A method and system for splicing infrared images of super-long string disc-shaped suspension porcelain insulators

Abstract

The invention discloses a method and system for splicing infrared images of super-long string disc-shaped suspension porcelain insulators. The method comprises the following steps: obtaining segmented infrared images of super-long string disc-shaped suspension porcelain insulators; preprocessing the obtained segmented infrared images; dividing the overlapping areas of the preprocessed segmented infrared images into a plurality of sub-block pairs; obtaining matching feature points of the sub-block pairs; calculating confidence coefficients by screening reliable matching feature points through an optimal sub-block search algorithm, and obtaining the sub-block pair with the highest confidence coefficient as the optimal sub-block pair by comparing the confidence coefficients of matching feature points between different sub-block pairs; and realizing alignment and splicing of the preprocessed segmented infrared images by transforming the preprocessed segmented infrared images corresponding to the optimal sub-block pairs. The invention adopts a designed optimal sub-block search algorithm to make the screened matching feature points more reliable, the spliced ​​images less distorted, and significantly reduce the mis-splicing of the insulator infrared images.

Description

Technical Field

[0001] The invention relates to the technical field of image processing, and in particular to a method and system for splicing infrared images of super-long string disc-shaped suspended porcelain insulators. Background Art

[0002] Extra-long string disc-shaped suspension porcelain insulators play a vital role in the power system. They withstand high voltage and isolate the live parts from the ground or other conductive parts. However, porcelain insulators are affected by various factors and environments during use, which leads to a decline in their performance and zero value degradation. Therefore, porcelain insulator failures can be effectively prevented through regular inspection and maintenance. At present, the zero value detection methods of porcelain insulators at home and abroad include spark gap method, insulation resistance method, distributed voltage method, power frequency withstand voltage method, leakage current method, ultraviolet corona detection method, ultrasonic detection method, etc., but the above methods are subject to greater restrictions in the actual operation and maintenance process due to safety risks, labor intensity, detection applicability, accuracy, efficiency and cost.

[0003] The infrared imaging method is a non-contact detection method that uses an infrared thermal imager to detect defects in porcelain insulators. It identifies possible anomalies or defects by capturing the thermal distribution image of the surface of the porcelain insulator. It has the advantages of non-contact detection, high safety, the ability to quickly cover a large area, and real-time imaging, which is easy to find potential problems. In actual operation and maintenance work, the temperature difference of plus or minus 0.5°C at the iron cap is generally used as the basis for infrared diagnosis of low and zero value insulators. Due to the long length of the super-long string disc-shaped suspension porcelain insulator, the existing industrial infrared camera has limited viewing angle and resolution, and it is impossible to capture clear infrared images of super-long string disc-shaped suspension magnetic insulators, which has caused certain difficulties in studying the infrared images of super-long string disc-shaped suspension magnetic insulators. Therefore, image stitching can be used to obtain the complete infrared image of the super-long string disc-shaped suspension porcelain insulator. However, due to the large number of periodic repetitions of the insulator shape features, the traditional image stitching algorithm cannot find the correct registration position. In addition, due to the change in the camera's viewing angle, the shape of the insulator at the edge of the insulator's infrared image will be perspectively deformed, resulting in errors in matching the feature points at the edge of the overlapping area of ​​the insulator's infrared image with the feature points in the middle of the overlapping area of ​​another image when using the deep learning method to perform feature matching. This will result in the problem of incorrect number of insulator pieces in the spliced ​​image. Summary of the invention

[0004] In view of the deficiencies in the prior art, the present invention provides a method and system for splicing infrared images of ultra-long strings of disc-shaped suspended porcelain insulators, which aims to solve the problem that it is currently impossible to use industrial infrared cameras to obtain complete infrared images of ultra-long strings of disc-shaped suspended magnetic insulators.

