Telegraph pole insulator defect detection method and device based on deep learning

Abstract

The invention discloses a telegraph pole insulator defect detection method and a telegraph pole insulator defect detection device based on deep learning, which are used for solving the technical problem of poor final telegraph pole insulator defect detection effect caused by the fact that the defect detection precision of an existing telegraph pole insulator defect detection method is still difficult to meet high-standard requirements in an actual inspection scene. The method comprises the following steps: acquiring a telegraph pole insulator image, performing enhancement operation on the telegraph pole insulator image, and outputting a target telegraph pole insulator image; inputting the target telegraph pole insulator image into an insulator defect detection model, wherein the insulator defect detection model comprises a lightweight multi-scale feature extraction unit and an adaptive feature aggregation unit; performing feature extraction on the target telegraph pole insulator image by adopting a lightweight multi-scale feature extraction unit, and outputting a target insulator multi-scale optimization feature map; and detecting the multi-scale optimization feature map of the target insulator through an adaptive feature aggregation unit, and generating a defect detection result of the telegraph pole insulator.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power systems, and in particular to a power pole insulator defect detection method and device based on deep learning. BACKGROUND

[0002] With the continuous improvement of the intelligent level of the power system, the state monitoring means of the power transmission line is increasingly rich, and the use condition of the insulator on the power pole has gradually become a key concern object in the power operation and maintenance. As an important component for ensuring the insulation performance and operation safety in the power system, once the insulator has defects such as damage, loss, and pollution, it is easy to cause power short circuit, trip, and even large-scale power outage accidents, which will pose a major threat to the safety of the power grid. Therefore, carrying out efficient and accurate insulator defect detection has important significance for improving the intelligent and proactive level of power equipment operation and maintenance.

[0003] Traditional insulator defect detection methods mostly rely on manual inspection or infrared detection means, which have many problems such as low efficiency, high cost, dependence on expert experience, and difficulty in covering a large area. Under complex environmental conditions, such as high altitude, strong wind, and weather influences such as rain and fog, manual inspection not only has high risk, but also is prone to omissions, and it is difficult to meet the detection needs of modern power grids in terms of efficiency and real-time performance.

[0004] In recent years, image recognition technology based on deep learning has made significant progress in intelligent inspection of power facilities, especially in the use of unmanned aerial vehicle images for automatic defect identification. However, in actual inspection scenarios, the insulator is small in size, detailed in structure, and easily disturbed by complex backgrounds, and its defect detection accuracy is still difficult to meet high standard requirements, resulting in poor final power pole insulator defect detection results. SUMMARY

[0005] The present application provides a power pole insulator defect detection method and device based on deep learning, which is used to solve the technical problem that the existing power pole insulator defect detection method still has poor defect detection accuracy in actual inspection scenarios, resulting in poor final power pole insulator defect detection results.

[0006] The first aspect of the present application provides a power pole insulator defect detection method based on deep learning, comprising:

[0007] Obtaining a power pole insulator image and performing an enhancement operation on the power pole insulator image to output a target power pole insulator image;

[0008] Inputting the target power pole insulator image into an insulator defect detection model, the insulator defect detection model comprising a lightweight multi-scale feature extraction unit and a self-adaptive feature aggregation unit;

[0009] The light-weight multi-scale feature extraction unit is used for feature extraction on the target power pole insulator image, and a target insulator multi-scale optimization feature map is output;

[0010] The adaptive feature aggregation unit is used for detection on the target insulator multi-scale optimization feature map, and a power pole insulator defect detection result is generated.

