Insulator anomaly detection method and device, medium and equipment

Abstract

The invention relates to the technical field of image processing, and particularly discloses an insulator anomaly detection method and device, a medium and equipment. The method comprises the following steps: carrying out image detection on a target image by utilizing a target insulator detection sub-model in a target detection model obtained by pre-training, and obtaining a target first area frame which corresponds to the target image and comprises an insulator object; performing image detection on the target image by using a target telegraph pole detection sub-model in a target detection model obtained by pre-training to obtain a target second area frame which corresponds to the target image and comprises a telegraph pole object; based on the target first area frame and the target second area frame, determining an angle difference value between an insulator object in the target first area frame and a telegraph pole object in the target second area frame; and based on the angle difference value and a predetermined angle difference threshold value, determining whether the insulator object in the target image is abnormal so as to obtain an abnormality detection result. According to the invention, the method can achieve the precise detection of the abnormality of the insulator.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to an insulator abnormality detection method, device, medium and equipment. Background Art

[0002] In the power distribution network, insulators are key components that are exposed to the natural environment for a long time and are prone to various defects, such as dirt accumulation, physical damage, installation tilt, and irregular binding. These defects may affect the safe and stable operation of the distribution line, so regular inspections and tests are essential to promptly detect and solve potential problems. Traditional inspection methods such as manual inspections or vehicle inspections have problems such as low efficiency, high cost, poor safety, and low accuracy.

[0003] With the development of computer vision technology, more and more power companies have begun to use automated inspection methods to detect the status of insulators. This method uses drones to photograph the operation of distribution stations and uses image recognition technology to automatically detect insulator defects. In particular, the tilt defect of insulators is one of the most common and critical problems that affect safe operation.

[0004] However, existing image-based automatic detection technologies usually perform tilt detection based on insulator images at fixed angles, while the angles of images taken by drones are variable, so when the viewing angle changes, inaccurate detection results will occur. Summary of the invention

[0005] In view of this, the present invention provides an insulator abnormality detection method, device, medium and equipment, the main purpose of which is to solve the problem of inaccurate detection results in current insulator tilt detection.

[0006] In order to solve the above problems, the present application provides an insulator abnormality detection method, comprising: Performing image detection on a target image using a target insulator detection sub-model in a pre-trained target detection model to obtain a target first region frame corresponding to the target image and containing an insulator object; Performing image detection on the target image using the target utility pole detection submodel in the pre-trained target detection model to obtain a target second region frame corresponding to the target image and containing the utility pole object; Based on the target first area frame and the target second area frame, determine an angle difference between an insulator object in the target first area frame and a utility pole object in the target second area frame; Based on the angle difference value and a predetermined angle difference threshold, it is determined whether the insulator object in the target image is abnormal to obtain an abnormality detection result.

[0007] In order to solve the above problems, the present application provides an insulator abnormality detection device, comprising: A first detection module is used to perform image detection on a target image using a target insulator detection sub-model in a pre-trained target detection model, and obtain a target first region frame corresponding to the target image and containing an insulator object; A second inspection module is used to perform image detection on the target image using a target utility pole detection submodel in a pre-trained target detection model to obtain a target second area frame corresponding to the target image and containing a utility pole object; An angle determination module, configured to determine an angle difference between an insulator object in the first target area frame and a utility pole object in the second target area frame based on the first target area frame and the second target area frame; The abnormality determination module is used to determine whether the insulator object in the target image is abnormal based on the angle difference value and a predetermined angle difference threshold value to obtain an abnormality detection result.

[0008] In order to solve the above problems, the present application provides a storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, the steps of any of the above-mentioned insulator abnormality detection methods are implemented.

[0009] To solve the above problems, the present application provides an electronic device, comprising at least a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any of the above-mentioned insulator abnormality detection methods when executing the computer program on the memory.

[0010] The insulator abnormality detection method, device, medium and equipment in the present application can accurately detect the target first area frame where the insulator object is located in the target image and the target second area frame where the electric pole object is located in the target image by using the target detection model. Subsequently, the offset angle / angle difference of the insulator relative to the electric pole can be accurately determined based on the target first area frame and the target second area frame, and then the predetermined angle difference threshold can be used to accurately determine whether the insulator is tilted according to the offset angle / angle difference, so as to obtain the abnormality detection result of the insulator, thereby improving the accuracy of the detection result. The problem that the insulator tilt detection in the prior art needs to be performed using images shot at a specific angle, and when the shooting angle changes / there is an error, the insulator tilt detection is inaccurate.

[0011] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented according to the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. Moreover, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings: Figure 1 This is a flow chart of an insulator abnormality detection method according to an embodiment of the present application; Figure 2 This is a structural block diagram of an insulator abnormality detection device according to an embodiment of the present application; Figure 3 This is a structural block diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0013] Various aspects and features of the present application are described herein with reference to the accompanying drawings.

[0014] It should be understood that various modifications may be made to the embodiments of the present application. Therefore, the above description should not be considered as limiting, but only as an example of an embodiment. Other modifications within the scope and spirit of the present application will occur to those skilled in the art.

[0015] The accompanying drawings, which are incorporated in and constitute a part of the specification, illustrate embodiments of the present application and, together with the general description of the present application given above and the detailed description of the embodiments given below, serve to explain the principles of the present application.

[0016] These and other characteristics of the present application will become apparent from the following description of a preferred form of embodiment given as a non-limiting example with reference to the accompanying drawings.

[0017] It should also be understood that although the present application has been described with reference to some specific examples, those skilled in the art will be able to readily implement many other equivalent forms of the present application.

[0018] The above and other aspects, features and advantages of the present application will become more apparent in view of the following detailed description when taken in conjunction with the accompanying drawings.

[0019] Specific embodiments of the present application are described hereinafter with reference to the accompanying drawings; however, it should be understood that the embodiments applied for are merely examples of the present application, which may be implemented in a variety of ways. Well-known and / or repeated functions and structures are not described in detail to avoid unnecessary or redundant details that obscure the present application. Therefore, the specific structural and functional details applied for herein are not intended to be limiting, but merely serve as a basis and representative basis for the claims to teach those skilled in the art to use the present application in a variety of ways with substantially any suitable detailed structure.

[0020] This specification may use the phrases "in one embodiment," "in another embodiment," "in yet another embodiment," or "in other embodiments," all of which may refer to one or more of the same or different embodiments according to the present application.

