Power grid inspection image automatic processing method based on image matching
By combining image preprocessing and feature extraction with a fault model trained by machine learning, the problem of UAV image recognition was solved, and high-precision, efficient, and automatic recognition of power grid inspection images was achieved.
Patent Information
- Application Number
- CN202510992152.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-18
- Publication Date
- 2025-11-04
AI Technical Summary
In existing technologies, the power grid inspection images acquired by inspection drones are affected by factors such as air pollution and lighting, making image recognition difficult and unable to accurately identify inspection targets, thus affecting defect identification and judgment.
An automatic image processing method for power grid inspection based on image matching is adopted, including image preprocessing, feature extraction and matching detection. Through Kalman filtering and median filtering for noise reduction, image dehazing and enhancement, Sobel edge detection and machine learning training of fault models, environmental interference is eliminated and accurate identification is achieved.
It improves image quality and recognition accuracy, enabling precise matching of the fault type and severity of inspection targets under environmental interference, automatically expanding the fault database, and improving recognition efficiency.
Smart Images
Figure CN120894299A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of power grid inspection, more particularly to a power grid inspection image automatic processing method based on image matching. BACKGROUND
[0002] With the development of economy and society, the role of power grid is becoming more and more important, and with the development of technology, power grid erection has been able to cross mountains and rivers, and it is very difficult to manually inspect and troubleshoot faults and hidden dangers between power transmission nodes on the mountain and across the water. Therefore, the way of unmanned aerial vehicle shooting image for inspection has been born.
[0003] The traditional processing and recognition method of the inspection image obtained by the inspection unmanned aerial vehicle adopts a simple convolutional neural network algorithm to recognize and process the picture. However, when the unmanned aerial vehicle obtains the image, it will be affected by many factors such as air haze, light and visibility, which will greatly affect the recognition difficulty of the image, resulting in that the obtained picture cannot accurately recognize the inspection target, thereby affecting the defect recognition and judgment of the inspection target.
[0004] In view of this, we propose a power grid inspection image automatic processing method based on image matching. SUMMARY
[0005] 1. Technical problem to be solved
[0006] The purpose of the present application is to provide a power grid inspection image automatic processing method based on image matching, which solves the technical problem that the traditional processing and recognition method of the inspection image obtained by the inspection unmanned aerial vehicle adopts a simple convolutional neural network algorithm to recognize and process the picture, and when the unmanned aerial vehicle obtains the image, it will be affected by many factors such as air haze, light and visibility, which will greatly affect the recognition difficulty of the image, resulting in that the obtained picture cannot accurately recognize the inspection target, thereby affecting the defect recognition and judgment of the inspection target. The technical effect is achieved.
[0007] 2. Technical scheme
[0008] The technical scheme of the present application provides a power grid inspection image automatic processing method based on image matching, which comprises the following steps:
[0009] S1, image preprocessing: obtaining a to-be-detected image corresponding to the power grid line at the current time, and performing image noise reduction and image enhancement preprocessing on the power grid line image;
[0010] S2, feature extraction: performing feature extraction on the preprocessed image to obtain a feature map;
[0011] S3, matching detection: comparing the processed power grid image feature map with a fault model, judging and outputting the fault type;
[0012] S4, report output: the information after image matching detection generates an inspection report through the automatic analysis system, facilitating subsequent maintenance and inspection.
[0013] Preferably, in step S1, the image denoising is first processed by Kalman filtering, and then the denoised image is denoised again by a median filter denoising method, to obtain a denoised image.
[0014] Preferably, the median filter denoising includes the following steps:
[0015] S11, set h ij is the gray value of point (i,j), A ij is the current working window, h min is the minimum gray value in S i,j , h max is the maximum gray value in S i,j , hmed is the median gray value in S i,j , set A max is the preset maximum allowed window;
[0016] S12, if h min < h med < h max , go to step S13, otherwise go to step S14;
[0017] S13, if h min < h ij < h max , output h ij , otherwise output h med ;
[0018] S14, increase the window size A ij , if A ij < A max , go to step S12, otherwise output h ij .