[0011] Optionally, the light-weight multi-scale feature extraction unit comprises a convolution module and a light-weight multi-scale feature enhancement module; the target insulator multi-scale optimization feature map comprises a first insulator multi-scale optimization feature map, a second insulator multi-scale optimization feature map and a third insulator multi-scale optimization feature map; and the light-weight multi-scale feature extraction unit is used for feature extraction on the target power pole insulator image, and a target insulator multi-scale optimization feature map is output, which comprises:

[0012] The convolution module is used for convolution operation on the target power pole insulator image, and a first power pole insulator convolution feature map is generated;

[0013] The light-weight multi-scale feature enhancement module is used for feature enhancement extraction on the first power pole insulator convolution feature map, and a first insulator multi-scale optimization feature map is output;

[0014] The convolution module is used for convolution operation on the first insulator multi-scale optimization feature map, and a second power pole insulator convolution feature map is generated;

[0015] The light-weight multi-scale feature enhancement module is used for feature enhancement extraction on the second power pole insulator convolution feature map, and a second insulator multi-scale optimization feature map is output;

[0016] The convolution module is used for convolution operation on the second insulator multi-scale optimization feature map, and a third power pole insulator convolution feature map is generated;

[0017] The light-weight multi-scale feature enhancement module is used for feature enhancement extraction on the third power pole insulator convolution feature map, and a third insulator multi-scale optimization feature map is output.

[0018] Optionally, the light-weight multi-scale feature enhancement module comprises a 1×1 two-dimensional convolution layer, a 3×3 two-dimensional convolution layer, a 3×3 depth convolution layer, a 5×5 depth convolution layer and a ReLU6 activation function layer; and the light-weight multi-scale feature enhancement module is used for feature enhancement extraction on the first power pole insulator convolution feature map, and a first insulator multi-scale optimization feature map is output, which comprises:

[0019] performing a two-dimensional convolution operation on the first power pole insulator convolution feature map by using a 1*1 two-dimensional convolution layer, to output a first power pole insulator two-dimensional convolution feature map;

[0020] inputting the first power pole insulator two-dimensional convolution feature map into a 3*3 deep convolution layer and a 5*5 deep convolution layer respectively to perform deep convolution, to output a first power pole insulator deep convolution feature map and a second power pole insulator deep convolution feature map;

[0021] taking the first power pole insulator deep convolution feature map as an input of a ReLU6 activation function layer, to output a first power pole insulator transformation feature map, and performing element-wise addition on the first power pole insulator transformation feature map and the second power pole insulator deep convolution feature map, to output a first power pole insulator addition feature map;

[0022] performing deep convolution on the first power pole insulator transformation feature map by using a 3*3 deep convolution layer, to output a third power pole insulator deep convolution feature map;

[0023] taking the first power pole insulator addition feature map as an input of a ReLU6 activation function layer, to generate a second power pole insulator transformation feature map;

[0024] performing batch normalization and channel attention calculation on the third power pole insulator deep convolution feature map and the second power pole insulator transformation feature map respectively in sequence, to output a first power pole insulator channel feature map and a second power pole insulator channel feature map;

[0025] performing element-wise addition on the first power pole insulator channel feature map and the second power pole insulator channel feature map, to output a second power pole insulator addition feature map, and performing two-dimensional convolution on the second power pole insulator addition feature map by using a 3*3 two-dimensional convolution layer, to output a second power pole insulator two-dimensional convolution feature map;

[0026] splicing the second power pole insulator two-dimensional convolution feature map, the second power pole insulator addition feature map and the first power pole insulator addition feature map, to output a power pole insulator splicing feature map, and performing two-dimensional convolution on the power pole insulator splicing feature map by using a 1*1 two-dimensional convolution layer, to generate a first insulator multi-scale optimization feature map.

[0027] Optionally, the adaptive feature aggregation unit comprises a dynamic feature reorganization module, a convolution module and a target detection head; the target insulator multi-scale optimization feature map is input into the adaptive feature aggregation unit for detection, to generate a power pole insulator defect detection result, comprising:

[0028] The dynamic feature reorganization module is adopted to perform feature reorganization on the third insulator multi-scale optimization feature map, a first insulator reorganization feature map is output, and upsampling is performed on the first insulator reorganization feature map to generate a first insulator up-sampling feature map;

[0029] The first insulator up-sampling feature map and the second insulator multi-scale optimization feature map are spliced to output a first insulator splicing feature map, and the dynamic feature reorganization module is adopted to perform feature reorganization on the first insulator splicing feature map to output a second insulator reorganization feature map;

