High-voltage disconnector positioning and state recognition method based on SURF feature matching

Abstract

A kind of high-voltage disconnecting switch positioning and state identification method based on SURF feature matching, first through SURF feature matching method cooperate with the image similarity based on hamming distance measurement accurate positioning switch position, through perceptual hashing algorithm as similar determination measure accurately identify switch state.DCT is used to reduce frequency, higher than average hash algorithm and difference hash algorithm, to meet the requirements of high-voltage switch state identification in accuracy, and in the aspect of calculation speed, in the case of ensuring robustness and matching accuracy, SURF algorithm is faster than SIFT algorithm in speed, hamming distance calculation is faster than Euclidean distance calculation, and inherits the characteristics of good adaptability to the influence of actual environment such as illumination and angle, overcome the problems of existing high-voltage switch state identification method in sample size, calculation cost, accuracy, the positioning and identification of switch have higher accuracy.

Description

Technical Field

[0001] This invention relates to the field of image matching technology in image processing and pattern recognition, and more specifically to a method for locating and identifying the status of a high-voltage disconnector switch based on SURF feature matching. Background Technology

[0002] With the rapid development of my country's economy, the power grid is expanding and voltage levels are increasing, placing higher demands on the safety and efficiency of power grid construction and operation. To meet the needs of unmanned substations and safe production operations, and to improve the online monitoring function of primary power equipment in unmanned substations, as well as the real-time fault early warning function, and to solve problems such as the inspection of disconnect switches under unmanned conditions, video images captured by image-capturing equipment are used. At the substation level, machine vision technology is employed to achieve online identification of the operating status of equipment such as disconnect switches. This improves the reliability of the substation monitoring system and broadens its scope.

[0003] Literature review reveals that numerous methods for identifying the status of high-voltage switches have been proposed. However, these methods generally rely on large training datasets and machine learning algorithms, resulting in significant computational costs. Furthermore, the limited number of training datasets necessitates further verification of the accuracy and reliability of machine learning-based high-voltage switch status monitoring. While SIFT feature matching stands out among traditional identification algorithms and exhibits strong robustness, it still falls short in terms of speed and accuracy.

[0004] In summary, existing high-voltage switch status identification methods still fall far short of the requirements for practical applications in terms of sample size, computational cost, and accuracy, and urgently need improvement. Summary of the Invention

[0005] To overcome the shortcomings of existing high-voltage switch status identification methods in terms of sample size, computational cost, and accuracy, this invention proposes a high-voltage disconnector location and status identification method based on SURF feature matching that is low in computational cost and high in accuracy under low sample conditions.

[0006] The technical solution adopted by this invention to solve its technical problem is:

[0007] A method for locating and identifying the status of a high-voltage disconnector switch based on SURF feature matching, the method comprising the following steps:

[0008] 1) After installing the camera, initialize the isolating switch so that it is in the closed state and the fully open state respectively;

[0009] 2) Acquire images from the camera's field of view, capture the position of the isolating switch, and use the image as a matching template;

[0010] 3) When the camera is in normal operation, it captures and saves the captured images in real time. The saved images are then matched using the matching template in the closed state and the matching template in the fully open state, respectively.

[0011] 4) For templates and images with high matching degree, the perceptual hashing algorithm is used to calculate the Hamming distance;

[0012] 5) If the calculated Hamming distance is lower than the set threshold, the display switch will be in the state of the corresponding matching template;

[0013] 6) If the calculated Hamming distances are all higher than the set threshold, the switch will be displayed as half open.

[0014] 7) Once the test results are displayed, return to step 3) to begin a new round of testing.

[0015] Furthermore, the process of step 3) is as follows:

[0016] 3.1) Use the images obtained after steps 1) and 2) as matching templates to perform SURF feature matching respectively;

[0017] 3.2) Save the image obtained after step 3.1) and extract the matching part, and use it together with the corresponding matching template in subsequent calculations.

[0018] Furthermore, in step 3.1), the SURF feature matching algorithm process is as follows:

[0019] 3.1.1) Establish scale space, determine feature points, and extract feature points;

[0020] 3.1.2) Determine the direction of feature points and generate feature vectors, that is, assign a principal direction to each feature point to ensure the invariance of feature vectors;

[0021] 3.1.3) Feature point matching and its optimization.