[0019] Preferably, in step S1, the image enhancement processing enhances the contrast, brightness and details of the image after image dehazing, so that the power grid facilities are more obvious, facilitating subsequent matching and analysis. The image dehazing is processed by the following formula:
[0020] J(x) = (I(x)-A) / t(x)+A,
[0021] Wherein, I(x) is the observed power grid image with haze; J(x) is the haze-removed clear power grid image without haze; t(x) is a transmission map, representing the light transmittance of a point in the image to the observer, usually taking a value between 0 and 1, and the estimated value is obtained by the local contrast of the image; A is atmospheric light, which is usually the area with higher brightness in the image, and is estimated by observing the sky or bright area in the distance of the power grid image.
[0022] Preferably, in step S2, the feature extraction adopts the Sobel edge detection method, and four different scale factors such as 1, 1.5, 2 and 2.5 are selected to perform multi-scale gradient calculation on each image. In each scale, the Sobel operator is used to calculate the gradient in the horizontal and vertical directions of the image to obtain the gradient amplitude and gradient direction of each pixel point. For the gradient direction of each pixel point, it is quantized to the corresponding direction index according to the direction interval to which it belongs. In each scale, the sum of the gradient amplitudes of the pixel points in each direction interval is calculated to obtain the weighted gradient amplitude of multiple directions. The weighted gradient amplitudes of multiple directions are normalized to obtain the weighted gradient direction histogram of each image block at the current scale. Each bin of the histogram represents the proportion of the weighted gradient amplitude of the corresponding direction interval. The weighted gradient direction histograms of each image at different scales are concatenated to obtain a feature vector.
[0023] Preferably, each element in the feature vector represents the proportion of the weighted gradient amplitude of the image in the corresponding scale and direction, reflecting the gradient distribution characteristics of the image in different scales and directions. The multi-scale direction histograms of all images are combined into a feature vector. The preprocessed power grid image data is divided into M images, and each image corresponds to a feature vector. The feature vectors of the M images are sequentially spliced in the order from left to right and from top to bottom to obtain a final feature vector. The final feature vector represents the multi-scale gradient direction distribution characteristics of the power grid equipment in the entire preprocessed power grid image data, which is used as the input feature for subsequent matching detection.
[0024] Preferably, the fault model is obtained by training through a machine learning method, specifically as follows:
[0025] S31, first, the images of historical fault defect types are obtained, all historical defect images are sorted into a data set, and then the data set is divided into a training set and a validation set. Then, an identification model is established. When the model is established, the initial fault defect model is obtained by combining the historical judgment results of the corresponding images after obtaining the historical fault defect type images. The constant value of the initial fault defect model is unknown. The constant value is obtained by training a plurality of historical fault defect type images and corresponding identification results. Then, the constant value is imported into the initial model, and the fault defect type image and the identification result are used as unknown values to obtain the fault defect type image.
[0026] S32, training and verification are performed using the training set and the verification set, during training, the pictures and corresponding recognition results of the training set are imported into the iterative optimization of constant value, the model after iterative optimization is verified using the verification set, during verification, the pictures of the verification set are input, and the model outputs the recognition result, finally, the recognition result is compared with the historical recognition result of the corresponding picture of the verification set to determine the optimization effect, that is, the fault defect model is obtained;
[0027] S33, after obtaining the reliable fault defect model, the inspection power grid image features obtained in step S2 are input into the fault defect model to automatically identify and output the fault defect type and judge the fault defect level.
[0028] Preferably, the fault defect model is divided into a plurality of sub-models, which are different according to the types of detection objects, including device aging, wire fracture, device pollution and wire overheating redness.
[0029] Preferably, when the detected fault type is not within the range defined by the model data set, the fault defect is manually confirmed by artificial manual confirmation, and a new fault type data set is created to include the fault defect image in the data set to form a new training set, and the fault defect model is trained again for subsequent matching of the same defect.
