu binarization on the image to distinguish the target image from the background image;

Step 4-4, perform an open operation on the image to smooth the boundary of the target image, eliminate small spikes, and disconnect narrow connections;

Step 4-5, perform contour scanning on the image, extract the two areas with the largest connected area, and divide them into left and right parts according to their positions.

Compared with the prior art, the present invention has the following significant advantages: (1) the present invention can monitor and capture the information of the slider-type circuit breaker in real time, automatically identify the state of the circuit breaker in the power system, increase the automation level of the power inspection robot, and improve the Work efficiency; (3) the present invention integrates robot positioning information and machine learning, so that the position repeatability is high (for example, the positioning is less than 5cm), and the changes in scale and rotation are small; (2) the present invention is under different lighting and attitude conditions. It can effectively complete the detection and recognition task of slider-type circuit breakers, improve the recognition accuracy of images in complex environments, and minimize the problems of missed detection and false detection.

Description of drawings

FIG. 1 is a schematic flow chart of the present invention.

Figure 2 is a schematic diagram of the collected template image.

Figure 3 is an image of the slider-type circuit breaker after preprocessing.

Figure 4 is the image of the slider-type circuit breaker after extracting the connected area, in which Figure 4(a) is the left part of the image after the slider-type circuit breaker extracts the connected area, and Figure 4(b) is the right-side image of the slider-type circuit breaker after extracting the connected area part of the image.

Detailed ways

The present invention will be described in further detail below with reference to the accompanying drawings.

As shown in Figure 1, a method for identifying a slider-type circuit breaker based on an inspection robot, the specific steps are:

Step 1. For each inspection point, select an image in the center of the circuit breaker taken at the inspection point as a template image, and use the number of slider-type circuit breaker images collected in advance to train the SVM multi-classifier. The specific method is as follows:

Step 1-1. Collect the number of pointer-type circuit breaker images in advance as a set of positive and negative training samples: (1) Collect the pictures of the pointer-type circuit breaker. For the circular picture showing the opening and closing state, the picture in the open state is used as the positive and negative training sample set: (1) Collect the pictures of the pointer circuit breaker For the sample set, the pictures in the combined state are regarded as the negative sample set; for the circular pictures showing the energy storage state, the pictures in the energy storage state are regarded as the positive sample set, and the pictures in the non-energy storage state are regarded as the negative sample set;

(2) Crop the picture and delete the redundant information outside the circular area on the pointer circuit breaker, specifically:

(3) Scale the image to be m pixels long and n pixels wide.

Step 1-2, extract the HOG features of the positive and negative training sample sets, specifically:

(1) Convert the color image to a grayscale image;

(2) Gamma correction is performed on the grayscale image to reduce local shadow and illumination changes in the image. The formula for Gamma correction is:

I(x,y)=I(x,y) ^gamma (1)

Among them, I(x,y) represents the pixel value of the xth row and the yth column of the image, and gamma takes a number between 0-1;

(3) Calculate the gradient of each pixel of the image according to the following formula:

G _x (x,y)=H(x+1,y)-H(x-1,y) (2)

G _y (x,y)=H(x,y+1)-H(x,y-1) (3)

Among them, G _x (x, y), G _y (x, y), H (x, y) represent the horizontal gradient, vertical gradient and pixel value at the pixel point (x, y) in the image, respectively. The gradient magnitude G(x,y) at (x,y) and the gradient direction α(x,y) can be obtained according to the following formula:

(4) Divide the image into square cells with a side length of a pixels, a is the greatest common factor of m and n, create a gradient direction histogram for each cell, and divide the gradient direction 360 degrees into k directions block, the direction range of the ith direction block is Count the gradient direction of each pixel in the cell. If the gradient direction belongs to a certain direction block, the count value of the corresponding direction block is added to the amplitude corresponding to this gradient;

(5) Combine the cells into blocks, and the normalized gradient histogram in the block rewrites the gradient histogram corresponding to each cell into a vector form, and concatenates all the gradient vectors in each block to form the gradient of the block Direction histogram vector; multiply the vector by the corresponding normalization factor, the calculation formula of the normalization factor is:

Among them, v represents a vector that has not been normalized, ||v|| ₂ represents the second-order norm of v, and e represents a constant;

(6) Concatenate the normalized vectors of all blocks in the image to obtain the HOG feature of the training sample set

Steps 1-3, assign sample labels to all positive and negative training sample sets, and send the HOG features and sample labels of the training sample set to SVM for training:

(1) The training goal of SVM is to find an optimal hyperplane that can classify positive and negative samples, and its mathematical form can be expressed as:

Among them, w represents the vector perpendicular to the hyperplane, ||w|| represents the norm of w, ξ _i represents the slack variable, which is a non-negative number, D is a parameter used to control the weight of the two terms in the objective function, x _i represents the HOG feature of the ith sample, _yi represents the sample label of the ith sample, and b represents a constant;

(2) Build the Lagrangian function:

where α _i represents the Lagrange multiplier, r _i =D-α _i , then let

The objective function is transformed into

Among them, d* represents the optimal value of the objective function;

(3) Minimize L with respect to w, b, ξ, namely:

Putting Equation (11) into Equation (8), the objective function is transformed into:

Among them, < _xi ,x _j > means to find the inner product of x _i , x _j ;

(4) Use the SMO algorithm to find the optimal value of the Lagrangian multiplier α _i , and use the heuristic algorithm to select a pair of Lagrangian multipliers α _i , α _j ; fix other parameters except α _i , α _j , determine the value of α _i under the condition that w takes the extreme value, and use α _i to represent α _j ; keep repeating until the objective function converges;

(5) Determine the optimal hyperplane according to the optimal value of the Lagrange multiplier:

in, Represents the optimal value of the Lagrange multiplier, w ^* , b ^* represent the direction of the optimal hyperplane and the offset from the origin, respectively;

(6) Obtain the classification decision function, that is, the trained SVM classifier:

.

Step 2. The inspection robot reaches the designated inspection point through positioning and navigation, obtains the image of the on-site slider-type circuit breaker and reads it in the form of a grayscale image, which is used for the detection and identification of the slider-type circuit breaker;

Step 3. Perform coarse positioning and precise positioning on the target area to be detected, and filter the target candidate area to obtain the slider-type circuit breaker image, specifically:

Step 3-1. Use Merlin Fourier transform and phase correlation technology to roughly locate the target circuit breaker area in the image to be detected;

Step 3-2, using the machine learning method to accurately locate the target circuit breaker area, send the image to be detected into the trained classifier, and obtain several target candidate areas;

Step 3-3, respectively intersect each target candidate area with the coarse positioning target pointer circuit breaker area and compare the parameter IOU, and perform perceptual hashing between each target candidate area image and the pointer circuit breaker area image in the template image Calculate, obtain the perceptual hash index, calculate the mutual information index between each target candidate area image and the template image, and filter the target candidate area to obtain a pointer circuit breaker, specifically:

(1) Scale the target candidate area image and the template image to the same size, perform cosine transformation, select the low-frequency area in the upper left corner of the cosine-transformed image, remove the DC component of the coordinate (0,0) to obtain the feature vector, and calculate the target candidate area. The Hamming distance of the feature vector of the image and the template image, as a perceptual hash index;

(2) The specific calculation formula of the intersection ratio parameter IOU is:

In the formula, C is the rough positioning target circuit breaker area, and n _i is the target candidate area;

(3) The calculation formula of the mutual information index I(G _(X) ,H _(Y) ) is:

G _(X) and H _(Y) are the number of grayscale pixels in the template image and the candidate image, respectively, and W and H are the width and height of the candidate region image, respectively.