[0021] The present application embodiment provides a method for detecting abnormality of an insulator, such as Figure 1 As shown, including: Step S101, performing image detection on a target image using a target insulator detection sub-model in a pre-trained target detection model, and obtaining a target first region frame corresponding to the target image and containing an insulator object; In this embodiment, the target detection model may be a YOLOV7 detection model, which may include: a target feature extraction module, a target insulator detection sub-model, and a target pole detection sub-model. Specifically, the target feature extraction module may be used to extract features from the target image to obtain a number of target feature images / target feature layers, and then the target insulator detection sub-model may be used to perform insulator detection on each target feature image / target feature layer, thereby accurately obtaining the target first region frame where the insulator is located; Step S102, performing image detection on the target image using the target utility pole detection submodel in the pre-trained target detection model to obtain a target second region frame corresponding to the target image and containing the utility pole object; In the specific implementation process of this step, the target electric pole detection sub-model can be used to perform electric pole detection on each target feature image / target feature layer corresponding to the target image, so as to accurately obtain the target second area frame of the area where the electric pole is located; Step S103, determining an angle difference between an insulator object in the first target area frame and a utility pole object in the second target area frame based on the first target area frame and the second target area frame; In the specific implementation process of this step, the first symmetry axis can be determined based on the target first area frame along the length direction of the target first area frame. At the same time, based on the target second area frame, the second symmetry axis can be determined along the length direction of the target second area frame. Then, the angle between the first symmetry axis and the second symmetry axis is determined to obtain the angle difference between the insulator and the pole. Alternatively, the first symmetry axis is determined based on the target first area frame along the length direction of the target first area frame, and then the offset angle of the target first area frame relative to the target image is determined based on the first symmetry axis to obtain the first angle; similarly, the second symmetry axis is determined based on the target second area frame along the length direction of the target second area frame, and then the offset angle of the target second area frame relative to the target image is determined based on the second symmetry axis to obtain the second angle, and finally the angle difference is determined based on the first angle and the second angle.

[0022] Step S104: determining whether the insulator object in the target image is abnormal based on the angle difference value and a predetermined angle difference threshold, so as to obtain an abnormality detection result.

[0023] In this step, the angle difference threshold value can be set and adjusted according to actual needs, for example, the angle difference threshold value can be pre-set to 1 degree, 2 degrees, 2.5 degrees, 5 degrees, 10 degrees, 15 degrees, etc. Then, the angle difference value can be compared with the angle difference threshold value, and when it is determined that the angle difference value is less than or equal to the angle difference threshold value, it is determined that the insulator is not tilted, so as to obtain a detection result that the insulator is not abnormal; on the contrary, when it is determined that the angle difference value is greater than the angle difference threshold value, it is determined that the insulator is tilted, so as to obtain a detection result that the insulator is abnormal.

[0024] The insulator abnormality detection method, device, medium and equipment in the present application can accurately detect the target first area frame where the insulator object is located in the target image and the target second area frame where the electric pole object is located in the target image by using the target detection model. Subsequently, the offset angle / angle difference of the insulator relative to the electric pole can be accurately determined based on the target first area frame and the target second area frame, and then the predetermined angle difference threshold can be used to accurately determine whether the insulator is tilted according to the offset angle / angle difference, so as to obtain the abnormality detection result of the insulator, thereby improving the accuracy of the detection result. The problem that the insulator tilt detection in the prior art needs to be performed using images shot at a specific angle, and when the shooting angle changes / there is an error, the insulator tilt detection is inaccurate.

[0025] Another embodiment of the present application provides an insulator abnormality detection method, including: In this embodiment, before detecting the target image, it also includes: using the feature extraction module in the pre-trained target detection model to extract features from the target image to obtain several target feature images for image detection. The process of obtaining each target feature image is: Step 1: Based on the target backbone feature extraction network in the feature extraction module, feature extraction is performed on the target image to obtain a number of valid feature images; In the specific implementation of this step, the several effective feature images include: a first effective feature image, a second effective feature image, and a third effective feature image; wherein the first effective feature image, the second effective feature image, and the third effective feature image have different sizes, the number of channels of the first effective feature image is smaller than that of the second effective feature image, and the number of channels of the second effective feature image and the third effective feature image are the same. Specifically, the first effective feature image is feat1=(80,80,512), the second effective feature image is feat2=(40,40,1024), and the third effective feature image is feat3=(20,20,1024).

[0026] Step 2: Based on the target feature pyramid network, feature fusion processing is performed on each of the effective feature images to obtain a target feature image corresponding to each effective feature image.

[0027] In this step, three target feature images can be obtained by subjecting the three effective feature images to feature fusion / feature enhancement processing.

[0028] In the specific implementation process, when performing feature fusion processing on each effective feature image to obtain the target feature image, the following steps are specifically included: Step (1), performing pooling processing on the third effective feature image to obtain a pooled third effective feature image p5; In this step, the third effective feature image feat3=(20,20,1024) can be pooled based on the spatial pyramid pooling layer / spatial pyramid pooling structure SPPCSPC to obtain the pooled third effective feature image p5.

[0029] Step (2), fusing the pooled third effective feature image p5 with the second effective feature image to obtain a first fused image; In the specific implementation process of this step, based on a 1*1 convolutional layer and an upsampling layer, the pooled third effective feature image p5 can be convolved and upsampled in sequence to obtain an upsampled third effective feature image; and then the upsampled third effective feature image is fused with the second effective feature image feat2=(40,40,1024) to obtain a first fused image.

[0030] Step (3), extracting features from the first fused image to obtain a first feature image P4=(40,40,256); In this step, the first feature image P4 can be obtained by performing feature extraction on the first fused image based on the multi-branch stacking module Multi_Concat_Block in the target feature extraction module.

[0031] Step (4), performing feature fusion on the first feature image and the first effective image to obtain a first target feature image P3_out corresponding to the first effective feature image; In the specific implementation process of this step, based on a 1*1 convolution layer and an upsampling layer, the first feature image P4 can be convolved and upsampled in sequence to obtain an upsampled first feature image; then the upsampled first feature image and the first effective image feat1=(80,80,512) are feature fused to obtain a second fused image, and then features are extracted from the second fused image to obtain a first target feature image P3_out corresponding to the first effective feature image, and the shape of the first target feature image P3_out is (80,80,128).

[0032] Step (5), stacking the first target feature image P3_out and the first feature image P4 to obtain a second target feature image P4_out corresponding to the second effective feature image; In this step, the first target feature image can be downsampled based on the transition module Transition_Block in the target feature extraction module to obtain the downsampled first target feature image; then the downsampled first target feature image P3_out is stacked with the first feature image P4 to obtain a first stacked image, and features are extracted from the first stacked image to obtain a second target feature image P4_out corresponding to the second effective feature image, and the shape of the second target feature image P4_out is (40, 40, 256).