[0005] To achieve the above object, the present invention provides the following technical solution: a method for splicing infrared images of an ultra-long string of disc-shaped suspension porcelain insulators, comprising the following steps:

[0006] Step S1: using a drone equipped with an infrared camera to take segmented infrared images of the super-long string of disc-shaped suspension porcelain insulators; the segmented infrared images include a grounding side infrared image and a high-voltage side infrared image of the super-long string of disc-shaped suspension porcelain insulators;

[0007] Among them, there are overlapping areas between the infrared image of the grounding side of the super-long string of disc-shaped suspension porcelain insulators and the porcelain insulators in the infrared image of the high-voltage side;

[0008] Step S2: construct a deep learning semantic segmentation network to preprocess the acquired segmented infrared images;

[0009] Step S3: Calculate the ground sampling distance GSD of the infrared camera carried by the UAV, and then calculate the overlapping area width of the pre-processed segmented infrared image according to the ground sampling distance, the interval between the shooting time of the ground side infrared image and the high voltage side infrared image in the pre-processed segmented infrared image, and the horizontal flight speed of the UAV; set a mask on the overlapping area and divide it into blocks to obtain a number of sub-block pairs;

[0010] Step S4: extracting and matching feature points of the obtained sub-block pairs to obtain matching feature points of the sub-block pairs;

[0011] Step S5: Screen the matching feature points of the sub-block pairs by using the optimal sub-block search algorithm, calculate the confidence coefficients of the remaining reliable matching feature points in the sub-block pairs after screening, and compare the confidence coefficients of the reliable matching feature points between different sub-block pairs to obtain the sub-block pair with the highest confidence coefficient as the optimal sub-block pair;

[0012] Step S6: Aligning and splicing the pre-processed segmented infrared images by transforming the pre-processed segmented infrared images corresponding to the optimal sub-blocks.

[0013] Furthermore, the specific process of screening the matching feature points of the sub-block pairs is as follows: screening the matching feature points of the sub-block pairs, using the nearest neighbor check to find the two matching feature points with the closest Euclidean distance in the sub-block pairs, setting the distance of the closest matching feature point to d1, and the distance of the second closest matching feature point to d2, calculating the ratio ratio=d1 / d2, and then setting a threshold to determine whether the matching feature points are reliable. When the ratio is less than the threshold value 0.5, it is determined to be reliable, otherwise, it is determined to be unreliable.

[0014] Furthermore, the confidence coefficient of the remaining reliable matching feature points in the sub-block pair after removing the unreliable matching feature points is calculated and expressed as:

[0015] ;

[0016] ;

[0017] In the formula, is the mean Euclidean distance of the matching feature points in the sub-block pair; γ is the number of matching feature points in the sub-block pair; For sub-blocks, The coordinates of the matching feature points in the preprocessed segmented infrared image; For sub-blocks, The coordinates of the matching feature points in the preprocessed segmented infrared image; is the confidence coefficient of the matching feature points in the sub-block pair.

[0018] Furthermore, the specific process of setting a mask and dividing the overlapping area into blocks to obtain a number of sub-block pairs is as follows: using a specific matrix to mask the overlapping area of ​​the preprocessed segmented infrared image, setting the value of the overlapping area in the preprocessed segmented infrared image to 1, and the value of the non-overlapping area to 0, and then dividing the overlapping area of ​​the preprocessed segmented infrared image into The sub-blocks have the same length and width. , respectively and , It represents the first sub-block divided in the overlapping area of ​​the ground side infrared image. represents the nth sub-block divided in the overlapping area of ​​the ground side infrared image; It represents the first sub-block divided in the overlapping area of ​​the high-voltage side infrared image. Indicates the nth sub-block divided in the overlapping area of ​​the high-voltage side infrared image. and As a pair of sub-blocks.

[0019] Furthermore, the specific process of step S4 is: using the Efficient LoFTR algorithm to perform feature point matching on multiple sub-block pairs in the overlapping area of ​​the pre-processed segmented infrared image to obtain matching feature points of the sub-block pairs.