The confidence of each candidate area is calculated by weighting the intersection ratio IOU, mutual information and perceptual hash pHash of each candidate area, where D is a constant:

Confidence=1-(pHash+1/I(G _(X) ,H _(y) ))/(IOU+D)(20)

Sort all candidate regions in descending order of confidence, find the region with the highest confidence, and use this region as the candidate detection result. If the IOU of the alternative detection result is less than the set threshold thresholdIOU and (pHash+1/I(G _(X) ,H _(Y) )) is greater than the threshold thresholdA, the coarse positioning target circuit breaker area determined in step 3-2 As the final target, otherwise the candidate detection result is used as the final target. The threshold value thresholdIOU ranges from 0.1 to 0.4, and the threshold value thresholdA ranges from 10 to 50.

Step 4. Preprocess the acquired image of the on-site slider-type circuit breaker, extract the two areas with the largest connected area, and divide them into left and right parts according to their positions, specifically:

Step 4-1. Perform histogram equalization on the grayscale image to increase the overall contrast of the image and make the image clearer;

Step 4-2, perform Gaussian filtering on the image to eliminate Gaussian noise on the image;

Step 4-3, perform Otsu binarization on the image to distinguish the target image from the background image;

Step 4-4, perform an open operation on the image to smooth the boundary of the target image, eliminate small spikes, and disconnect narrow connections;

Steps 4-5, perform contour scanning on the image, and extract the area with the largest connected area.

Step 4-6: Detect the center of the circle, calculate the gradient of the graph, and determine the circle line. In the two-dimensional Hough space, give the gradient straight line of all graphs, and perform non-maximum suppression in the 4 neighborhoods, and set a threshold, The point whose accumulated sum is greater than this threshold in Hough space corresponds to the center of the circle;

Step 4-7. Detect the radius of the circle, calculate the distance from the center of a circle to all the circle lines, find the value with the same distance, and calculate the number of the same value. Only when the number of the same value is greater than a certain threshold is considered as the circle radius corresponding to the center of the circle , and detect its corresponding circle radius in the same way for another circle center.

Step 5. Perform pixel adjustment on the two separated areas respectively, slide on the image with a sliding window with a length of m pixels and a width of n pixels, extract the HOG feature from the window, and send the calculated HOG feature operator to the SVM multi-classification get the final recognition result.

Example 1

A method for identifying a slider-type circuit breaker based on an inspection robot, comprising the following steps:

Step 1. For each inspection point, select an image of the circuit breaker in the center of the inspection point as a template image, as shown in Figure 2, use the slider-type circuit breaker image data set collected in advance to train SVM multi-classification device;

Step 1-1. Prepare a training sample set, specifically:

(1) Collect 100,000 pictures containing the slider-type circuit breaker, take the pictures in the energy storage state as the positive sample set, and the pictures in the non-energy storage state as the negative sample set;

(2) Crop the picture, delete the redundant information outside the slider display window area on the slider-type circuit breaker

(3) Scale the picture into a rectangle whose length and width are both 48 pixels;

Step 1-2, extract the HOG features of positive and negative samples, specifically:

(1) Convert the color image to a grayscale image;

(2) Gamma correction is performed on the grayscale image to reduce local shadow and illumination changes in the image. Equation (1) is the Gamma correction formula, where gamma is 0.5;

(3) Calculate the horizontal and vertical gradients of each pixel of the image according to formula (2) (3), and then calculate the gradient magnitude and gradient direction at the pixel point (x, y) according to formula (4) (5);

(4) Divide the image into square cells with a side length of 8 pixels, and create a gradient direction histogram for each cell. Divide the gradient direction into 9 direction blocks 360 degrees, and count the gradient directions of each pixel in the cell. If the gradient direction belongs to a certain direction block, the count value of the corresponding direction block is added to the amplitude corresponding to this gradient. Combined into a block with a side length of 16 pixels, the gradient histogram is normalized within the block to reduce the influence of light, shadow and edge on the gradient; the HOG feature is obtained by concatenating the normalized vectors of all blocks in the image. ;

Steps 1-3, assign sample labels to all positive and negative samples, and send the HOG features and sample labels of positive and negative samples into SVM for training. The specific steps are:

(1) The training goal of SVM is to find an optimal hyperplane that can classify positive and negative samples, and its mathematical form can be expressed by equation (7);