[0030] Preferably, when the detected fault type is not within the range defined by the model data set, the fault defect is manually confirmed by artificial manual confirmation, and a new fault type data set is created to include the fault defect image in the data set to form a new training set, and the fault defect model is trained again for subsequent matching of the same defect.
[0031] 3. Beneficial effects
[0032] One or more technical solutions provided in the technical scheme of the present application have at least the following technical effects or advantages:
[0033] 1. The present application processes the collected power grid inspection image by Kalman filtering method to reduce noise of the obtained image, then performs denoising processing on the denoised image by median filtering method, and then enhances the contrast, brightness and details of the image after image defogging, so that the power grid facilities are more obvious, the image quality is improved, and more details of the image are better reserved, facilitating subsequent identification and analysis;
[0034] 2. The multi-scale gradient calculation is performed by the Sobel edge detection method, the gradient direction is weighted and quantized, and the multi-scale direction histogram is obtained as the device feature, which can effectively extract the power grid related device features in the image, enhance the feature representation ability, and improve the image recognition accuracy;
[0035] 3. The processed power grid inspection image features and the fault defect model trained from historical fault images are trained to remove the influence of environmental and other interference factors, so that when matching the inspection image, the influence of interference factors can be eliminated, and the power grid inspection fault or not, fault type and degree can be accurately matched, achieving the purpose of automatic matching and identification, effectively eliminating the influence of environment and weather factors on the shooting picture, making the identification result more accurate and efficient, and at the same time, the fault picture database is automatically expanded and trained when a new fault type is identified, making the subsequent fault comparison more efficient, providing a high-precision and efficient matching method for inspection image processing. BRIEF DESCRIPTION OF DRAWINGS
[0036] Figure 1 The flowchart of the power grid inspection image automatic processing method based on image matching disclosed in a preferred embodiment of the present application is shown in the figure. DETAILED DESCRIPTION
[0037] The present application will be further described in detail below in conjunction with the accompanying drawings.
[0038] Reference Figure 1 The present application provides a power grid inspection image automatic processing method based on image matching, which comprises the following steps:
[0039] S1, image preprocessing: obtaining the image to be detected corresponding to the power grid line at the current time, and performing image noise reduction and image enhancement preprocessing on the power grid line image;
[0040] S2, feature extraction: extracting features from the preprocessed image to obtain a feature map;
[0041] S3, matching detection: comparing the processed power grid image feature map with the fault model to determine and output the fault type;
[0042] S4, report output: the information after image matching detection generates an inspection report through an automatic analysis system, which is convenient for subsequent maintenance and inspection.
[0043] After the power grid inspection image is denoised, the features are extracted, so as to effectively eliminate the influence of interference factors on the inspection picture, and then the inspection target defect type is identified by comparing the processed image with the fault defect model. The fault model obtained by processing the image and historical faults can eliminate the interference of interference factors, make the identification result more accurate and efficient, and provide a high-precision and efficient identification method for inspection image processing.
[0044] Further, in step S1, the image denoising is firstly processed by Kalman filtering method, and then the denoised image is processed again by median filtering denoising method to obtain the denoised image.
[0045] Further, the median filtering denoising comprises the following steps:
[0046] S11, set h ij is the gray scale of the point (i, j), A ij is the current working window, h min is the gray scale minimum value in S i,j , h max is the gray scale maximum value in S i,j , hmed is the gray scale median value in S i,j , set A max is the preset allowed maximum window.
[0047] S12, if h min < h med < h max , go to step S13, otherwise go to step S14;
[0048] S13, if h min < h ij < h max , output h ij , otherwise output h med .
[0049] S14, increase the window size A ij , if A ij < A max , go to step S12, otherwise output h ij .
[0050] By changing the window size of the filter dynamically in the filtering process according to the pre-set condition, the size of the filtering window is changed according to the pre-set condition, and whether the current pixel is noise is also judged according to certain conditions. If it is, the current pixel is replaced by the neighborhood median; if not, it is not changed. Through the above method, not only the salt and pepper noise with larger probability can be filtered out, but also the details of the image can be better protected, which is convenient for subsequent feature extraction and fault matching.