[0020] Furthermore, the specific process of step S6 is: using the USAC_MAGSAC robust estimation algorithm to calculate the optimal transformation matrix of the matching feature points in the optimal sub-block pair, and using the optimal transformation matrix to transform one image in the corresponding preprocessed segmented infrared image into the coordinate system of another image to complete the alignment and splicing of the preprocessed segmented infrared images.

[0021] Furthermore, the specific process of step S2 is as follows:

[0022] Step S2.1: construct an infrared image dataset using the acquired segmented infrared images, and use the labelme image annotation tool to annotate the porcelain insulator targets in the segmented infrared images in the infrared image dataset, and then divide the infrared image dataset into a training set, a test set, and a validation set in a ratio of 8:1:1;

[0023] Step S2.2: Construct a deep learning semantic segmentation network, use the training set, test set and validation set to train, test and validate the deep learning semantic segmentation network to obtain a trained deep learning semantic segmentation network, and then use the trained deep learning semantic segmentation network to remove the background of the segmented infrared image in the external image data set to obtain the preprocessed segmented infrared image.

[0024] Furthermore, the ground sampling distance of the infrared camera carried by the drone is calculated and expressed as:

[0025] ;

[0026] In the formula, is the flight altitude of the drone, is the width of the infrared camera sensor, is the focal length of the infrared camera, is the width of the segmented infrared image after preprocessing;

[0027] According to the ground sampling distance, the interval between the shooting time of the ground side infrared image and the high voltage side infrared image in the pre-processed segmented infrared image and the horizontal flight speed of the drone Calculate the overlap width of the preprocessed segmented infrared image , expressed as:

[0028] .

[0029] A splicing system for infrared images of super-long disc-shaped suspension porcelain insulators, comprising:

[0030] An image acquisition module is used to take segmented infrared images of the super-long string of disc-shaped suspension porcelain insulators using a drone equipped with an infrared camera; the segmented infrared images include grounding side infrared images and high-voltage side infrared images of the super-long string of disc-shaped suspension porcelain insulators;

[0031] The preprocessing module is used to build a deep learning semantic segmentation network and preprocess the acquired segmented infrared images;

[0032] The block division module is used to calculate the ground sampling distance GSD of the infrared camera carried by the UAV, and then calculate the overlapping area width of the pre-processed segmented infrared image according to the ground sampling distance, the interval between the shooting time of the ground side infrared image and the high voltage side infrared image in the pre-processed segmented infrared image, and the horizontal flight speed of the UAV; set a mask and divide the overlapping area into blocks to obtain a number of sub-block pairs;

[0033] An extraction and matching module is used to extract and match feature points of the obtained sub-block pairs to obtain matching feature points of the sub-block pairs;

[0034] A screening module is used to screen the matching feature points of the sub-block pairs by using an optimal sub-block search algorithm, calculate the confidence coefficients of the remaining reliable matching feature points in the sub-block pairs after screening, and compare the confidence coefficients of the reliable matching feature points between different sub-block pairs to obtain the sub-block pair with the highest confidence coefficient as the optimal sub-block pair;

[0035] The splicing module is used to realize the alignment and splicing of the pre-processed segmented infrared images by transforming the pre-processed segmented infrared images corresponding to the optimal sub-blocks.

[0036] An electronic device comprises a processor, a memory and a bus, wherein the processor and the memory are connected via the bus, wherein the memory is used to store a set of program codes, and the processor is used to call the program codes stored in the memory to execute a method for splicing infrared images of super-long string disc-shaped suspension porcelain insulators.

[0037] Compared with the existing technology, the present invention has the following beneficial effects:

[0038] (1) The present invention uses the Efficient LoFTR algorithm to extract feature points from the infrared image of the insulator, so that a large number of effective feature points can be extracted from the infrared image of the insulator with a large number of repeated patterns, and the efficiency and accuracy are higher.