(2) Construct the Lagrangian function, such as formula (8), and transform the objective function into formula (10) according to formula (9);

(3) Minimize L for w, b, ξ, and convert the objective function into formula (12);

(4) Use the SMO algorithm to find the optimal value of the Lagrange multiplier α _i ;

(5) Determine the optimal hyperplane according to the optimal value of the Lagrange multiplier and equation (13);

(6) Obtain the classification decision function formula (14), that is, the trained SVM classifier:

Step 2. The inspection robot reaches the designated inspection point through positioning and navigation, with a navigation error of 6cm, obtains an image of a slider-type circuit breaker and reads it in the form of a grayscale image for detection and identification;

Step 3. Perform coarse positioning and precise positioning on the target area to be detected, and filter the target candidate area to obtain a slider-type circuit breaker;

Step 3-1. Use Merlin Fourier transform and phase correlation technology to roughly locate the target circuit breaker area in the image to be detected;

Step 3-2, use the machine learning Adaboost classifier trained in advance to accurately locate the image to be detected, and obtain several target candidate regions;

Step 3-3, respectively intersect each target candidate area and the rough positioning target slider-type circuit breaker area and compare the parameter IOU; perceive each target candidate area image and the slider-type circuit breaker area image in the template image Hash calculation to obtain the perceptual hash index; calculate the mutual information index of each target candidate area image and the template image, and filter the target candidate area to obtain a slider circuit breaker;

(1) Calculate the perceptual hash pHash indicator. The candidate area image obtained by the classifier and the circuit breaker area image in the intercepted template image are calculated by perceptual hash pHash. The perceptual hash pHash calculation is to scale the two images to a size of 32*32, perform cosine transformation, select the 8*8 area in the upper left corner of the cosine-transformed image, and remove the DC component of the coordinate (0,0) to obtain 63 dimensions. Feature vector, calculate the Hamming distance of the feature vector of image A and image B, as the perceptual hash pHash indicator;

(2) Use formulas (16) to (19) to calculate the mutual information index;

(3) Calculate the intersection ratio parameter IOU using formula (15), and obtain three intersection ratio parameter indicators (0.7, 0.0, 0.0) respectively;

(4) The three indicators of the intersection ratio IOU, mutual information and perceptual hash pHash of each candidate area are weighted according to formula (20) to obtain the confidence level of the candidate area;

Sort all candidate regions in descending order of confidence, and find the region with the highest confidence as the final region.

Step 4. Preprocess the acquired image of the on-site slider-type circuit breaker. The processed image is shown in Figure 3. The two areas with the largest connected area are extracted and divided into left and right parts according to their positions, as shown in Figure 4. Show;

Step 5. Perform pixel adjustment on the two separated areas respectively, slide on the image with a sliding window with a length and width of 48 pixels, extract the HOG feature from the window, and send the extracted HOG feature into the SVM for judgment. Final test result.

Claims (9)

1. A method for identifying a slider type circuit breaker based on an inspection robot is characterized by comprising the following steps:

step 1, selecting a picture in the middle of the circuit breaker shot at each inspection point as a template picture for each inspection point, and training an SVM (support vector machine) multi-classifier by utilizing a slider type circuit breaker picture number set collected in advance;

step 2, the inspection robot reaches a specified inspection point through positioning and navigation, acquires an on-site slider type circuit breaker image and reads the image in a gray scale pattern form for detection and identification of the slider type circuit breaker;

step 3, carrying out coarse positioning and accurate positioning on a target area to be detected, and screening a target candidate area to obtain a slider type circuit breaker image;

step 4, preprocessing the acquired on-site slider type circuit breaker image, extracting two areas with the largest communication area, and dividing the two areas into a left part and a right part according to positions;

and 5, respectively carrying out pixel adjustment on the two separated regions, sliding a sliding window with the length of m pixels and the width of n pixels on the image, extracting HOG characteristics from the window, and sending the HOG characteristic operator obtained by calculation into the SVM multi-classifier to obtain a final recognition result.