[0051] Further, in step S1, the image enhancement processing enhances the contrast, brightness and details of the image after image defogging, so that the power grid facilities are more obvious, which is convenient for subsequent matching and analysis. The image defogging is processed by the following formula:
[0052] J(x) = (I(x)-A) / t(x)+A,
[0053] Wherein, I(x) is the observed image of power grid with haze; J(x) is the clear image of power grid without haze after defogging; t(x) is a transmission map, representing the light transmittance from a point of the image to the observer, usually taking a value between 0 and 1, and the estimated value is obtained by the local contrast of the image; A is atmospheric light, which is usually the area with higher brightness in the image, and is estimated by observing the sky or bright area in the distance of the image of power grid.
[0054] By defogging and filtering the image, a clearer device scene image is obtained, the image quality is improved, more details of the image are better preserved, and subsequent image analysis is facilitated.
[0055] Further, in step S2, the feature extraction adopts a Sobel edge detection method, four different scale factors such as 1, 1.5, 2 and 2.5 are selected, multi-scale gradient calculation is performed on each image, at each scale, the Sobel operator is used to calculate the gradient in the horizontal and vertical directions of the image, the gradient amplitude and gradient direction of each pixel point are obtained, for the gradient direction of each pixel point, it is quantized to the corresponding direction index according to the direction interval to which it belongs, at each scale, the sum of the gradient amplitudes of the pixel points in each direction interval is calculated, the weighted gradient amplitudes of multiple directions are obtained, the weighted gradient amplitudes of multiple directions are normalized, and the weighted gradient direction histogram of each image block at the current scale is obtained, each bin of the histogram represents the proportion of the weighted gradient amplitude of the corresponding direction interval, and the weighted gradient direction histograms of each image at different scales are concatenated to obtain a feature vector.
[0056] Further, each element in the feature vector represents the proportion of the weighted gradient amplitude of the image in the corresponding scale and direction, reflecting the gradient distribution characteristics of the image in different scales and directions, the multi-scale direction histograms of all images are combined into a feature vector, the preprocessed power grid image data is divided into M images, each image corresponds to a feature vector, the feature vectors of the M images are sequentially spliced in the order from left to right and from top to bottom to obtain a final feature vector, which represents the multi-scale gradient direction distribution characteristics of the power grid equipment in the entire preprocessed power grid image data, and serves as an input feature for subsequent matching detection.
[0057] By performing multi-scale gradient calculation on the preprocessed image data and weighting and quantizing the gradient direction, a multi-scale direction histogram is obtained as the device feature, which can effectively extract the power grid related device features in the image, enhance the feature representation capability, and improve the accuracy of image recognition.
[0058] Further, the fault model is obtained by training through a machine learning method, specifically:
[0059] S31, first acquire the image of the historical fault defect type, arrange all historical defect images into a data set, then divide the data set into a training set and a validation set, and then establish an identification model. When the model is established, first acquire the historical fault defect type image, and then combine the corresponding picture historical judgment result to obtain an initial fault defect model. The constant value of the initial fault defect model is unknown. First, use multiple sets of historical fault defect type pictures and corresponding identification results to train and obtain the constant value. Then, import the constant value into the initial model and use the fault defect type picture as the unknown value to obtain the fault defect type picture;
[0060] S32, train and validate using the training set and the validation set. When training, import the pictures and corresponding identification results of the training set to perform iterative optimization of the constant value. The model after iterative optimization is validated using the validation set. When validating, input the pictures of the validation set and output the identification results from the model. Finally, compare the identification results with the historical identification results of the corresponding pictures of the validation set to determine the optimization effect, i.e., obtain the fault defect model.
[0061] S33, after obtaining a reliable fault defect model, input the image features of the power grid inspection obtained in step S2 into the fault defect model to automatically identify and output the fault defect type and judge the fault defect level.