[0039] (2) The present invention further divides the overlapping area into different sub-blocks on the basis of masking the overlapping area when extracting matching feature points, and designs a formula for calculating the confidence coefficients of matching points in different sub-blocks in the same image, so that the screened matching feature points are more reliable, the spliced ​​image is less distorted, the splicing is smoother, the quality and reliability are higher, and the mis-splicing of insulator infrared images is significantly reduced. The problem that when using the EfficientLoFTR algorithm to perform feature matching on the overlapping area, the feature points at the edge of the overlapping area of ​​the infrared image are mismatched with the feature points in the middle part of the overlapping area of ​​another image due to perspective deformation, resulting in the presence of an error in the number of insulator pieces in the spliced ​​image is solved.

[0040] (3) The present invention pre-processes the infrared image of the ultra-long string disc-shaped suspension porcelain insulator by adopting a deep learning semantic segmentation network, removes the background in the infrared image, and avoids the influence of background debris on the accuracy of image stitching. Compared with the existing technology, the number of insulator pieces in the infrared image of the insulator spliced ​​by the above operation is more accurate, the number of mis-splicing is less, and the reliability is higher. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 The figure is a flow chart of the method of the present invention.

[0042] Figure 2 It is a schematic diagram of the block structure of the overlapping area of ​​the segmented infrared image of the present invention. DETAILED DESCRIPTION

[0043] like Figure 1 As shown, the present invention provides a technical solution: a method for splicing infrared images of super-long string disc-shaped suspension porcelain insulators, comprising the following steps:

[0044] Step S1: using a drone equipped with an infrared camera to take segmented infrared images of an ultra-long string of disc-shaped suspension porcelain insulators; the segmented infrared images include a grounding side infrared image and a high-voltage side infrared image of the ultra-long string of disc-shaped suspension porcelain insulators.

[0045] Among them, there is an overlapping area between the infrared image of the grounding side of the super-long string of disc-shaped suspension porcelain insulators and the porcelain insulators in the infrared image of the high-voltage side.

[0046] In the process of using a drone equipped with an infrared camera to take segmented infrared images of ultra-long strings of disc-shaped suspended porcelain insulators, it is necessary to ensure that the porcelain insulators in the segmented infrared images are visually relatively horizontal; the drone only moves horizontally.

[0047] Step S2: Construct a deep learning semantic segmentation network to preprocess the acquired segmented infrared images.

[0048] Step S2.1: Use the acquired segmented infrared images to construct an infrared image dataset, and use the labelme image annotation tool to annotate the porcelain insulator targets in the segmented infrared images in the infrared image dataset, and then divide the infrared image dataset into a training set, a test set, and a validation set in a ratio of 8:1:1.

[0049] Step S2.2: Construct a deep learning semantic segmentation network, use the training set, test set and validation set to train, test and validate the deep learning semantic segmentation network to obtain a trained deep learning semantic segmentation network, and then use the trained deep learning semantic segmentation network to remove the background of the segmented infrared image in the external image data set to obtain the preprocessed segmented infrared image.

[0050] Among them, the deep learning semantic segmentation network in this embodiment adopts the U-net model, and can also adopt the SegNet model or the DeepLab v3 / v3+ model.

[0051] Step S3: Calculate the ground sampling distance (GSD) of the infrared camera carried by the UAV, and then calculate the overlapping area width of the pre-processed segmented infrared image according to the ground sampling distance, the interval between the shooting time of the ground side infrared image and the high-voltage side infrared image in the pre-processed segmented infrared image, and the horizontal flight speed of the UAV; set a mask on the overlapping area and divide it into blocks to obtain a number of sub-block pairs.

[0052] Among them, the ground sampling distance of the infrared camera carried by the drone is calculated and expressed as:

[0053] ;

[0054] In the formula, is the flight altitude of the drone, is the width of the infrared camera sensor, is the focal length of the infrared camera, is the width of the segmented infrared image after preprocessing.