[0030] The second insulator reorganization feature map is up-sampled to generate a second insulator up-sampling feature map, and the second insulator up-sampling feature map and the first insulator multi-scale optimization feature map are spliced to output a second insulator splicing feature map;

[0031] The dynamic feature reorganization module is adopted to perform feature reorganization on the second insulator splicing feature map to generate a third insulator reorganization feature map, and the convolution module is adopted to perform convolution operation on the third insulator reorganization feature map to output a first insulator convolution feature map;

[0032] The first insulator convolution feature map and the first insulator splicing feature map are spliced to generate a third insulator splicing feature map, and the dynamic feature reorganization module is adopted to perform feature reorganization on the third insulator splicing feature map to output a fourth insulator reorganization feature map;

[0033] The convolution module is adopted to perform convolution operation on the fourth insulator reorganization feature map to output a second insulator convolution feature map, and the second insulator convolution feature map and the first insulator reorganization feature map are spliced to output a fourth insulator splicing feature map;

[0034] The dynamic feature reorganization module is adopted to perform feature reorganization on the fourth insulator splicing feature map to generate a fifth insulator reorganization feature map;

[0035] The fifth insulator reorganization feature map, the fourth insulator reorganization feature map and the third insulator reorganization feature map are respectively input into a target detection head to output a power pole insulator defect detection result.

[0036] Optionally, the dynamic feature reorganization module comprises a 3*3 deformable convolution network layer, a re-parameterization convolution layer and a 1*1 two-dimensional convolution layer; and the dynamic feature reorganization module is adopted to perform feature reorganization on the third insulator multi-scale optimization feature map to output a first insulator reorganization feature map, which comprises:

[0037] The third insulator multi-scale optimization feature map is subjected to deformable convolution by using a 3*3 deformable convolution network layer to generate an insulator deformable convolution feature map, and the insulator deformable convolution feature map is subjected to two-dimensional convolution by using a 1*1 two-dimensional convolution layer to output a first insulator two-dimensional convolution feature map;

[0038] The first insulator two-dimensional convolution feature map is subjected to dynamic channel attention calculation and batch normalization in sequence to generate an insulator normalization feature map, and the insulator normalization feature map is subjected to reparameterization convolution by using a reparameterization convolution layer to output a first insulator reparameterization convolution feature map;

[0039] The insulator normalization feature map is subjected to maximum pooling to output a first insulator maximum pooling feature map, and the first insulator maximum pooling feature map is subjected to reparameterization convolution by using a reparameterization convolution layer to output a second insulator reparameterization convolution feature map;

[0040] The first insulator maximum pooling feature map is subjected to maximum pooling to output a second insulator maximum pooling feature map, and the second insulator maximum pooling feature map is subjected to reparameterization convolution by using a reparameterization convolution layer to output a third insulator reparameterization convolution feature map;

[0041] The second insulator maximum pooling feature map is subjected to maximum pooling to output a third insulator maximum pooling feature map, and the third insulator maximum pooling feature map is subjected to reparameterization convolution by using a reparameterization convolution layer to output a fourth insulator reparameterization convolution feature map;

[0042] The first insulator reparameterization convolution feature map, the second insulator reparameterization convolution feature map, the third insulator reparameterization convolution feature map and the fourth insulator reparameterization convolution feature map are spliced to generate a target insulator splicing feature map;

[0043] The target insulator splicing feature map is subjected to two-dimensional convolution by using a 1*1 two-dimensional convolution layer to generate a first insulator reorganization feature map.

[0044] Optionally, the model training process of the insulator defect detection model is specifically as follows:

[0045] A plurality of model training insulator images of a power pole are acquired, and the plurality of model training insulator images of the power pole are screened and labeled to output an insulator defect detection original data set;

[0046] The insulator defect detection original data set is subjected to an enhancement operation to generate an insulator defect detection data set;

[0047] The preset loss function is used to train an initial insulator defect detection model according to the insulator defect detection data set, and a trained insulator defect detection model is determined.