[0022] In step 3.1.1), the process of establishing the scale space is as follows:

[0023] 3.1.1.1) Establish the scale space, that is, perform Gaussian filtering on any point p(x, y) in image I to obtain the Hessian matrix H(x, σ) with the corresponding scale value σ:

[0024]

[0025] Among them, L xx (x, σ), L xy (x, σ), L yy (x, σ) is the second-order partial derivative at point p.

[0026] The Gaussian function is defined as:

[0027]

[0028] The determinant of the Hessian matrix is ​​then expressed as:

[0029] Det(H) = L xx L yy -L xy L xy

[0030] To improve computational speed, the SURF algorithm uses a box filter instead of a Gaussian filter.

[0031] 3.1.1.2) The SURF algorithm starts with a 9×9 box filter and expands its size. The 9×9 box filter is the filter template of the Gaussian second-order differential function with σ=1.2 after discretization and cropping. In SURF, the image remains unchanged, and only the size of the Gaussian filter window is changed to obtain images of different scales, thus forming a scale space. The scale space is divided into 4 groups, each with 4 layers. First, the initial filter value is 9×9, corresponding to the Gaussian scale σ=1.2. The value of each layer in the first group increases by 6 compared to the previous layer. It can be deduced that the size of the first group is 9×9, 15×15, 21×21, 27×27. The size difference of each layer in the second group is 12. The value of the first layer in each group is equal to the value of the second layer in the previous group. That is to say, the filter size of the second group is 15×15, 27×27, 39×39, 51×51, and so on. Let D be the value after the image is convolved with the box filter template. xx D xy D yy ;

[0032] 3.1.1.3) Further solving yields the determinant of the Hessian matrix, which can be simplified as follows:

[0033] Det(H) = L xx L yy -L xy L xy =(A-Bω)C

[0034] Where A = D xx D yy B=D xy D xy ω is the weighting coefficient, expressed as:

[0035]

[0036] so:

[0037] Det(H) = D xx D yy -(ωD xy )

[0038] 3.1.1.4) After the above calculation, the matrix image of the original image can be obtained at a certain point in the spatial scale. The SURF algorithm compares each processed pixel value with the surrounding 26 points. If it is the maximum or minimum value, it is initially determined to be a feature point, and then points with values ​​less than a certain threshold are removed.

[0039] In step 3.1.2), the process of generating the feature vector is as follows:

[0040] 3.1.2.1) To achieve rotation invariance, in the neighborhood of a feature point, within a region with a radius of 6σ, where σ is the scale value of the feature point, the horizontal and vertical Haar features of each point set are statistically analyzed. Then, a fan-shaped region with an angle of π / 3 is used to scan the entire circular region once around the point of interest at certain intervals. Finally, the vector sum of the Haar wavelet responses of the image points within each angle region of the fan-shaped region is obtained, and the direction of the longest vector is selected as the main direction of the feature point.

[0041] 3.1.2.2) After the direction of the feature points is determined, the feature vector needs to be generated. The feature point description vector reflects the gradient change of the gray value around the feature point. If the information of a feature point is more prominent, the corresponding gradient change will be larger. The gray gradient value changes more near the edge of the image, while the gradient value changes less significantly in a smoother area. Therefore, in order to improve the accuracy of subsequent matching, those points with prominent information should be extracted when extracting feature points.

[0042] In step 3.1.3), the feature point matching process is as follows:

[0043] 3.1.3.1) In step 3.1.1.3), after obtaining the trace of the Hessian matrix and a series of feature point values, positive values ​​are grouped together and negative values ​​are grouped together. If two feature points are in the same group, the Euclidean distance between the two descriptive vectors is calculated. If the values ​​are equal, the two points are a pair of matching points. If they are not equal, the two points are not a pair of matching points, and the Euclidean distance is not calculated. This reduces the workload and improves the efficiency of matching.

[0044] The Euclidean distance between the two description vectors is

[0045]

[0046] 3.1.3..2) In an n-dimensional vector, X ikX is the k-th element of the vector describing the i-th feature point in the target image; jk The j-th feature point in the image to be matched describes the k-th element of the vector. For a point A in the image, calculate the Euclidean distance between A and feature points B and C in the image to be matched. The minimum distance and the second minimum distance are ab and ac, respectively. If ab ≤ ηac, η = 0.8. By searching in this way, A and B can be determined to be a pair of matching points.