[0033] Step (6): stacking the second target feature image P4_out and the third effective feature image p5 after pooling to obtain a third target feature image corresponding to the third effective feature image.

[0034] In this step, the second target feature image P4_ou can be downsampled based on the transition module Transition_Block in the target feature extraction module to obtain the downsampled second target feature image; then the downsampled second target feature image P4_ou is stacked with the pooled third effective feature image p5 to obtain a second stacked image, and then feature extraction is performed on the second stacked image to obtain a third target feature image P5_out (20,20,512) corresponding to the third effective feature image, and the shape of the third target feature image is (20,20,512).

[0035] In this embodiment, by adopting the above scheme, each target feature image corresponding to the target image can be accurately and quickly obtained, which lays a foundation for subsequent accurate insulator area frame detection and utility pole area frame detection based on each target feature image.

[0036] On the basis of the above embodiments, another embodiment of the present application provides an insulator abnormality detection method. In this embodiment, when performing image detection to obtain a target first area frame, that is, when executing step S101, the specific process is as follows: based on the target insulator detection sub-model, insulator detection is performed on each of the target feature images to obtain at least one first priori frame corresponding to each feature point in each target feature image and a first detection parameter corresponding to the first priori frame; based on at least one first priori frame corresponding to each feature point in each target feature image, merge to obtain at least one first target priori frame corresponding to each feature point in the target image and a first target detection parameter corresponding to the first target priori frame; based on each of the first target detection parameters, determine the insulator feature point from each feature point, and merge the first target priori frames corresponding to each insulator feature point to obtain an insulator prediction frame corresponding to the target image to obtain the target first area frame.

[0037] In the specific implementation process of this embodiment, each feature point in each target feature image can be detected based on the target insulator detection submodel / first detection head (Yolo Head) in the target YOLOV7 model, so as to obtain several prior frames corresponding to each feature point. For example, three prior frames corresponding to each feature point are obtained. Among them, each prior frame corresponds to detection information, and the detection information includes: regression parameters, first detection parameters of objects containing objects, and second detection parameters of object types. In the specific implementation process of this embodiment, the three first prior frames in the three target feature images corresponding to the same feature point can be merged respectively, so as to obtain three first target test frames corresponding to each feature point in the target image. A first prediction frame is determined from the three first target prior frames, and then the insulator feature points of the same type of insulator are determined from each feature point of the target image according to the first detection parameters and the second detection parameters in the first target detection information corresponding to each prediction frame, and the first prediction frames corresponding to the insulator feature points are merged to obtain the insulator prediction frame, that is, the target first area frame is obtained.

[0038] In this embodiment, the target insulator detection sub-model / first detection head (Yolo Head) introduces RepConv in a predetermined part of the network, thereby reducing the complexity of the network while keeping the prediction performance unchanged. For each feature layer / each target feature image, the number of channels is adjusted by a convolution, and the final number of channels is related to the number of categories that need to be distinguished. When performing target first area box detection, for each feature point on each target feature image / layer, 3 prior boxes of different sizes / different sizes and the corresponding first detection information are predicted. Since the categories of insulators detected in this embodiment include 7 types, the final dimension should be 39 = 3×(6+7), where 39 can be split into 3 13s, corresponding to the 13 parameters of the prior box, and 13 can be split into 5+1+7. The first five parameters are used to determine the regression parameters (center coordinates, length, width, angle) of each feature point, namely (xc, yc, w, h, theta). After adjusting the regression parameters, the prediction box can be obtained; the sixth parameter is used to determine whether each feature point contains an object; the last seven parameters are used to determine the type of object contained in each feature point.

[0039] In this embodiment, after obtaining three prior frames for each feature point in the target image, a prior frame with the largest probability value can be selected from the three prior frames as the prediction frame of the feature point, and then the prediction frames whose feature points are insulator points are merged to obtain the insulator prediction frame corresponding to the target feature image, that is, the target first area frame is obtained.

[0040] Similarly, in this embodiment, when performing image detection to obtain the target second area frame, that is, when executing step S102, telephone pole detection can be performed on each of the target feature images based on the target telephone pole detection submodel to obtain at least one second priori frame corresponding to each feature point in each target feature image and second detection information corresponding to the second priori frame; based on at least one second priori frame corresponding to each feature point in each target feature image, merge to obtain at least one second target priori frame corresponding to each feature point in the target image and second target detection information corresponding to the second target priori frame; based on each of the second target detection information, determine the telephone pole feature point from each feature point, and merge the second target priori frames corresponding to each telephone pole feature point to obtain a telephone pole prediction frame corresponding to the target image, so as to obtain the target second area frame.

[0041] In the specific implementation process of this embodiment, each feature point in each target feature image can be detected based on the target pole detection submodel / second detection head (Yolo Head) in the target YOLOV7 model, so as to obtain several a priori frames corresponding to each feature point. For example, three a priori frames corresponding to each feature point are obtained. Among them, each a priori frame corresponds to detection information, and the detection information includes: regression parameters, second detection parameters of objects containing objects, and second detection parameters of object types. In the specific implementation process of this embodiment, the three second a priori frames in the three target feature images corresponding to the same feature point can be merged respectively, so as to obtain three second target detection frames corresponding to each feature point in the target image. A second prediction frame is determined from the three second target a priori frames, and then the pole feature points are further determined from each feature point of the target image according to the second detection parameters and the second detection parameters in the second target detection information corresponding to each prediction frame, and the second prediction frames corresponding to the pole feature points are merged to obtain the pole prediction frame, that is, the target second area frame.

[0042] In this embodiment, when performing target second area frame detection, three prior frames of different sizes will be predicted for each feature point on each target feature image / layer. Since there is only one type of telephone pole detected in this embodiment, the final dimension is 21 = 3×(6+1). Among them, 21 can be split into 3 7s, corresponding to the 7 parameters of the 3 prior frames, and 7 can be split into 5+1+1. The first 5 parameters are used to determine the regression parameters of each feature point, the 6th parameter is used to determine whether each feature point contains an object; the 7th parameter is used to determine the type of object contained in each feature point.

[0043] In this embodiment, after obtaining three prior frames for each feature point in the target image, a prior frame with the largest probability value can be selected from the three prior frames as the prediction frame of the feature point, and then the prediction frames of the feature points being the utility pole feature points are merged to obtain the utility pole prediction frame corresponding to the target feature image, that is, to obtain the target second area frame.