[0055] Among them, according to the ground sampling distance, the interval between the shooting time of the ground side infrared image and the high voltage side infrared image in the pre-processed segmented infrared image and the horizontal flight speed of the drone Calculate the overlap width of the preprocessed segmented infrared image , expressed as:

[0056] .

[0057] like Figure 2 As shown in FIG. 1 , the specific process of setting a mask and dividing the overlapping area to obtain a number of sub-block pairs is as follows: using a specific matrix to mask the overlapping area of ​​the preprocessed segmented infrared image, setting the value of the overlapping area in the preprocessed segmented infrared image to 1, and setting the value of the non-overlapping area to 0, so that when extracting features, only the overlapping area of ​​the preprocessed segmented infrared image can be extracted; then the overlapping area of ​​the preprocessed segmented infrared image is evenly divided into The sub-blocks have the same length and width. , respectively and , It represents the first sub-block divided in the overlapping area of ​​the ground side infrared image. represents the nth sub-block divided in the overlapping area of ​​the ground side infrared image; It represents the first sub-block divided in the overlapping area of ​​the high-voltage side infrared image. Indicates the nth sub-block divided in the overlapping area of ​​the high-voltage side infrared image. and As a pair of sub-blocks.

[0058] The number of sub-blocks divided in the overlapping area is determined according to the number of porcelain insulator pieces in the overlapping area and the pre-processed segmented infrared image.

[0059] Step S4: extracting and matching feature points of the obtained sub-block pairs to obtain matching feature points of the sub-block pairs.

[0060] The specific process of step S4 is: using the Efficient LoFTR algorithm to perform feature point matching on multiple sub-block pairs in the overlapping area of ​​the pre-processed segmented infrared image to obtain matching feature points of the sub-block pairs.

[0061] Among them, the Efficient LoFTR algorithm detects and matches feature points of images based on the visual model (Transformer), introduces position encoding to help the visual model understand the relative positions of different elements in the segmented infrared image, and provides a clear position information for the visual model, so that it can learn the relative position in the segmented infrared image and process image features from a global perspective. This allows the Efficient LoFTR algorithm to not only rely on shape features in the image for feature extraction, but also extract a large number of feature points in weak texture areas and pattern repetition areas.

[0062] Position encoding is used in the visual model to provide position information for sub-block pairs so that the visual model can understand the relative positions of these sub-block pairs in the image and extract features of the sub-block pairs from a global perspective, thereby effectively capturing the potential structures and details in the sub-block pairs. This embodiment uses the sin function for position encoding, which is expressed as:

[0063] ;

[0064] In the formula, , They respectively represent the row position and column position of the feature point in the sub-block pair in the preprocessed segmented infrared image; Represents the dimension index, which determines the frequency of the sine or cosine function; is the dimension of the visual model (Transformer); Encode the positions of the matching feature points extracted from the sub-block pairs.

[0065] Step S5: Screen the matching feature points of the sub-block pairs through the optimal sub-block search algorithm, calculate the confidence coefficients of the remaining reliable matching feature points in the sub-block pairs after screening, and compare the confidence coefficients of the reliable matching feature points between different sub-block pairs to obtain the sub-block pair with the highest confidence coefficient as the optimal sub-block pair.

[0066] Among them, the specific process of screening the matching feature points of the sub-block pairs is as follows: screening the matching feature points of the sub-block pairs, using the nearest neighbor ratio check to find the two matching feature points with the closest Euclidean distance in the sub-block pair, setting the distance of the closest matching feature point to d1, and the distance of the second closest matching feature point to d2, calculating the ratio ratio=d1 / d2, and then setting a threshold of 0.5 to determine whether the matching feature point is reliable. If the ratio is less than the threshold of 0.5, the matching feature point is considered to be reliable, otherwise, the matching feature point is considered to be wrong, and the correctness of the matching feature point is verified by the nearest neighbor ratio check, thereby removing unreliable matching feature points in the sub-block pairs.