[0047] In step 3.1.2.2), generating the SURF feature vector includes the following steps:

[0048] 3.1.2.2.1) Construct a quadrilateral region of length 20σ×20σ centered on the point of interest, and then rotate the horizontal x-axis of this region to the main direction of the point of interest, where σ is the scale of the feature point;

[0049] 3.1.2.2.2) Divide the square region established in the previous step into a 4×4 sub-region, and then calculate the wavelet effect within a 5×5 range in each sub-region. This feature includes the Haar wavelet's response to the horizontal and vertical directions and the absolute value of the response sum.

[0050] 3.1.2.2.3) Calculate the four-dimensional features of each sub-region and add them together. Accumulate the 16 sub-regions to obtain the final 64-dimensional feature vector descriptor.

[0051] In step 4), the perceptual hash algorithm process is as follows:

[0052] 4.1) The image size was uniformly scaled to 32*32, resulting in a total of 1024 pixels;

[0053] 4.2) Convert the image to grayscale to standardize the input for the next step, and convert all non-single-channel images to single-channel grayscale.

[0054] 4.3) Calculate the Discrete Cosine Transform (DCT), that is, calculate the 32×32 data matrix corresponding to the discrete cosine transform of the 32×32 data matrix. The DCT formula is:

[0055]

[0056] 4.4) Reduce the DCT, i.e., take the 8×8 sub-region in the upper left corner of the 32×32 data matrix obtained in step 9.3);

[0057] 4.5) Calculate the average value. That is, through step 9.4), we can obtain an 8×8 integer matrix G. Calculate the average value of all elements in this matrix, and assume that its value is a.

[0058] 4.6) Further reduce the DCT; if it is greater than the average value, record it as 1; otherwise, record it as 0.

[0059] 4.7) Obtain the image information, combine 64 information bits, and maintain consistency regardless of their order;

[0060] 4.8) After obtaining the phash value of the image, compare the Hamming distance between the phash values ​​of the two images.

[0061] In step 4), the Hamming distance calculation formula is:

[0062]

[0063] In the formula, i = 0, 1, ..., n-1, and x and y are both n-bit codes. Indicates XOR.

[0064] The technical concept of this invention is: a method for locating and identifying the status of high-voltage disconnecting switches based on SURF feature matching, specifically a method for locating and identifying the status of high-voltage disconnecting switches based on SURF feature matching under low sample conditions, low computational cost, and high accuracy. This method first accurately locates the switch position using SURF feature matching combined with an image similarity metric based on Hamming distance. Finally, it accurately identifies the switch status using a perceptual hash algorithm (pHash) as a similarity criterion. DCT (Discrete Cosine Transform) is used to reduce frequency, achieving higher accuracy than average hash (aHash) and difference hash (dHash) algorithms, thus meeting the accuracy requirements for high-voltage switch status identification. Furthermore, in terms of computational speed, while ensuring robustness and matching accuracy, the SURF algorithm is faster than the SIFT algorithm, and Hamming distance calculation is faster than Euclidean distance calculation. It also inherits the characteristic of good adaptability to the influence of actual environmental factors such as lighting and angle, overcoming the shortcomings of existing high-voltage switch status identification methods in terms of sample size, computational cost, and accuracy, and achieving high accuracy in switch location and identification.

[0065] The beneficial effects of this invention are manifested in that it improves upon existing high-voltage switch status identification methods in terms of sample size, computational cost, and accuracy. Current methods for identifying high-voltage switch status generally utilize large training datasets and machine learning algorithms, resulting in high computational costs. Furthermore, due to the relatively small number of training data samples, the accuracy and reliability of the machine learning-based high-voltage switch status monitoring methods require further verification. Even the commonly used SIFT feature matching algorithm in traditional algorithms suffers from limitations in speed and accuracy. Therefore, this invention adopts an improved approach to traditional image recognition algorithms, including the SURF feature matching algorithm and a hash algorithm. The SURF algorithm enhances the real-time identification speed based on SIFT, while the perceptual hash algorithm compensates for the accuracy deficiency of the SURF algorithm. This solves the problems of insufficient sample size, computational cost, and accuracy in high-voltage switch status identification methods, enabling real-time identification of high-voltage switches. Attached Figure Description

[0066] Figure 1 This is a flowchart of a high-voltage disconnector location and status identification method based on SURF feature matching.