[0044] In this embodiment, the first target area frame and the second target area frame can be accurately predicted by adopting the above method, thereby laying a foundation for subsequently accurately determining the angle difference based on the first target area frame and the second target area frame.

[0045] On the basis of the above embodiments, another embodiment of the present application provides an insulator abnormality detection method. In this embodiment, after obtaining the target first area frame and the target second area frame, step S103 can be executed, that is, based on the target first area frame and the target second area frame, the angle difference between the insulator object in the target first area frame and the electric pole object in the target second area frame is determined. In this embodiment, when determining the angle difference, the following method can be specifically adopted: According to the first angle theta corresponding to the first target area frame and the angle theta corresponding to the second target area frame, the angle difference is determined to obtain the angle difference of the insulator object relative to the electric pole. That is to say, when obtaining the first target area frame, since the first target area frame is obtained based on the merging of the regression parameters of each prediction frame, the first target area frame will correspond to the target regression parameter, which also includes the center coordinates, length, width, angle), that is, (xc, yc, w, h, theta), so the angle in the target regression parameter of the first target area frame is used as the first angle of the first target area frame. Similarly, according to the target regression parameter corresponding to the second target area frame, the second angle corresponding to the second target area frame is determined, that is, the second angle of the second target area frame relative to the target image is determined. Subsequently, the angle difference between the first angle and the second angle can be determined, and it can be compared with the target angle difference threshold to determine whether the insulator is tilted, and then obtain the abnormal detection result.

[0046] In the specific implementation process of this embodiment, the angle difference can also be determined based on the symmetry axis of the first target area frame and the symmetry axis of the target second area frame. That is to say, after obtaining the target first area frame, the first symmetry axis corresponding to the first target area frame can be determined based on the center coordinates, length, and width of the target first area frame. Similarly, the second symmetry axis corresponding to the second target area frame is determined based on the center coordinates, length, and width of the target second area frame. The angle difference can be determined based on the first symmetry axis and the second symmetry axis, and compared with the target angle difference threshold to determine whether the insulator is tilted, thereby obtaining an abnormal detection result.

[0047] In this embodiment, the first angle of the insulator is Angle_Ins, the second angle of the pole is Angel_Pole, and the angle threshold is set to 13°.

[0048] In this embodiment, by adopting the above method to obtain the target second area frame, the determination of the target second area frame can be made more reasonable and accurate, laying a foundation for subsequent accurate detection of insulator tilt / abnormality based on the target second area frame.

[0049] Based on the above embodiment, another embodiment of the present application provides an insulator abnormality detection method. In the specific implementation process of this embodiment, before performing image detection, the target detection model can also be pre-trained to obtain the target detection model. The training process of the target detection model in this embodiment is as follows: Step 1: Obtain several sample images; In this step, several images can be collected in advance for the installation conditions of the seven types of insulators, so as to obtain several sample images. Specifically, a collection of drone inspection images can be obtained. Ensure that the number of insulators of each type is balanced to eliminate the impact of color and category distribution on the detection results. Among them, the categories of insulators are shown in Table 1 below.

[0050] Table 1: Serial number Insulator Type 1 Pin Insulator_Single Umbrella_White 2 Pin Insulator_Double Umbrella_White 3 Pin Insulator_Triple Umbrella_White 4 Pin Insulator_Single Umbrella_Brown 5 Pin Insulator_Double Umbrella_Brown 6 Pin Insulator_Triple Umbrella_Brown 7 Post Insulator_White Step 2: label the insulator object and the electric pole object in the sample image with a sample area frame / selection frame; In this step, you can use the labelme tool to use the cross-marking method to mark the box area where the seven types of insulators and utility poles are located. Find the four points of the target outer contour, first connect the lines of the opposite sides, connect the four vertices to form a cross, and then convert the cross mark into a rotated rectangle facing the target. That is, use a diagonal line as the baseline, and then make perpendicular lines from the other two points to the baseline, translate and complete to form a rectangle, and obtain the first label box corresponding to the insulator object and the second label box corresponding to the utility pole; Step 3: Based on the first label frame, determine the first regression parameter label corresponding to the first label frame, the first object existence parameter label of whether the object is included, and the first category parameter label of the included object class to obtain the insulator label information; at the same time, based on the second label frame, determine the second label regression parameter label corresponding to the second label frame, the second object existence parameter label of whether the object is included, and the second category parameter label of the included object class to obtain the pole label information; In this step, the rotation angle θ can be calculated according to the four vertex coordinates of the first sample frame / the second sample frame. The long side representation method (format: label, center coordinate, length, width, angle θ) is used. The long side of the rectangular frame is defined as the height, and the short side of the rectangular frame is defined as the width. At this time, the value range of θ is [0, θ) Step 4: Divide the sample images according to the ratio of training set: validation set: test set = 0.7:0.15:0.15.

[0051] Step 5: extract features from each sample in the training set based on the initial feature extraction module in the initial detection model to obtain several target sample feature images corresponding to each sample image; In this step, for each sample image in the training set, validation set, and test set, each sample image may be resized in advance to obtain a sample image of a predetermined size, for example, a sample image of a size of 1024×1024. Then, each sample image of a size of 1024×1024 in the training set is input into the initial detection model, and the initial feature extraction module in the initial detection model is used to extract features from each sample image to obtain a number of target sample feature images corresponding to each sample image for image detection.

[0052] In this step, when the initial feature extraction module in the initial detection model performs feature extraction on each sample image to obtain a number of target sample feature images corresponding to each sample image for image detection, it specifically includes: based on the target backbone feature extraction network in the feature extraction module, performing feature extraction on each sample image to obtain a number of sample effective feature images; based on the target feature pyramid network, performing feature fusion processing on each of the sample effective feature images to obtain a target sample feature image corresponding to each sample effective feature image.

[0053] In this embodiment, the process of obtaining the target sample characteristic image is similar to the principle of obtaining the target characteristic image in the above embodiment, and will not be repeated here.

[0054] Step 6: Use the initial insulator detection sub-model in the initial detection model to perform image detection on each sample image in the training set to obtain insulator detection information corresponding to each sample image; wherein the insulator detection information includes: a first predicted regression parameter, a predicted first object existence parameter, and a predicted first category parameter; In this step, the specific process of obtaining insulator detection information is as follows: (1) Based on the initial insulator detection sub-model, each target sample feature image of the same sample image is detected to obtain a number of prior frames corresponding to each feature point in each target sample feature image; (2) For the same sample image, based on the several prior frames corresponding to each feature point in each target sample feature image, determine several prior frames (specifically, three prior frames) corresponding to each feature point in the sample image to obtain insulator detection information corresponding to the sample image; In this step, since there can be three target sample feature images, the nine prior frames corresponding to the same feature point in the three target sample feature images can be merged respectively, so as to obtain three prior frames corresponding to the feature points in the sample image, and obtain the first predicted regression parameter, the predicted first object existence parameter and the predicted first category parameter of each prior frame, so as to obtain the insulator detection information of the sample image.