[0067] Among them, the confidence coefficient of the remaining reliable matching feature points in the sub-block pair after removing the unreliable matching feature points is calculated and expressed as:

[0068] ;

[0069] ;

[0070] In the formula, is the mean Euclidean distance of the matching feature points in the sub-block pair; γ is the number of matching feature points in the sub-block pair; For sub-blocks, The coordinates of the matching feature points in the preprocessed segmented infrared image; For sub-blocks, The coordinates of the matching feature points in the preprocessed segmented infrared image; is the confidence coefficient of the matching feature points in the sub-block pair.

[0071] To demonstrate the effectiveness and reliability of the optimal sub-block algorithm, as shown in Table 1, the feature point matching rate of the original sub-block pairs in the overlapping area is compared. It can be seen that the feature point matching rate within the sub-block pairs screened by the optimal sub-block algorithm is higher, and the overall efficiency of the acquired feature points is higher.

[0072] Table 1 Comparison of feature point matching rates of original sub-block pairs and filtered sub-block pairs in overlapping areas

[0073]

[0074] Step S6: Aligning and splicing the pre-processed segmented infrared images by transforming the pre-processed segmented infrared images corresponding to the optimal sub-blocks.

[0075] The specific process of step S6 is: using the USAC_MAGSAC robust estimation algorithm to calculate the optimal transformation matrix of the matching feature points in the optimal sub-block pair, and using the optimal transformation matrix to transform one image in the corresponding preprocessed segmented infrared image into the coordinate system of another image to complete the alignment and splicing of the preprocessed segmented infrared images.

[0076] In this example, an affine transformation is used to transform one image in the corresponding preprocessed segmented infrared image into the coordinate system of another image.

[0077] Among them, in order to make the stitched image quality higher and further eliminate the influence of perspective distortion caused by shooting angle change on image stitching, in the process of stitching the segmented infrared images after preprocessing, the partial area with the porcelain insulator image is selected to be retained and then matrix transformed; half of the overlapping area of ​​the segmented infrared image close to the edge is discarded, and the remaining part is retained and then stitched.

[0078] A splicing system for infrared images of super-long disc-shaped suspension porcelain insulators, comprising:

[0079] The image acquisition module is used to use a drone equipped with an infrared camera to take segmented infrared images of super-long string disc-shaped suspension porcelain insulators; the segmented infrared images include grounding side infrared images and high-voltage side infrared images of super-long string disc-shaped suspension porcelain insulators.

[0080] The preprocessing module is used to build a deep learning semantic segmentation network and preprocess the acquired segmented infrared images.

[0081] The blocking module is used to calculate the ground sampling distance GSD of the infrared camera carried by the UAV, and then calculate the overlapping area width of the pre-processed segmented infrared image according to the ground sampling distance, the interval between the shooting time of the ground side infrared image and the high-voltage side infrared image in the pre-processed segmented infrared image, and the horizontal flight speed of the UAV; set a mask and block the overlapping area to obtain several sub-block pairs.

[0082] The extraction and matching module is used to extract and match feature points of the obtained sub-block pairs to obtain matching feature points of the sub-block pairs.

[0083] The screening module is used to screen the matching feature points of the sub-block pairs through the optimal sub-block search algorithm, calculate the confidence coefficients of the remaining reliable matching feature points in the sub-block pairs after screening, and compare the confidence coefficients of the reliable matching feature points between different sub-block pairs to obtain the sub-block pair with the highest confidence coefficient as the optimal sub-block pair.

[0084] The splicing module is used to realize the alignment and splicing of the pre-processed segmented infrared images by transforming the pre-processed segmented infrared images corresponding to the optimal sub-blocks.

[0085] An electronic device comprises a processor, a memory and a bus, wherein the processor and the memory are connected via the bus, wherein the memory is used to store a set of program codes, and the processor is used to call the program codes stored in the memory to execute a method for splicing infrared images of super-long string disc-shaped suspension porcelain insulators.