[0048] The second aspect of the present application provides a power pole insulator defect detection device based on deep learning, comprising:

[0049] An acquisition module is configured to acquire a power pole insulator image, perform an enhancement operation on the power pole insulator image, and output a target power pole insulator image.

[0050] An input module is configured to input the target power pole insulator image into an insulator defect detection model, wherein the insulator defect detection model comprises a lightweight multi-scale feature extraction unit and an adaptive feature aggregation unit.

[0051] A feature extraction module is configured to perform feature extraction on the target power pole insulator image by using the lightweight multi-scale feature extraction unit, and output a target insulator multi-scale optimized feature map.

[0052] A detection module is configured to perform detection on the target insulator multi-scale optimized feature map by using the adaptive feature aggregation unit, and generate a power pole insulator defect detection result.

[0053] The third aspect of the present application provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer program is executed by the processor to make the processor execute the steps of the power pole insulator defect detection method based on deep learning according to any one of the above aspects.

[0054] The fourth aspect of the present application provides a computer readable storage medium, which stores a computer program, and the computer program is executed to implement the steps of the power pole insulator defect detection method based on deep learning according to any one of the above aspects.

[0055] The fifth aspect of the present application provides a computer program product, which comprises a computer program stored on a non-transitory computer readable storage medium, and the computer program comprises program instructions, wherein when the program instructions are executed by a computer, the computer executes the steps of the power pole insulator defect detection method based on deep learning according to any one of the above aspects.

[0056] From the above technical solutions, the present application has the following advantages:

[0057] The scheme provides a power pole insulator defect detection method based on deep learning, acquires a power pole insulator image, and performs an enhancement operation on the power pole insulator image to output a target power pole insulator image; the target power pole insulator image is input into an insulator defect detection model, the insulator defect detection model comprising a lightweight multi-scale feature extraction unit and a self-adaptive feature aggregation unit; the target power pole insulator image is subjected to feature extraction by using the lightweight multi-scale feature extraction unit to output a target insulator multi-scale optimized feature map; the target insulator multi-scale optimized feature map is detected by using the self-adaptive feature aggregation unit to generate a power pole insulator defect detection result; based on the scheme, the original image quality is improved by image enhancement preprocessing, and a clear foundation is laid for subsequent feature extraction, the lightweight multi-scale feature extraction unit can capture key features of insulators that are easily ignored due to small size and detailed structure, avoid small target feature loss, and the self-adaptive feature aggregation unit can actively suppress the influence of complex background interference on feature judgment, improve the aggregation and utilization efficiency of effective features, and the three form a synergistic effect, greatly improving the power pole insulator defect detection precision, thereby improving the final power pole insulator defect detection effect. BRIEF DESCRIPTION OF DRAWINGS

[0058] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0059] Figure 1 A step flow chart of a power pole insulator defect detection method based on deep learning provided by the first embodiment of the present application is provided.

[0060] Figure 2 A structure schematic diagram of an insulator defect detection model provided by the first embodiment of the present application is provided.

[0061] Figure 3 A structure schematic diagram of a lightweight multi-scale feature extraction unit provided by the first embodiment of the present application is provided.

[0062] Figure 4 A structure schematic diagram of a lightweight multi-scale feature extraction unit provided by the first embodiment of the present application is provided.

[0063] Figure 5 A structure schematic diagram of a self-adaptive feature aggregation unit provided by the first embodiment of the present application is provided.

[0064] Figure 6 A structure schematic diagram of a dynamic feature recombination module provided by the first embodiment of the present application is provided.

[0065] Figure 7 A step flow chart of model training of the insulator defect detection model provided for the second embodiment of the present application is provided.

[0066] Figure 8 A schematic diagram of the insulator image provided for the second embodiment of the present application is provided.

[0067] Figure 9 A schematic diagram of the insulator defect detection model output result diagram provided for the second embodiment of the present application is provided.

[0068] Figure 10 A flowchart of a power pole insulator defect detection method based on deep learning provided for the second embodiment of the present application is provided.