[0067] Figure 2 This is a flowchart of feature matching based on the SURF algorithm.

[0068] Figure 3 This is a flowchart based on the perceptual hashing algorithm.

[0069] Figure 4 This is a schematic diagram of a box filter and a Gaussian filter.

[0070] Figure 5 This is a schematic diagram of the size distribution of the box filter in scale space.

[0071] Figure 6 This is a diagram of the feature point localization process.

[0072] Figure 7 This is a diagram illustrating the process of determining the main direction of feature points.

[0073] Figure 8 This is a schematic diagram of a Haar wavelet filter.

[0074] Figure 9 This is a diagram of the feature descriptor generation process. Detailed Implementation

[0075] The present invention will now be further described with reference to the accompanying drawings.

[0076] Reference Figures 1-9 A method for locating and identifying the status of a high-voltage disconnector based on SURF feature matching includes the following steps:

[0077] 1) After installing the camera, initialize the isolating switch so that it is in the closed state and the fully open state respectively;

[0078] 2) Acquire images from the camera's field of view, capture the position of the isolating switch, and use the image as a matching template;

[0079] 3) When the camera is in normal operation, it captures and saves the captured images in real time. The saved images are then matched using the matching template in the closed state and the matching template in the fully open state, respectively.

[0080] The process of step 3) is as follows:

[0081] 3.1) Using the images obtained after steps 1 and 2 as matching templates, perform SURF feature matching respectively, such as... Figure 2 As shown, SURF feature matching includes the following steps:

[0082] 3.1.1) Establish scale space, determine feature points, and extract feature points;

[0083] Establish the scale space, that is, perform Gaussian filtering on any point p(x, y) in image I to obtain the Hessian matrix H(x, σ) with the corresponding scale value σ:

[0084]

[0085] Where L xx (x, σ), L xy (x, σ), L yy (x, σ) is the second-order partial derivative at point p. The Gaussian function is defined as:

[0086]

[0087] The determinant of the Hessian matrix can then be expressed as:

[0088] Det(H) = L xx L yy -L xy L xy

[0089] To improve computational speed, the SURF algorithm uses a box filter instead of a Gaussian filter, such as... Figure 4 As shown.

[0090] The SURF algorithm begins with a 9×9 box filter and then expands its size. The 9×9 box filter is the filter template obtained by discretizing and cropping the second-order differential function of Gaussian when σ is 1.2. In SURF, we keep the image unchanged and only change the size of the Gaussian filter window to obtain images of different scales, thus forming a scale space, such as... Figure 5 As shown.

[0091] The scale space is divided into 4 groups, each with 4 layers. First, the initial filter value is 9×9, corresponding to a Gaussian scale σ = 1.2. In the first group, the value of each subsequent layer increases by 6 compared to the previous layer, so the sizes of the first group can be calculated as 9×9, 15×15, 21×21, and 27×27. In the second group, the size difference between each layer is 12, and the value of the first layer in each group is equal to the value of the second layer in the previous group. That is, the filter sizes in the second group are 15×15, 27×27, 39×39, and 51×51, and so on. Let D be the value after convolving the image with the bounding box filter template. xx D xy D yy ;

[0092] Further solving yields the determinant of the Hessian matrix, which can be simplified as follows:

[0093] Det(H) = L xx L yy -L xy L xy =(A-Bω)C

[0094] Where A = D xx D yy B=D xy D xy ω is the weighting coefficient, expressed as:

[0095]

[0096] so:

[0097] Det(H) = D xx D yy -(ωD xy )

[0098] After the above calculations, a matrix image of the original image can be obtained at a certain point in the spatial scale. The SURF algorithm compares each processed pixel value with 26 surrounding points; if it is the maximum or minimum value, it is initially identified as a feature point. Then, points with values ​​less than a certain threshold are removed, such as... Figure 6 As shown.