[0055] Step 7: Use the initial utility pole detection sub-model in the initial detection model to perform image detection on each target sample feature image corresponding to each sample image in the training set to obtain utility pole detection information corresponding to the utility pole in each sample image; In this step, the specific process of obtaining the pole detection information is as follows: (1) Based on the initial utility pole detection sub-model, each target sample feature image of the same sample image is detected to obtain several prior frames corresponding to each feature point in each target sample feature image; (2) For the same sample image, based on the several prior frames corresponding to each feature point in each target sample feature image, determine several prior frames corresponding to each feature point in the sample image (specifically, three prior frames may be included) to obtain the electric pole detection information corresponding to the sample image; In this step, since there can be three target sample feature images, the nine prior frames corresponding to the same feature point in the three target sample feature images can be merged respectively, so as to obtain three prior frames corresponding to each feature point in the sample image, and obtain the second predicted regression parameter, the predicted second object existence parameter and the predicted second category parameter of each prior frame, so as to obtain the telephone pole detection information of the sample image.

[0056] Step 8. Determine the insulator loss value and the pole loss value based on the insulator label information, the pole label information, the insulator detection information and the pole detection information corresponding to the same sample image, and determine whether a predetermined training stop condition is met; and when the predetermined training stop condition is not met, adjust the parameters of the initial feature extraction module, the initial insulator detection sub-model and the initial pole detection sub-model in the initial detection model based on the insulator loss value and the pole loss value to obtain the current detection model, and re-perform image detection on each sample image based on the current detection model to obtain the current insulator detection information and the current pole detection information corresponding to each sample image; until it is determined that the current detection model meets the predetermined training stop condition, stop training and use the current detection model as the target detection model.

[0057] In this step, the initial detection model can specifically be an initial YOLOV7 model. The predetermined training stop condition is that the current number of training rounds reaches a predetermined training round threshold. Specifically, the training round threshold can be set to 300 rounds, 500 rounds, etc. In each round of training, the insulator detection information and the pole detection information are obtained based on the current detection model prediction / detection, and then the insulator loss value and the pole loss value are determined based on the insulator label information, pole label information, insulator detection information and pole detection information corresponding to the same sample image, and the model parameters in the current detection model are adjusted according to the insulator loss value and the pole loss value to obtain the next detection model for the next round.

[0058] When determining the insulator loss value and the utility pole loss value based on the insulator label information, the utility pole label information, the insulator detection information, and the utility pole detection information corresponding to the same sample image, the processes are as follows: Among them, the calculation process / determination process of the insulator loss value is as follows: (1) Based on each first target feature point in the first label frame of each sample image, determine the first real frame corresponding to each first target feature point and the sub-label information corresponding to each first real frame; wherein each sub-label information includes: a first regression parameter sub-label of the first real frame, a first object existence parameter sub-label of whether an object is included, and a first category parameter sub-label of the included object class. The first target feature point in this step refers to: an image point of an insulator in the sample image.

[0059] In this step, when determining the sub-label information corresponding to each first real frame, it can be determined specifically according to the insulating sub-label information of the sample image.

[0060] (4) For the same sample image, based on the first true frame of each first target feature point, a plurality of prior frames of all feature points in the sample image are screened to obtain a first positive sample prior frame corresponding to each first target feature point, and each of the first positive sample prior frames is used as a first prediction frame, and the remaining prior frames are used as first negative sample prior frames; In the specific implementation of this step, each first true frame can be matched with the prior frames (3 prior frames) on all feature points to determine the aspect ratio; if the ratio difference between any prior frame and each first true frame is greater than a threshold, the prior frame is determined to be the first negative sample prior frame. If the ratio difference between a prior frame and a first true frame is small, the prior frame is determined to be the first positive sample prior frame.

[0061] Then, based on SimOTA adaptive matching, the specific process is: calculate the Cost cost matrix between each first real frame and each prior frame of each feature point. Cost consists of two parts: overlap degree and category prediction accuracy. The higher the overlap degree, the smaller the Cost; the higher the category prediction accuracy, the smaller the Cost; then the prior frame with a cost value Cost less than a predetermined threshold is used as the positive sample prior frame of the first real frame. In other words, if the cost value Cost of a prior frame and a first real frame is less than a predetermined threshold, the prior frame is determined to be the first positive sample prior frame / first prediction frame of the first real frame.

[0062] (5) For the same sample image, the KLD loss is calculated based on the first regression parameter sub-label corresponding to each first true frame and the first predicted regression parameter of the first predicted frame corresponding to each first true frame to obtain the insulator regression loss Reg; (6) For the same sample image, based on the prediction results of whether each first prediction box contains an object (the first object existence parameter obtained by prediction) and the prediction results of whether each second negative sample prior box contains an object (the first object existence parameter obtained by prediction), the cross entropy loss is calculated to obtain the insulator object existence loss Obj; (7) For the same sample image, the cross entropy loss is calculated based on the category prediction results of each positive sample prior box / prediction box (the first category parameter obtained by prediction) and the label type of the corresponding true box to obtain the insulator category loss Cls.

[0063] (8) The insulator regression losses Reg corresponding to each sample image are merged to obtain the target insulator regression loss; the insulator object existence losses Obj corresponding to each sample image are merged to obtain the target insulator object existence loss; the insulator category losses Cls corresponding to each sample image are merged to obtain the target insulator category loss; finally, the target insulator regression loss, the target insulator object existence loss and the target insulator category loss are weightedly summed to obtain the insulator loss value.

[0064] Similarly, the calculation process / determination process of the pole loss value in this step is as follows: (1) Based on each second target feature point in the second label frame of each sample image, determine the second real frame corresponding to each second target feature point and the sub-label information corresponding to each second real frame; wherein each sub-label information includes: a second regression parameter sub-label of the second real frame, a second object existence parameter sub-label of whether an object is included, and a second category parameter sub-label of the included object class. The second target feature point in this step refers to: the image point of the electric pole in the sample image.

[0065] In this step, when determining the sub-label information corresponding to each second real frame, it can be determined specifically according to the telephone pole label information of the sample image.