2. The inspection robot-based slider type circuit breaker recognition method according to claim 1, wherein the specific method for training the SVM multiple classifiers in the step 1 is as follows:

step 1-1, collecting a slider type circuit breaker image number set as a positive and negative training sample set;

step 1-2, extracting HOG characteristics of a positive and negative training sample set;

and 1-3, endowing all positive and negative training sample sets with sample labels, and sending the HOG characteristics of the training sample sets and the sample labels into the SVM for training.

3. The inspection robot-based slider type circuit breaker recognition method according to claim 2, wherein the specific method for preparing the training sample set in the step 1-1 is as follows:

(1) collecting pictures of the slider type circuit breaker, and taking the pictures in an energy storage state as a positive sample set and the pictures in an energy non-storage state as a negative sample set;

(2) cutting pictures, and deleting redundant information outside a slider display window area on the slider type circuit breaker;

(3) the picture is scaled to m pixels long and n pixels wide, with m and n ranging from 36-64.

4. The inspection robot-based slider type circuit breaker identification method according to claim 2, wherein the specific method for extracting the positive and negative training sample set HOG features in the step 1-2 is as follows:

(1) converting the color image into a gray image;

(2) gamma correction is carried out on the gray level image, the local shadow and illumination change of the image are reduced, and the formula of the Gamma correction is as follows:

I(x,y)＝I(x,y)^gamma(1)

wherein I (x, y) represents the pixel value of the x row and the y column of the image, and gamma takes a number between 0 and 1;

(3) the gradient of each pixel of the image is calculated according to the following formula:

G_x(x,y)＝H(x+1,y)-H(x-1,y) (2)

G_y(x,y)＝H(x,y+1)-H(x,y-1) (3)

wherein G is_x(x,y)，G_y(x, y), H (x, y) respectively represents the horizontal gradient, the vertical gradient and the pixel value of the pixel point (x, y) in the image, and the gradient magnitude G (x, y) and the gradient direction α (x, y) of the pixel point (x, y) can be obtained according to the following formula:

(4) dividing an image into square cells with the side length of a pixel, wherein a is the maximum common factor of m and n, creating a gradient direction histogram for each cell, dividing the gradient direction into k direction blocks by 360 degrees, and the direction range of the ith direction block isCounting the gradient direction of each pixel in the cell, and if the gradient direction belongs to a certain direction block, adding the count value of the corresponding direction block to the amplitude value corresponding to the gradient;

(5) combining the unit cells into blocks, rewriting the gradient histogram corresponding to each unit cell into a vector form by the intra-block normalized gradient histogram, and connecting all gradient vectors in each block in series to form a gradient direction histogram vector of the block; multiplying the vector by a corresponding normalization factor, wherein the calculation formula of the normalization factor is as follows:

wherein v represents a vector that has not been normalized, | v | | | luminance₂A norm of order 2 representing v, e representing a constant;

(6) and connecting the normalized vectors of all the blocks in the image in series to obtain the HOG characteristic of the training sample set.

5. The inspection robot-based slider type circuit breaker recognition method according to claim 2, wherein the specific method of feeding the HOG features and sample labels of the positive and negative training sample sets into the SVM for training in the step 1-3 is as follows:

(1) the training target of the SVM is to find an optimal hyperplane which can realize classification of positive and negative samples, and the mathematical form of the optimal hyperplane can be expressed as follows:

where w represents a vector perpendicular to the hyperplane, | | w | | | represents the norm of w, ξ_iRepresenting a relaxation variable, being a non-negative number, D being a parameter controlling the weight of two terms in the objective function, x_iRepresenting HOG characteristics, y, of the ith sample_iA sample label representing the ith sample, b represents a constant;

(2) constructing a Lagrangian function:

wherein, α_iRepresenting the Lagrange multiplier, r_i＝D-α_iThen order

Transformation of objective function into

Wherein d is^*Representing an optimal value of the objective function;

(3) let L minimize for w, b, ξ, i.e.:

by bringing equation (11) into equation (8), the objective function is transformed into:

wherein,<x_i,x_j>expression to x_i,x_jInner product of (d);