[0099] 3.1.2) Determine the direction of feature points and generate feature vectors, that is, assign a principal direction to each feature point to ensure the invariance of feature vectors;

[0100] To achieve rotation invariance, within a neighborhood of a feature point, within a region of radius 6σ (σ being the scale value of the feature point), the horizontal and vertical Haar features of each point set are statistically analyzed. Then, a sector-shaped region with an angle of π / 3 is scanned around the point of interest at regular intervals, completing one revolution of the entire circular region. Finally, the vector sum of the Haar wavelet responses of the image points within each angular region of this sector is obtained, and the direction of the longest vector is selected as the principal direction of the feature point. Figure 7 As shown.

[0101] Once the orientation of the feature points is determined, the final step is to generate feature vectors. Feature vectors describe the gradient changes in grayscale values ​​around the feature point. If a feature point is particularly prominent, its gradient change will be larger. Typically, grayscale gradients are larger near image edges, while they are less pronounced in smoother areas. Therefore, to improve the accuracy of subsequent matching, feature points should be extracted based on their prominent information.

[0102] The generation of SURF feature vectors involves the following three steps:

[0103] Construct a quadrilateral region of length 20σ×20σ centered on the point of interest, and then rotate the horizontal x-axis of this region to the main direction of the point of interest, where σ is the scale of the feature point;

[0104] The square region established in the previous step is divided into 4×4 sub-regions. Then, wavelet effects within a 5×5 range are calculated in each sub-region. This feature includes the Haar wavelet's response to the horizontal and vertical directions, as well as the absolute value of the response sum. The Haar wavelet filter is as follows: Figure 8 As shown;

[0105] The four-dimensional features of each sub-region are calculated and summed. The 16 sub-regions are then combined to obtain the final 64-dimensional feature vector descriptor, such as... Figure 9 As shown.

[0106] 3.1.3) Feature point matching and its optimization;

[0107] Step 3.1.1) has already obtained the trace of the Hessian matrix. After obtaining a series of feature point values, positive values ​​are grouped together, and negative values ​​are grouped together. If two feature points are in the same group, the Euclidean distance between these two descriptive vectors is calculated. If the values ​​are equal, then these two points are a pair of matching points. If they are not equal, then the two points are not a pair of matching points, and the Euclidean distance is not calculated. This reduces the workload and improves the efficiency of matching.

[0108] The Euclidean distance between the two description vectors is

[0109]

[0110] In an n-dimensional vector, X ik x is the k-th element of the vector describing the i-th feature point in the target image; jk The j-th feature point in the image to be matched is described by the k-th element of the vector. Specifically, for a point A in the image, the Euclidean distances between A and feature points B and C in the image to be matched are calculated. The minimum distance and the second minimum distance are ab and ac, respectively. If ab ≤ ηac (η = 0.8), and this process is repeated, then A and B can be determined to be a pair of matching points.

[0111] 3.2) Save the image obtained after the operation in step 2.1) and extract the matching part, and use it together with the corresponding matching template in subsequent calculations.

[0112] 4) For templates and images with high matching degrees, a perceptual hashing algorithm is used to calculate the Hamming distance, such as... Figure 3 As shown; the perceptual hashing algorithm includes the following steps:

[0113] 4.1) The image size was uniformly scaled to 32*32, resulting in a total of 1024 pixels;

[0114] 4.2) Convert the image to grayscale to standardize the input for the next step, and convert all non-single-channel images to single-channel grayscale.

[0115] 4.3) Calculate the Discrete Cosine Transform (DCT), that is, calculate the 32×32 data matrix corresponding to the discrete cosine transform of the 32×32 data matrix. The DCT formula is:

[0116]

[0117] 4.4) Reduce the DCT, i.e., take the 8×8 sub-region in the upper left corner of the 32×32 data matrix obtained in step 9.3);

[0118] 4.5) Calculate the average value. That is, through step 9.4), we can obtain an 8×8 integer matrix G. Calculate the average value of all elements in this matrix, and assume that its value is a.

[0119] 4.6) Further reduce the DCT; if it is greater than the average value, record it as 1; otherwise, record it as 0.

[0120] 4.7) Obtain the image information, combine 64 information bits, and keep the order consistent.