[0069] Figure 11 A structural block diagram of a power pole insulator defect detection device based on deep learning provided for the third embodiment of the present application is provided. DETAILED DESCRIPTION

[0070] The embodiment of the present application provides a power pole insulator defect detection method and device based on deep learning, which is used to solve the technical problem that the defect detection accuracy of the existing power pole insulator defect detection method is still difficult to meet the high standard requirement in the actual inspection scene, resulting in poor final power pole insulator defect detection effect.

[0071] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application. It should be noted that in the optional embodiments of the present application, the object information and other related data involved need to be authorized or agreed by the object when the embodiments of the present application are applied to specific products or technologies, and the collection, use and processing of the related data need to comply with the relevant laws, regulations and standards of the country and region. That is, the data related to the object in the embodiments of the present application need to be obtained with the authorization and consent of the object, the authorization and consent of the relevant department, and the compliance with the relevant laws, regulations and standards of the country and region. If personal information is involved in the embodiments, the consent of the individual needs to be obtained for the acquisition of all personal information, and the separate consent of the information subject needs to be obtained for the sensitive information, and the embodiments also need to be implemented with the authorization and consent of the object.

[0072] Please refer to Figure 1 , Figure 1A step flow chart of a power pole insulator defect detection method based on deep learning is provided for the first embodiment of the present application.

[0073] The present application provides a power pole insulator defect detection method based on deep learning, comprising:

[0074] Step 101, acquire the power pole insulator image, and perform enhancement operation on the power pole insulator image, output the target power pole insulator image.

[0075] It should be noted that the image of the insulator is collected by the unmanned aerial vehicle inspection equipment, and the collection process needs to cover various weather conditions (such as sunny, cloudy, rainy and foggy days) and different time periods (sunrise, noon and evening). Then, data rotation, flipping, cropping, contrast adjustment, noise addition and other enhancement operations are taken on the collected images to obtain the target power pole insulator image.

[0076] Step 102, input the target power pole insulator image into the insulator defect detection model, and the insulator defect detection model comprises a lightweight multi-scale feature extraction unit and a self-adaptive feature aggregation unit.

[0077] It should be noted that the insulator defect detection model proposed by the present application is composed of a lightweight multi-scale feature extraction unit and a self-adaptive feature aggregation unit, and the structure is as shown in Figure 2 .

[0078] Step 103, the lightweight multi-scale feature extraction unit is used to extract features from the target power pole insulator image, and output the target insulator multi-scale optimization feature map.

[0079] The lightweight multi-scale feature extraction unit comprises a convolution module and a lightweight multi-scale feature enhancement module.

[0080] The target insulator multi-scale optimization feature map comprises a first insulator multi-scale optimization feature map, a second insulator multi-scale optimization feature map and a third insulator multi-scale optimization feature map.

[0081] It should be noted that the target insulator multi-scale optimization feature map is input into the lightweight multi-scale feature extraction unit, and different levels and different scales of insulator multi-scale optimization feature maps F3 (i.e. the first insulator multi-scale optimization feature map), insulator multi-scale optimization feature map F5 (i.e. the second insulator multi-scale optimization feature map) and insulator multi-scale optimization feature map F7 (i.e. the third insulator multi-scale optimization feature map) can be obtained, and the structure of the lightweight multi-scale feature extraction unit is as shown in Figure 3As shown, the lightweight multi-scale feature extraction unit is composed of three structurally identical convolution modules and three structurally identical lightweight multi-scale feature reinforcement modules. The convolution module (ConvBNSiLU module, Convolution-BatchNormalization-Sigmoid Linear Unit) is composed of a Conv2d layer (Convolution 2-dimensional) with a convolution kernel size of 3x3 and a step of 2, a BN layer (Batch Normalization layer), and a SiLU activation function in series. The ConvBNSiLU module can be used for downsampling to compress the feature map size, reduce the subsequent calculation amount, and provide initial feature support for deep semantics.