[0066] (4) For the same sample image, based on the second true frame of each second target feature point, a plurality of prior frames of all feature points in the sample image are screened to obtain a second positive sample prior frame corresponding to each second target feature point, and each of the second positive sample prior frames is used as the second predicted frame, and the remaining prior frames are used as the second negative sample prior frames; In the specific implementation process of this step, each second real frame can be matched with the prior frames (3 prior frames) on all feature points to determine the aspect ratio; if the ratio difference between any prior frame and each second real frame is greater than a threshold, the prior frame is determined to be a second negative sample prior frame. If the ratio difference between a prior frame and a second real frame is small, the prior frame is determined to be a second positive sample prior frame.

[0067] Then, based on SimOTA adaptive matching, the specific process is: calculate the Cost cost matrix between each second real frame and each prior frame of each feature point. Cost consists of two parts: overlap degree and category prediction accuracy. The higher the overlap degree, the smaller the Cost; the higher the category prediction accuracy, the smaller the Cost; then the prior frame with a cost value Cost less than a predetermined threshold is used as the positive sample prior frame of the second real frame. In other words, if the cost value Cost of a prior frame and a second real frame is less than a predetermined threshold, the prior frame is determined to be the second positive sample prior frame / second prediction frame of the second real frame.

[0068] (5) For the same sample image, the KLD loss is calculated based on the second regression parameter sub-label corresponding to each second true frame and the second predicted regression parameter of the second predicted frame corresponding to each second true frame to obtain the electric pole regression loss Reg; (6) For the same sample image, based on the prediction results of whether each second prediction box contains an object (the second object existence parameter obtained by prediction) and the prediction results of whether each second negative sample prior box contains an object (the second object existence parameter obtained by prediction), the cross entropy loss is calculated to obtain the pole object existence loss Obj; (7) For the same sample image, the cross entropy loss is calculated based on the category prediction results of each positive sample prior frame / prediction frame (the second category parameter obtained by prediction) and the label type of the corresponding true frame to obtain the pole category loss Cls.

[0069] (8) Combine the telephone pole regression losses Reg corresponding to each sample image to obtain the target telephone pole regression loss; combine the telephone pole object existence losses Obj corresponding to each sample image to obtain the target telephone pole object existence loss; combine the telephone pole category losses Cls corresponding to each sample image to obtain the target telephone pole category loss; finally, perform weighted summation on the target telephone pole regression loss, target telephone pole object existence loss and target telephone pole category loss to obtain the telephone pole loss value.

[0070] In this embodiment, after obtaining the insulator loss value and the pole loss value, the difference between the model prediction value and the true value can be measured. Then, according to the calculated total loss function and using the back propagation algorithm, the parameters of the model will be adjusted according to the size of the loss value to minimize the difference, thereby improving the accuracy of the model. That is, the model parameters in the insulator detection sub-model can be adjusted according to the insulator loss value, the model parameters in the pole detection sub-model can be adjusted based on the pole loss value, and the model parameters in the feature extraction model can also be adjusted based on the insulator loss value and the pole loss value.

[0071] In this embodiment, after obtaining the target detection model, the target detection model can also be verified and tested based on the sample images in the verification set and the test set. The specific verification process, test process and the above-mentioned training process type will not be repeated here.

[0072] Based on the insulator anomaly detection method in this application, the distribution network drone inspection images provided by Shanxi Power Grid Company are used as the experimental dataset. This dataset contains drone inspection images in various distribution network scenarios, with rich scene changes and fault modes. The dataset has a total of 1,360 images, and the number of various labels and their ROIs are shown in Table 2.

[0073] Table 2: Tag Name Label meaning quantity pin_ins_1_w Pin insulator _ Single umbrella _ White 1392 pin_ins_2_w Pin insulator _ Double Umbrella _ White 711 pin_ins_3_w Pin insulator _ Three umbrellas _ White 241 pin_ins_1_b Pin insulator _ Single umbrella _ brown 275 pin_ins_2_b Pin insulator _ Double Umbrella _ brown 586 pin_ins_3_b Pin insulator _ Three umbrellas _ brown 230 post_ins_w Post insulator _ White 476 Pole Telephone pole 1384 The entire dataset is divided into three data subsets: training set, validation set, and test set. Specifically, the dataset is randomly divided according to the ratio of 0.7:0.15:0.15, of which 70% of the dataset is used to train the model, 15% of the dataset is used to verify the model performance and adjust the model hyperparameters, and the remaining 15% of the dataset is used to test the generalization ability of the model. In the process of dividing the data subsets, the above division method can ensure that the image distribution in each set is as consistent as possible, so as to reduce the impact of data deviation on the experimental results.

[0074] Performance Indicators In order to comprehensively evaluate the performance of the detection model in the distribution network fault detection task, accuracy, recall and F1 value are used as the main performance evaluation indicators. Accuracy reflects the ability of the model to correctly identify fault images, while recall reflects the sensitivity of the model to fault images. F1 value is the harmonic mean of accuracy and recall, which can comprehensively reflect the overall performance of the model.

[0075] Because the output results of the YOLOV7 rotating box target detection model / target YOLOV7 detection model simultaneously obtain the detection target category information, location information and tilt angle information, the evaluation should first judge the tilt angle based on the correct recognition.

[0076] In the test of the target YOLOV7 detection model, the classification, location information, and angle information of insulators and poles can be obtained at the same time, providing comprehensive detection results. Specifically, the target YOLOV7 detection model showed good performance in the insulator recognition task. The performance indicators of the insulator and pole recognition results are shown in Table 3. The test indicators of each type of insulator recognition reached more than 75%. This result meets the requirements of type recognition and lays a solid foundation for subsequent defect detection tasks.

[0077] Table 3:

[0078] Based on the high accuracy of multiple types of insulator identification, tilt defect detection is further carried out. The experiment aims to detect possible tilt defects of insulators by using the angle information of the identified insulators and poles. As shown in Table 4, the performance indicators of the insulator tilt defect detection results are shown.

[0079] Table 4:

[0080] As shown in Table 4, by adopting the target detection model in this application, multiple types of insulator tilt defects can be basically detected. The effectiveness of the target detection model in this application in the tilt defect detection task is verified, and certain technical support is also provided for the actual distribution network operation and maintenance work.

[0081] Another embodiment of the present application provides an insulator abnormality detection device, such as Figure 2 As shown, including: A first detection module 11 is used to perform image detection on a target image using a target insulator detection sub-model in a pre-trained target detection model, and obtain a target first region frame corresponding to the target image and containing an insulator object; The second inspection module 12 is used to perform image detection on the target image using the target electric pole detection submodel in the pre-trained target detection model to obtain a target second area frame corresponding to the target image and containing the electric pole object; An angle determination module 13 is used to determine the angle difference between the insulator object in the first target area frame and the utility pole object in the second target area frame based on the first target area frame and the second target area frame; The abnormality determination module 14 is used to determine whether the insulator object in the target image is abnormal based on the angle difference value and a predetermined angle difference threshold value, so as to obtain an abnormality detection result.