(4) lagrange multiplier α using SMO algorithm_iUsing a heuristic algorithm to select a pair of lagrange multipliers α_i,α_jFixing device α_i,α_jDetermining α under the condition that w is extreme, among other parameters_iIs taken from α_iRepresentation α_j(ii) a Repeating the steps until the target function is converged;

(5) determining an optimal hyperplane according to the optimal value of the Lagrange multiplier:

wherein,representing the optimum value of the Lagrange multiplier, w^*,b^*Squares representing respectively optimal hyperplanesOffset to and from the origin;

(6) obtaining a classification decision function, namely a trained SVM classifier:

。

6. the inspection robot-based slider type circuit breaker identification method according to claim 1, wherein the specific steps of positioning and screening the target area in the step 3 are as follows:

step 3-1, roughly positioning a target circuit breaker region in a picture to be detected by utilizing Mellin Fourier transform and phase correlation technology;

3-2, accurately positioning the target circuit breaker region by using a machine learning method, and sending the image to be detected into a trained classifier to obtain a plurality of target candidate regions;

and 3-3, respectively solving a merging ratio parameter IOU of each target candidate area and the coarse positioning target slide block type circuit breaker area, performing perceptual hash calculation on each target candidate area image and the slide block type circuit breaker area image in the template image to obtain perceptual hash indexes, calculating mutual information indexes of each target candidate area image and the template image, and screening the target candidate areas to obtain the slide block type circuit breakers.

7. The inspection robot-based slider type circuit breaker identification method according to claim 6, wherein the step 3-3 senses a hash index, an intersection ratio parameter IOU and a mutual information index I (G)_(X),H_(Y)) The specific calculation methods are respectively as follows:

(1) scaling the target candidate area image and the template image to the same size, performing cosine transform, selecting a low-frequency area at the upper left corner of the image after cosine transform, removing direct current components of coordinates (0,0) to obtain a characteristic vector, and calculating the Hamming distance of the characteristic vector of the target candidate area image and the characteristic vector of the template image to be used as a perceptual hash index;

(2) the specific calculation formula of the intersection ratio parameter IOU is as follows:

wherein C is a coarse positioning target breaker area, n_iA target candidate area is obtained;

(3) mutual information index I (G)_(X),H_(Y)) The calculation formula of (2) is as follows:

G_(X)、H_(Y)the number of grayscale pixels of the template image and the candidate image, respectively, and W, H the width and height of the candidate area image, respectively.

8. The method for identifying the sliding block type circuit breaker of the power inspection robot according to claim 6, wherein the specific method for screening the target candidate area to obtain the sliding block type circuit breaker in the step 3-3 is as follows:

weighting three indexes of the intersection ratio IOU, mutual information and perceptual hash pHash of each candidate region to obtain the confidence coefficient of the candidate region, wherein D is a constant:

Confidence＝1-(pHash+1/I(G_(X),H_(y)))/(IOU+D) (20)

according to the sequence of the confidence degrees of all the candidate regions from large to small, the region with the maximum confidence degree is obtained and is used as the candidate detection result, and if the candidate detection resultsThe IOU of the fruit satisfies the condition that the IOU is less than the set threshold value and is (pHash +1/I (G)_(X),H_(Y)) When the circuit breaker area is larger than the threshold, the circuit breaker area with the rough positioning target determined in the step 3-2 is taken as a final target, otherwise, the alternative detection result is taken as the final target.

9. The inspection robot-based slider type circuit breaker identification method according to claim 1, wherein the image preprocessing in the step 4 comprises the following specific steps:

step 4-1, histogram equalization is carried out on the gray level image, and the overall contrast of the image is increased, so that the image is clearer;

4-2, carrying out Gaussian filtering on the image to eliminate Gaussian noise on the image;

4-3, carrying out Otsu binarization on the image to distinguish the target image from the background image;

4-4, performing opening operation on the image to smooth the boundary of the target image, eliminating tiny spikes and disconnecting narrow connection;

and 4-5, carrying out contour scanning on the image, and extracting two areas with the largest communication area.