[0121] 4.8) After obtaining the phash value of the image, compare the Hamming distance between the phash values ​​of the two images. Generally, images with a Hamming distance of less than 10 are considered similar images.

[0122] The formula for calculating the Hamming distance is:

[0123]

[0124] In the formula, i = 0, 1, ..., n-1, and x and y are both n-bit codes. Indicates XOR.

[0125] 5) If the calculated Hamming distance is lower than the set threshold of 10, the display switch will be in the state of the corresponding matching template.

[0126] 6) If the calculated Hamming distances are all higher than the set threshold of 10, the switch will be displayed as half open.

[0127] 7) Once the test results are displayed, return to step 3) to begin a new round of testing.

[0128] The above specific embodiments are used to explain and illustrate the present invention, but not to limit the present invention. Any modifications and changes made to the present invention within the spirit and scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A method for locating and identifying the status of a high-voltage disconnector based on SURF feature matching, characterized in that, The method includes the following steps: 1) After installing the camera, initialize the isolating switch so that it is in the closed state and the fully open state respectively; 2) Acquire images from the camera's field of view, capture the position of the isolating switch, and use the image as a matching template; 3) When the camera is in normal operation, it captures and saves the captured images in real time. The saved images are then matched using SURF features with both the closed and fully open matching templates. 4) For templates and images with high matching degrees, a perceptual hashing algorithm is used to calculate the Hamming distance; the process of the perceptual hashing algorithm is as follows: 4.1) The image size was uniformly scaled to 32*32, resulting in a total of 1024 pixels; 4.2) Convert the image to grayscale to standardize the input for the next step, and convert all non-single-channel images to single-channel grayscale. 4.3) Calculate the Discrete Cosine Transform (DCT), that is, calculate the 32×32 data matrix corresponding to the discrete cosine transform of the 32×32 data matrix. The DCT formula is: ； 4.4) Reduce the DCT, i.e., take the 8×8 sub-region in the upper left corner of the 32×32 data matrix obtained in step 9.3); 4.5) Calculate the average value. That is, through step 9.4), we can obtain an 8×8 integer matrix G. Calculate the average value of all elements in this matrix, and assume that its value is a. 4.6) Further reduce the DCT; if it is greater than the average value, record it as 1; otherwise, record it as 0. 4.7) Obtain the image information, combine 64 information bits, and maintain consistency regardless of the order; 4.8) After obtaining the phash values ​​of the images, compare the Hamming distance between the phash values ​​of the two images; The formula for calculating Hamming distance is: ， In the formula, i = 0, 1, ..., n-1, x and y are both n-bit codes, and ⊕ represents XOR; 5) If the calculated Hamming distance is lower than the set threshold, the display switch will be in the state of the corresponding matching template; 6) If the calculated Hamming distances are all higher than the set threshold, the switch will be displayed as half open. 7) Once the test results are displayed, return to step 3) to begin a new round of testing.

2. The high-voltage disconnector location and status identification method based on SURF feature matching according to claim 1, characterized in that: The process of step 3) is as follows: 3.1) Use the images obtained after steps 1) and 2) as matching templates to perform SURF feature matching respectively; 3.2) Save the image obtained after the operation in step 3.1) and extract the matching part, and use it together with the corresponding matching template in subsequent calculations.

3. The high-voltage disconnector location and status identification method based on SURF feature matching according to claim 2, characterized in that: In step 3.1), the SURF feature matching algorithm proceeds as follows: 3.1.1) Establish scale space, determine feature points, and extract feature points; 3.1.2) Determine the direction of feature points and generate feature vectors, that is, assign a principal direction to each feature point to ensure the invariance of feature vectors; 3.1.3) Feature point matching and its optimization.