[0082] In the specific implementation process of this embodiment, the insulator abnormality detection device also includes: an extraction module, which is used to use the feature extraction module in the pre-trained target detection model to extract features from the target image and obtain several target feature images for image detection.

[0083] In the specific implementation process of this embodiment, the extraction module is specifically used to: based on the target backbone feature extraction network in the feature extraction module, perform feature extraction on the target image to obtain a number of valid feature images; based on the target feature pyramid network, perform feature fusion processing on each of the valid feature images to obtain a target feature image corresponding to each valid feature image.

[0084] In a specific implementation of this embodiment, the plurality of effective feature images include: a first effective feature image, a second effective feature image, and a third effective feature image; wherein the first effective feature image, the second effective feature image, and the third effective feature image have different sizes, the number of channels of the first effective feature image is smaller than that of the second effective feature image, and the number of channels of the second effective feature image and the third effective feature image are the same; The extraction module is specifically used for: Performing pooling processing on the third effective feature image to obtain a pooled third effective feature image; Fusing the pooled third effective feature image with the second effective feature image to obtain a first fused image; Performing feature extraction on the first fused image to obtain a first feature image; Performing feature fusion and feature extraction on the first feature image and the first effective image to obtain a first target feature image corresponding to the first effective feature image; The first target feature image is stacked with the first feature image to obtain a second target feature image corresponding to the second effective feature image; The second target feature image and the pooled third effective feature image are stacked to obtain a third target feature image corresponding to the third effective feature image.

[0085] In the specific implementation process of this embodiment, the first detection module is specifically used to: Perform insulator detection on each of the target feature images based on the target insulator detection sub-model, and obtain at least one first a priori box corresponding to each feature point in each target feature image and first detection information corresponding to the first a priori box; Based on at least one first a priori frame corresponding to each feature point in each target feature image, merging is performed to obtain at least one first target a priori frame corresponding to each feature point in the target image and first target detection information corresponding to the first target a priori frame; Based on each of the first target detection information, an insulator feature point is determined from each feature point, and the first target prior frames corresponding to each insulator feature point are merged to obtain an insulator prediction frame corresponding to the target image, so as to obtain the target first area frame.

[0086] In the specific implementation process of this embodiment, the second detection module is specifically used to: Performing electric pole detection on each of the target feature images based on the target electric pole detection submodel, obtaining at least one second priori frame corresponding to each feature point in each target feature image and second detection information corresponding to the second priori frame; Based on at least one second prior frame corresponding to each feature point in each target feature image, merging is performed to obtain at least one second target prior frame corresponding to each feature point in the target image and second target detection information corresponding to the second target prior frame; Based on each of the second target detection information, the utility pole feature points are determined from each of the feature points, and the second target prior frames corresponding to each of the utility pole feature points are merged to obtain a utility pole prediction frame corresponding to the target image, so as to obtain the target second area frame.

[0087] During the specific implementation of this embodiment, the first detection module and the abnormality determination module are specifically used to: compare the angle difference value with the angle difference threshold; when the angle difference value is less than or equal to the angle difference threshold, determine that there is no abnormality in the insulator object; when the angle difference value is greater than the angle difference threshold, determine that there is an abnormality in the insulator object.

[0088] In the specific implementation of this embodiment, the insulator abnormality detection device also includes a training module, and the training model is used to train the initial detection model to obtain the target detection model.

[0089] The insulator abnormality detection device in the present application can accurately detect the target first area frame where the insulator object is located in the target image, and the target second area frame where the electric pole object is located in the target image by using the target detection model. Subsequently, the offset angle / angle difference of the insulator relative to the electric pole can be accurately determined based on the target first area frame and the target second area frame, and then the predetermined angle difference threshold can be used to accurately determine whether the insulator is tilted according to the offset angle / angle difference, so as to obtain the abnormality detection result of the insulator, thereby improving the accuracy of the detection result. The problem that the insulator tilt detection in the prior art needs to be performed using images shot at a specific angle, and when the shooting angle changes / there is an error, the insulator tilt detection is inaccurate.

[0090] Another embodiment of the present application provides a storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, the following method steps are implemented: Step 1: Perform image detection on a target image using a target insulator detection sub-model in a pre-trained target detection model to obtain a target first region frame corresponding to the target image and containing an insulator object; Step 2: Perform image detection on the target image using the target electric pole detection submodel in the pre-trained target detection model to obtain a target second area frame corresponding to the target image and containing the electric pole object; Step 3: Based on the target first area frame and the target second area frame, determine the angle difference between the insulator object in the target first area frame and the utility pole object in the target second area frame; Step 4: Based on the angle difference value and a predetermined angle difference threshold, determine whether the insulator object in the target image is abnormal to obtain an abnormality detection result.

[0091] The specific implementation process of the above method steps can refer to the embodiments of any of the above-mentioned insulator abnormality detection methods, and this embodiment will not be repeated here.

[0092] The storage medium in the present application can accurately detect the target first area frame where the insulator object is located in the target image, and the target second area frame where the electric pole object is located in the target image by using the target detection model. Subsequently, the offset angle / angle difference of the insulator relative to the electric pole can be accurately determined based on the target first area frame and the target second area frame, and then the predetermined angle difference threshold can be used to accurately determine whether the insulator is tilted according to the offset angle / angle difference, so as to obtain the abnormal detection result of the insulator, thereby improving the accuracy of the detection result. The problem that the insulator tilt detection in the prior art requires the use of images shot at a specific angle, and the insulator tilt detection is inaccurate when the shooting angle changes / there is an error, is solved.

[0093] Another embodiment of the present application provides an electronic device, such as Figure 3 As shown, at least a memory 1 and a processor 2 are included, the memory 1 stores a computer program, and the processor 2 implements the following method steps when executing the computer program on the memory 1: Step 1: Perform image detection on a target image using a target insulator detection sub-model in a pre-trained target detection model to obtain a target first region frame corresponding to the target image and containing an insulator object; Step 2: Perform image detection on the target image using the target electric pole detection submodel in the pre-trained target detection model to obtain a target second area frame corresponding to the target image and containing the electric pole object; Step 3: Based on the target first area frame and the target second area frame, determine the angle difference between the insulator object in the target first area frame and the utility pole object in the target second area frame; Step 4: Based on the angle difference value and a predetermined angle difference threshold, determine whether the insulator object in the target image is abnormal to obtain an abnormality detection result.