[0086] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for splicing infrared images of super-long series of disc-shaped suspension porcelain insulators, characterized in that: The steps include: Step S1: using a drone equipped with an infrared camera to take segmented infrared images of the super-long string of disc-shaped suspension porcelain insulators; the segmented infrared images include a grounding side infrared image and a high-voltage side infrared image of the super-long string of disc-shaped suspension porcelain insulators; Among them, there are overlapping areas between the infrared image of the grounding side of the super-long string of disc-shaped suspension porcelain insulators and the porcelain insulators in the infrared image of the high-voltage side; Step S2: construct a deep learning semantic segmentation network to preprocess the acquired segmented infrared images; Step S3: Calculate the ground sampling distance GSD of the infrared camera carried by the UAV, and then calculate the overlapping area width of the pre-processed segmented infrared image according to the ground sampling distance, the interval between the shooting time of the ground side infrared image and the high voltage side infrared image in the pre-processed segmented infrared image, and the horizontal flight speed of the UAV; set a mask on the overlapping area and divide it into blocks to obtain a number of sub-block pairs; Step S4: extracting and matching feature points of the obtained sub-block pairs to obtain matching feature points of the sub-block pairs; Step S5: Screen the matching feature points of the sub-block pairs by using the optimal sub-block search algorithm, calculate the confidence coefficients of the remaining reliable matching feature points in the sub-block pairs after screening, and compare the confidence coefficients of the reliable matching feature points between different sub-block pairs to obtain the sub-block pair with the highest confidence coefficient as the optimal sub-block pair; Step S6: Aligning and splicing the pre-processed segmented infrared images by transforming the pre-processed segmented infrared images corresponding to the optimal sub-blocks.

2. The method for splicing infrared images of a super-long string of disc-shaped suspension porcelain insulators according to claim 1 is characterized in that: The specific process of screening the matching feature points of the sub-block pairs is as follows: screen the matching feature points of the sub-block pairs, use the nearest neighbor check to find the two matching feature points with the closest Euclidean distance in the sub-block pairs, set the distance of the closest matching feature point to d1, and the distance of the second closest matching feature point to d2, calculate the ratio ratio = d1 / d2, and then set a threshold to determine whether the matching feature points are reliable. When the ratio is less than the threshold value 0.5, it is determined to be reliable, otherwise, it is determined to be unreliable.

3. The method for splicing infrared images of a super-long string of disc-shaped suspension porcelain insulators according to claim 2 is characterized in that: After removing the unreliable matching feature points, the confidence coefficient of the remaining reliable matching feature points in the sub-block pair is calculated and expressed as: ； ； In the formula, is the mean Euclidean distance of matching feature points in the sub-block pairs; γ is the number of matching feature points in the sub-block pair; For sub-blocks, The coordinates of the matching feature points in the preprocessed segmented infrared image; For sub-blocks, The coordinates of the matching feature points in the preprocessed segmented infrared image; is the confidence coefficient of the matching feature points in the sub-block pair.

4. The method for splicing infrared images of an ultra-long string of disc-shaped suspension porcelain insulators according to claim 3 is characterized in that: The specific process of setting a mask and dividing the overlapping area to obtain several sub-block pairs is as follows: using a specific matrix to mask the overlapping area of ​​the preprocessed segmented infrared image, setting the value of the overlapping area in the preprocessed segmented infrared image to 1, and the value of the non-overlapping area to 0, and then dividing the overlapping area of ​​the preprocessed segmented infrared image into The sub-blocks have the same length and width. , respectively and , It represents the first sub-block divided in the overlapping area of ​​the ground side infrared image. represents the nth sub-block divided in the overlapping area of ​​the ground side infrared image; It represents the first sub-block divided in the overlapping area of ​​the high-voltage side infrared image. Indicates the nth sub-block divided in the overlapping area of ​​the high-voltage side infrared image. and As a pair of sub-blocks.