4. The high-voltage disconnector location and status identification method based on SURF feature matching according to claim 3, characterized in that: In step 3.1.1), the process of establishing the scale space is as follows: 3.1.1.1) Establish the scale space, that is, for any point in image I Gaussian filtering is performed to obtain the Hessian matrix with a corresponding scale value of σ. for: ， in, , , The second partial derivative at point p , , The Gaussian function is defined as: ， The determinant of the Hessian matrix is ​​then expressed as: ， To improve computational speed, the SURF algorithm uses a box filter instead of a Gaussian filter. 3.1.1.2) The SURF algorithm starts with a 9×9 box filter and expands its size. The 9×9 box filter is the filter template of the second-order differential function of Gaussian with σ = 1.2 after discretization and cropping. In SURF, the image remains unchanged, and only the size of the Gaussian filter window is changed to obtain images of different scales, thus forming a scale space. The scale space is divided into 4 groups, each with 4 layers. First, the initial filter value is 9×9, corresponding to the Gaussian scale σ = 1.

2. The value of each layer in the first group increases by 6 compared to the previous layer. It can be deduced that the size of the first group is 9×9, 15×15, 21×21, 27×27. The size difference of each layer in the second group is 12. The value of the first layer in each group is equal to the value of the second layer in the previous group. That is to say, the filter size of the second group is 15×15, 27×27, 39×39, 51×51, and so on. Let D be the value after the image is convolved with the box filter template. xx D xy D yy ; 3.1.1.3) Further solving yields the determinant of the Hessian matrix, which can be simplified as follows: ， in, ω is the weighting coefficient, expressed as: ， so: ； 3.1.1.4) After the above calculation, the matrix image of the original image can be obtained at a certain point in the spatial scale. The SURF algorithm compares each processed pixel value with the surrounding 26 points. If it is the maximum or minimum value, it is initially determined to be a feature point, and then points with values ​​less than a certain threshold are removed.

5. The high-voltage disconnector location and status identification method based on SURF feature matching according to claim 3, characterized in that: In step 3.1.2), the process of generating the feature vector is as follows: 3.1.2.1) To achieve rotation invariance, in the neighborhood of a feature point, within a region with a radius of 6σ, where σ is the scale value of the feature point, the horizontal and vertical Haar features of each point set are statistically analyzed. Then, a fan-shaped region with an angle of π / 3 is used to scan the entire circular region once around the point of interest at certain intervals. Finally, the vector sum of the Haar wavelet responses of the image points within each angle region of the fan-shaped region is obtained, and the direction of the longest vector is selected as the main direction of the feature point. 3.1.2.2) Once the direction of the feature point is determined, the feature vector needs to be generated. The feature point description vector reflects the gradient change of the gray value around the feature point.

6. The high-voltage disconnector location and status identification method based on SURF feature matching according to claim 4, characterized in that: In step 3.1.3), the feature point matching process is as follows: 3.1.3.1) In step 3.1.1.3), after obtaining the trace of the Hessian matrix and a series of feature point values, positive values ​​are grouped together and negative values ​​are grouped together. If two feature points are in the same group, the Euclidean distance between these two descriptive vectors is calculated. If the values ​​are equal, the two points are a pair of matching points. If they are not equal, the two points are not a pair of matching points, and the Euclidean distance is not calculated. This reduces the workload and improves the matching efficiency. The Euclidean distance between the two description vectors is ， 3.1.3..2) In an n-dimensional vector, X ik X is the k-th element of the vector describing the i-th feature point in the target image; jk The j-th feature point in the image to be matched describes the k-th element of the vector. For a point A in the image, calculate the Euclidean distance between A and feature points B and C in the image to be matched. The minimum distance and the second minimum distance are ab and ac, respectively. If ab ≤ ηac, η = 0.

8. By searching in this way, A and B can be determined to be a pair of matching points.

7. The high-voltage disconnector location and status identification method based on SURF feature matching according to claim 5, characterized in that: In step 3.1.2.2), the SURF feature vector is generated. Includes the following steps: 3.1.2.2.1) Construct a quadrilateral region of length 20σ×20σ centered on the point of interest, and then rotate the horizontal x-axis of this region to the main direction of the point of interest, where σ is the scale of the feature point; 3.1.2.2.2) Divide the square region established in the previous step into a 4×4 sub-region, and then calculate the wavelet effect within a 5×5 range in each sub-region. This feature includes the Haar wavelet's response to the horizontal and vertical directions and the absolute value of the response sum. 3.1.2.2.3) Calculate the four-dimensional features of each sub-region and add them together. Accumulate the 16 sub-regions to obtain the final 64-dimensional feature vector descriptor.

Images