[0094] The specific implementation process of the above method steps can refer to the embodiments of any of the above-mentioned insulator abnormality detection methods, and this embodiment will not be repeated here.

[0095] The storage medium in the present application can accurately detect the target first area frame where the insulator object is located in the target image, and the target second area frame where the electric pole object is located in the target image by using the target detection model. Subsequently, the offset angle / angle difference of the insulator relative to the electric pole can be accurately determined based on the target first area frame and the target second area frame, and then the predetermined angle difference threshold can be used to accurately determine whether the insulator is tilted according to the offset angle / angle difference, so as to obtain the abnormal detection result of the insulator, thereby improving the accuracy of the detection result. The problem that the insulator tilt detection in the prior art requires the use of images shot at a specific angle, and the insulator tilt detection is inaccurate when the shooting angle changes / there is an error, is solved.

[0096] The above embodiments are only exemplary embodiments of the present application and are not intended to limit the present application. The protection scope of the present application is defined by the claims. Those skilled in the art may make various modifications or equivalent substitutions to the present application within the essence and protection scope of the present application, and such modifications or equivalent substitutions shall also be deemed to fall within the protection scope of the present application.

Claims

1. A method for detecting abnormality of an insulator, characterized in that: include: Performing image detection on a target image using a target insulator detection sub-model in a pre-trained target detection model to obtain a target first region frame corresponding to the target image and containing an insulator object; Performing image detection on the target image using the target utility pole detection submodel in the pre-trained target detection model to obtain a target second region frame corresponding to the target image and containing the utility pole object; Based on the target first area frame and the target second area frame, determine an angle difference between an insulator object in the target first area frame and a utility pole object in the target second area frame; Based on the angle difference value and a predetermined angle difference threshold, it is determined whether the insulator object in the target image is abnormal to obtain an abnormality detection result.

2. The method according to claim 1, characterized in that Before detecting the target image, the method further includes: The feature extraction module in the pre-trained target detection model is used to extract features from the target image to obtain several target feature images for image detection.

3. The method according to claim 2, characterized in that The feature extraction module in the pre-trained target detection model is used to extract features from the target image to obtain a number of target feature images for image detection, specifically including: Based on the target backbone feature extraction network in the feature extraction module, feature extraction is performed on the target image to obtain a number of effective feature images; Based on the target feature pyramid network, feature fusion processing is performed on each of the effective feature images to obtain a target feature image corresponding to each effective feature image.

4. The method according to claim 3, characterized in that The plurality of effective feature images include: a first effective feature image, a second effective feature image, and a third effective feature image; wherein the first effective feature image, the second effective feature image, and the third effective feature image have different sizes, the number of channels of the first effective feature image is smaller than that of the second effective feature image, and the number of channels of the second effective feature image and the third effective feature image are the same; The method of performing feature fusion processing on the effective feature images based on the target feature pyramid network to obtain target feature images corresponding to each effective feature image specifically includes: Performing pooling processing on the third effective feature image to obtain a pooled third effective feature image; Fusing the pooled third effective feature image with the second effective feature image to obtain a first fused image; Performing feature extraction on the first fused image to obtain a first feature image; Performing feature fusion and feature extraction on the first feature image and the first effective image to obtain a first target feature image corresponding to the first effective feature image; The first target feature image is stacked with the first feature image to obtain a second target feature image corresponding to the second effective feature image; The second target feature image and the pooled third effective feature image are stacked to obtain a third target feature image corresponding to the third effective feature image.

5. The method according to claim 2, characterized in that The method of performing image detection on a target image by using a target insulator detection sub-model in a pre-trained target detection model to obtain a target first region frame corresponding to the target image and containing an insulator object specifically includes: Perform insulator detection on each of the target feature images based on the target insulator detection sub-model, and obtain at least one first a priori box corresponding to each feature point in each target feature image and first detection information corresponding to the first a priori box; Based on at least one first a priori frame corresponding to each feature point in each target feature image, merging is performed to obtain at least one first target a priori frame corresponding to each feature point in the target image and first target detection information corresponding to the first target a priori frame; Based on each of the first target detection information, an insulator feature point is determined from each feature point, and the first target prior frames corresponding to each insulator feature point are merged to obtain an insulator prediction frame corresponding to the target image, so as to obtain the target first area frame.

6. The method according to claim 2, characterized in that The method of performing image detection on the target image using the target electric pole detection submodel in the pre-trained target detection model to obtain a target second area frame corresponding to the target image and containing the electric pole object specifically includes: Performing electric pole detection on each of the target feature images based on the target electric pole detection submodel, obtaining at least one second priori frame corresponding to each feature point in each target feature image and second detection information corresponding to the second priori frame; Based on at least one second prior frame corresponding to each feature point in each target feature image, merging is performed to obtain at least one second target prior frame corresponding to each feature point in the target image and second target detection information corresponding to the second target prior frame; Based on each of the second target detection information, the utility pole feature points are determined from each of the feature points, and the second target prior frames corresponding to each of the utility pole feature points are merged to obtain a utility pole prediction frame corresponding to the target image, so as to obtain the target second area frame.

7. The method according to claim 1, characterized in that The determining whether the insulator object in the target image is abnormal based on the angle difference value and a predetermined angle difference threshold value to obtain an abnormality detection result specifically includes: comparing the angle difference value with the angle difference threshold; When the angle difference value is less than or equal to the angle difference threshold, determining that there is no abnormality in the insulator object; When the angle difference value is greater than the angle difference threshold, it is determined that an abnormality exists in the insulator object.

8. An insulator abnormality detection device, characterized in that: include: A first detection module is used to perform image detection on a target image using a target insulator detection sub-model in a pre-trained target detection model, and obtain a target first region frame corresponding to the target image and containing an insulator object; A second inspection module is used to perform image detection on the target image using a target utility pole detection submodel in a pre-trained target detection model to obtain a target second area frame corresponding to the target image and containing a utility pole object; An angle determination module, configured to determine an angle difference between an insulator object in the first target area frame and a utility pole object in the second target area frame based on the first target area frame and the second target area frame; The abnormality determination module is used to determine whether the insulator object in the target image is abnormal based on the angle difference value and a predetermined angle difference threshold value to obtain an abnormality detection result.

9. A storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the insulator abnormality detection method according to any one of claims 1 to 7 are implemented.

10. An electronic device, characterized in that: The device at least comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the insulator abnormality detection method according to any one of claims 1 to 7 when executing the computer program on the memory.

Images