[0082] Further, step 103 can include the following sub-steps:

[0083] S31, a convolution module is used to perform convolution operation on the target power pole insulator image to generate a first power pole insulator convolution feature map;

[0084] It should be noted that the target power pole insulator image is input into the ConvBNSiLU module to generate the insulator feature map F2 (i.e. the first power pole insulator convolution feature map).

[0085] S32, a lightweight multi-scale feature reinforcement module is used to perform feature reinforcement extraction on the first power pole insulator convolution feature map to output a first insulator multi-scale optimization feature map;

[0086] The lightweight multi-scale feature reinforcement module includes a 1x1 two-dimensional convolution layer (Conv2d 1x1), a 3x3 two-dimensional convolution layer (Conv2d 3x3), a 3x3 depth convolution layer (DWConv 3x3, Depthwise Convolution), a 5x5 depth convolution layer (DWConv 5x5), and a ReLU6 activation function layer (Rectified Linear Unit 6).

[0087] It should be noted that F2 is input into the LMRConv module to generate the insulator multi-scale optimization feature map F3 (i.e. the first insulator multi-scale optimization feature map), wherein the present application designs a lightweight multi-scale feature reinforcement module (Lightweight Multi-Scale Feature Reinforcement Convolution Module, LMRConv), and the structure of the LMRConv module is as shown in Figure 4As shown, the LMRConv module reduces the model calculation burden through a lightweight structure, while introducing multi-scale convolution and channel attention mechanism, effectively enhancing the feature expression ability of the insulator defect target. In addition, the LMRConv module adopts a lightweight convolution structure and a multi-scale parallel branch design, significantly reducing the model calculation amount while enhancing the expression ability of the insulator crack and damage defects. Through channel compression, depth separable convolution and attention fusion, the perception accuracy of small defect targets is improved, effectively adapting to the characteristics of sparse targets and complex background in insulator images, and realizing efficient feature extraction with low computational overhead.

[0088] Further, step S32 can include the following sub-steps:

[0089] S321, a 1x1 two-dimensional convolution layer is used to perform a two-dimensional convolution operation on the first power pole insulator convolution feature map, and a first power pole insulator two-dimensional convolution feature map is output;

[0090] S322, the first power pole insulator two-dimensional convolution feature map is input into a 3x3 depth convolution layer and a 5x5 depth convolution layer respectively for depth convolution, and a first power pole insulator depth convolution feature map and a second power pole insulator depth convolution feature map are output;

[0091] S323, the first power pole insulator depth convolution feature map is taken as the input of a ReLU6 activation function layer, a first power pole insulator transformed feature map is output, and the first power pole insulator transformed feature map and the second power pole insulator depth convolution feature map are element-wise added, outputting a first power pole insulator added feature map;

[0092] S324, a 3x3 depth convolution layer is used to perform depth convolution on the first power pole insulator transformed feature map, and a third power pole insulator depth convolution feature map is output;

[0093] S325, the first power pole insulator added feature map is taken as the input of a ReLU6 activation function layer, and a second power pole insulator transformed feature map is generated;

[0094] S326, batch normalization and channel attention calculation are performed on the third power pole insulator depth convolution feature map and the second power pole insulator transformed feature map in sequence, respectively, outputting a first power pole insulator channel feature map and a second power pole insulator channel feature map;

[0095] S327, the first power pole insulator channel feature map and the second power pole insulator channel feature map are element-wise added, outputting a second power pole insulator added feature map, and a 3x3 two-dimensional convolution layer is used to perform two-dimensional convolution on the second power pole insulator added feature map, outputting a second power pole insulator two-dimensional convolution feature map;

[0096] S328, splice the second electric pole insulator two-dimensional convolution feature map, the second electric pole insulator addition feature map and the first electric pole insulator addition feature map, output an electric pole insulator splicing feature map, and perform two-dimensional convolution on the electric pole insulator splicing feature map through a 1x1 two-dimensional convolution layer to generate a first insulator multi-scale optimization feature map.