[0062] Further, the fault defect model is divided into multiple sub-models, which are different according to the types of detection objects, including device aging, wire breakage, device pollution, and wire overheating redness.
[0063] Further, when the detected fault type is not within the range defined by the model data set, manually confirm the fault defect, create a new fault type data set, and include the fault defect image in the data set to form a new training set. Then, retrain the fault defect model to match the same defect in the future.
[0064] By using the processed power grid inspection image features and the fault defect model trained from the historical fault images, the influence of environmental interference factors is removed during training, so that the matching of the inspection image is not affected by the interference factors, thereby accurately matching the fault type and degree of the inspection power grid, achieving the purpose of automatic matching and identification, effectively eliminating the influence of environmental factors such as weather on the shooting pictures, making the identification result more accurate and efficient, and at the same time, the fault picture database is automatically expanded and trained when a new fault type is identified, making the subsequent fault comparison more efficient, providing a high-precision and efficient matching method for inspection image processing.
[0065] Further, when reporting output, automatically generate maintenance suggestions or alarm notifications according to the specific location of the fault image, the fault type, and the severity.
[0066] By matching the inspection image to identify fault information, output matching report is beneficial to rapid maintenance
[0067] In summary, by collecting the power grid inspection image, first through Kalman filtering method for image noise reduction processing, then through the median filter denoising method for denoising image again denoising processing, and then image defogging enhancement image contrast, brightness and details, make the power grid facilities more obvious, improve the image quality, better retain the image of more details, then through the Sobel edge detection method for multi-scale gradient calculation, the gradient direction is weighted quantization, get multi-scale direction histogram as equipment characteristics, can effectively extract the image of power grid related equipment characteristics, enhance the feature representation ability, improve the precision of image recognition, and then the processed power grid inspection image features, and the historical fault image training fault defect model, training time has removed the influence of environmental and other interference factors, so as to match the inspection image, can not be affected by interference factors, so as to accurately match the power grid inspection has no fault, and the fault type and degree, to achieve the purpose of automatic matching identification, effectively eliminate the influence of environmental weather and other factors on the shooting picture, make the recognition result more accurate, higher recognition efficiency, at the same time, the fault picture database in the identification of new fault type automatically expand the database training, so that the subsequent fault comparison efficiency is higher, for the inspection image processing provides high precision and efficient matching method.
Claims
1. A method for automatic processing of power grid inspection images based on image matching, characterized in that, Comprising the following steps: S1, image preprocessing: obtaining the image to be detected corresponding to the power grid line at the current time, and performing image noise reduction and image enhancement preprocessing on the power grid line image; S2, feature extraction: performing feature extraction on the preprocessed image to obtain a feature map; S3, matching detection: comparing the processed power grid image feature map with the fault model to judge and output the fault type; S4, report output: the information after image matching detection generates an inspection report through an automatic analysis system, facilitating subsequent maintenance and inspection.
2. The image matching based power grid inspection image automatic processing method according to claim 1, characterized in that: In step S1, the image noise reduction is first processed by the Kalman filter method, and then the denoised image is further denoised by the median filter denoising method to obtain a denoised image.
3. The image matching based power grid inspection image automatic processing method according to claim 2, characterized in that: The median filter denoising includes the following steps: S11, Let h ij Let A be the gray level of point (i,j). ij For the current working window, h min For S i,j The minimum gray level in h max For S i,j The maximum grayscale value in the image, hmed is S i,j The median gray level, let A be. max This is the preset maximum allowed number of doors and windows; S12, if h min <h med <h max go to step S13, otherwise go to step S14; S13, if h min <h ij <h max , output h ij , otherwise output h med ; S14, increase window A ij size, if A ij <A max then go to step S12, otherwise output h ij .