5. The method for splicing infrared images of a super-long string of disc-shaped suspension porcelain insulators according to claim 4 is characterized in that: The specific process of step S4 is: using the Efficient LoFTR algorithm to perform feature point matching on multiple sub-block pairs in the overlapping area of ​​the pre-processed segmented infrared image to obtain matching feature points of the sub-block pairs.

6. The method for splicing infrared images of an ultra-long string of disc-shaped suspension porcelain insulators according to claim 5 is characterized in that: The specific process of step S6 is: using the USAC_MAGSAC robust estimation algorithm to calculate the optimal transformation matrix of the matching feature points in the optimal sub-block pair, and using the optimal transformation matrix to transform one image in the corresponding preprocessed segmented infrared image into the coordinate system of another image to complete the alignment and splicing of the preprocessed segmented infrared images.

7. The method for splicing infrared images of an ultra-long string of disc-shaped suspension porcelain insulators according to claim 1 is characterized in that: The specific process of step S2 is: Step S2.1: construct an infrared image dataset using the acquired segmented infrared images, and use the labelme image annotation tool to annotate the porcelain insulator targets in the segmented infrared images in the infrared image dataset, and then divide the infrared image dataset into a training set, a test set, and a validation set in a ratio of 8:1:1; Step S2.2: Construct a deep learning semantic segmentation network, use the training set, test set and validation set to train, test and validate the deep learning semantic segmentation network to obtain a trained deep learning semantic segmentation network, and then use the trained deep learning semantic segmentation network to remove the background of the segmented infrared image in the external image data set to obtain the preprocessed segmented infrared image.

8. The method for splicing infrared images of an ultra-long string of disc-shaped suspension porcelain insulators according to claim 1 is characterized in that: Calculate the ground sampling distance of the infrared camera carried by the drone, expressed as: ； In the formula, is the flight altitude of the drone, is the width of the infrared camera sensor, is the focal length of the infrared camera, is the width of the segmented infrared image after preprocessing; According to the ground sampling distance, the interval between the shooting time of the ground side infrared image and the high voltage side infrared image in the pre-processed segmented infrared image and the horizontal flight speed of the drone Calculate the overlap width of the preprocessed segmented infrared image , expressed as: 。 9. A splicing system for infrared images of super-long series of disc-shaped suspended porcelain insulators, characterized in that: include: An image acquisition module is used to take segmented infrared images of the super-long string of disc-shaped suspension porcelain insulators using a drone equipped with an infrared camera; the segmented infrared images include grounding side infrared images and high-voltage side infrared images of the super-long string of disc-shaped suspension porcelain insulators; The preprocessing module is used to build a deep learning semantic segmentation network and preprocess the acquired segmented infrared images; The block division module is used to calculate the ground sampling distance GSD of the infrared camera carried by the UAV, and then calculate the overlapping area width of the pre-processed segmented infrared image according to the ground sampling distance, the interval between the shooting time of the ground side infrared image and the high voltage side infrared image in the pre-processed segmented infrared image, and the horizontal flight speed of the UAV; set a mask and divide the overlapping area into blocks to obtain a number of sub-block pairs; An extraction and matching module is used to extract and match feature points of the obtained sub-block pairs to obtain matching feature points of the sub-block pairs; A screening module is used to screen the matching feature points of the sub-block pairs by using an optimal sub-block search algorithm, calculate the confidence coefficients of the remaining reliable matching feature points in the sub-block pairs after screening, and compare the confidence coefficients of the reliable matching feature points between different sub-block pairs to obtain the sub-block pair with the highest confidence coefficient as the optimal sub-block pair; The splicing module is used to realize the alignment and splicing of the pre-processed segmented infrared images by transforming the pre-processed segmented infrared images corresponding to the optimal sub-blocks.

10. An electronic device, characterized in that: It includes a processor, a memory and a bus, wherein the processor and the memory are connected via the bus, wherein the memory is used to store a set of program codes, and the processor is used to call the program codes stored in the memory to execute the method for splicing infrared images of super-long string disc-shaped suspension porcelain insulators as described in any one of claims 1-8.

Images