4. The image matching based power grid inspection image automatic processing method according to claim 1, characterized in that: In step S1, the image enhancement processing enhances the contrast, brightness and details of the image after image dehazing, making the power grid facilities more obvious, facilitating subsequent matching and analysis. The image dehazing is processed by the following formula: J(x) = (I(x) - A) / t(x) + A, Where I(x) is the observed power grid image with haze; J(x) is the clear power grid image after dehazing; t(x) is the transmission map, representing the light transmittance from a point in the image to the observer, usually taking a value between 0 and 1, and the estimated value is obtained by the local contrast of the image; A is the atmospheric light, which is usually the area with high brightness in the image, and is estimated by observing the sky or bright area in the distance of the power grid image.
5. The image matching based automatic processing method of power grid inspection images according to claim 1, characterized in that: In step S2, the feature extraction adopts the Sobel edge detection method, selects four different scale factors such as 1, 1.5, 2 and 2.5, performs multi-scale gradient calculation on each image, at each scale, uses the Sobel operator to calculate the gradient in the horizontal and vertical directions of the image, obtains the gradient amplitude and gradient direction of each pixel point, for the gradient direction of each pixel point, quantizes it to the corresponding direction index according to its belonging direction interval, at each scale, calculates the sum of the gradient amplitudes of the pixel points in each direction interval to obtain the weighted gradient amplitude of multiple directions, normalizes the weighted gradient amplitude of multiple directions to obtain the weighted gradient direction histogram of each image block at the current scale, each bin of the histogram represents the proportion of the weighted gradient amplitude of the corresponding direction interval, and the weighted gradient direction histograms of each image at different scales are concatenated to obtain a feature vector.
6. The image matching based power grid inspection image automatic processing method according to claim 5, characterized in that: Each element in the feature vector represents the weighted gradient amplitude proportion of the image in the corresponding scale and direction, reflecting the gradient distribution characteristics of the image in different scales and directions, the multi-scale direction histograms of all images are combined into a feature vector, the preprocessed power grid image data is divided into M images, each image corresponds to a feature vector, the feature vectors of the M images are sequentially spliced in the order from left to right and from top to bottom to obtain a final feature vector, which represents the multi-scale gradient direction distribution characteristics of the entire preprocessed power grid image data, serving as the input feature for subsequent matching detection.
7. The image matching based power grid inspection image automatic processing method according to claim 1, characterized in that: The fault model is obtained by training through a machine learning method, specifically: S31, first, the images of historical fault defect types are acquired, all historical defect images are sorted into a data set, the data set is then divided into a training set and a verification set, and then an identification model is established. When the model is established, first, the historical fault defect type images are acquired, and the initial fault defect model is obtained by combining the historical judgment results of the corresponding images. The constant value of the initial fault defect model is unknown. The constant value is obtained by training a plurality of historical fault defect type images and corresponding identification results, and then the constant value is imported into the initial model, and the fault defect type image and the identification result are used as unknown values to obtain the fault defect type image; S32, the training set and the verification set are used for training and verification. During training, the images and corresponding identification results of the training set are imported to iteratively optimize the constant value. The model after iterative optimization is verified using the verification set. During verification, the images of the verification set are used as input, and the model outputs the identification result. Finally, the identification result is compared with the historical identification result of the corresponding image of the verification set to determine the optimization effect, i.e., the fault defect model is obtained; S33, after obtaining the reliable fault defect model, the image features of the power grid obtained in step S2 are input into the fault defect model to automatically identify and output the fault defect type and judge the fault defect level.
8. The image matching based power grid inspection image automatic processing method according to claim 7, characterized in that: The fault defect model is divided into a plurality of sub-models, which are different according to the types of detection objects, including device aging, wire fracture, device pollution, and wire overheating and redness.
9. The image matching based power grid inspection image automatic processing method according to claim 7, characterized in that: When the detected fault type is not within the defined range of the model data set, the fault defect is manually confirmed by artificial, a new fault type data set is created, the fault defect image is included in the data set to form a new training set, and the fault defect model is retrained, so as to match the same defect in the future.
10. The image matching based automatic processing method of power grid inspection images according to claim 1, characterized in that: When the report is output, a maintenance suggestion or an alarm notification is automatically generated according to the specific position of the fault image, the fault type, and the severity.