[0097] It should be noted that F2 is taken as the input feature map of the LMRConv module. For ease of description, F2 is denoted as X1 here. X1 is first input into a 1x1 two-dimensional convolution layer to perform a convolution operation with a convolution kernel size of 1x1, generating an insulator feature map X2 (i.e., a first electric pole insulator two-dimensional convolution feature map); this operation is used for channel mapping and compression, which not only reduces the calculation burden of the multi-branch structure, but also helps to unify the feature dimension and adapt to different scale branch inputs.

[0098] Further, X2 is processed in two branches:

[0099] In the first branch, X2 is input into a 3x3 deep convolution layer to perform a 3x3 deep separable convolution operation, generating an insulator feature map X3 (i.e., a first electric pole insulator deep convolution feature map); this deep separable convolution operation helps to extract local detail features of insulator edges and textures. X3 is input into a ReLU6 activation function layer to perform a nonlinear activation operation, generating an insulator feature map X4 (a first electric pole insulator transformation feature map); this activation function maintains numerical stability in low-precision calculation and enhances the expression ability of the lightweight model. X4 is input into a 3x3 deep convolution layer to perform a 3x3 deep separable convolution operation, generating an insulator feature map X5 (i.e., a third electric pole insulator deep convolution feature map); this cascaded structure can gradually enhance the sensitivity to small defects of insulator cracks. X5 is input into a Batch Norm layer to perform a batch normalization operation, generating an insulator feature map X6; this operation accelerates model convergence and alleviates the gradient vanishing problem. X6 is input into an ECA layer (Efficient Channel Attention, channel attention mechanism) to perform a channel attention operation, generating an insulator feature map X7 (i.e., a first electric pole insulator channel feature map); this mechanism focuses on key feature channels of insulator defects through local cross-channel interaction.

[0100] In the second branch, X2 is input into a 5x5 depth convolution layer to perform a depth separable convolution operation with a kernel size of 5x5, to generate an insulator feature map X8 (i.e., a second pole insulator depth convolution feature map); the large kernel convolution can capture the large-range structural features of the insulator. X4 and X8 are added element by element to generate an insulator feature map X9 (i.e., a first pole insulator addition feature map); the fusion operation integrates local details and global context information, enhancing the multi-scale defect representation capability. X9 is input into a ReLU6 activation function layer to perform a nonlinear activation operation, to generate an insulator feature map X10 (i.e., a second pole insulator transformation feature map); the nonlinear transformation of the fused features enhances the feature expression capability. X10 is input into a Batch Norm layer (Batch Normalization) to perform a batch normalization operation, to generate an insulator feature map X11; X11 is input into an ECA layer to perform a channel attention operation, to generate an insulator feature map X12 (i.e., a second pole insulator channel feature map); the channel attention operation dynamically adjusts the weights of each channel, highlighting the feature channels related to the aging and flashover defects of the insulator.

[0101] Then, X7 and X12 are added element by element to generate an insulator feature map X13 (i.e., a second pole insulator addition feature map); X13 is input into a 3x3 two-dimensional convolution layer to perform a convolution operation with a kernel size of 3x3, to generate an insulator feature map X14 (i.e., a second pole insulator two-dimensional convolution feature map); the standard convolution realizes cross-channel information interaction, optimizing the discriminability of the fused features. X14, X9, and X13 are input into a Concat layer (concatenation) to perform a channel dimension concatenation operation, to generate an insulator feature map X15 (i.e., a pole insulator concatenation feature map); X15 is input into a 1x1 two-dimensional convolution layer to perform a convolution operation with a kernel size of 1x1, to generate an insulator feature map X16 (i.e., a first insulator multi-scale optimization feature map) as the final output of the LMRConv module.

[0102] wherein, DWConv (Depthwise Convolution) is a depth separable convolution, which is a lightweight operation that decomposes the standard convolution into channel-wise convolution and point-wise convolution, capable of reducing the computational complexity while retaining the effective feature extraction capability, and is widely used in target detection tasks in mobile terminals or embedded devices.

[0103] The ECA (Efficient Channel Attention) module is a high-efficiency channel attention mechanism that captures the interaction between channels through local one-dimensional convolution, avoids complex fully connected operations, and realizes explicit enhancement of key channel features, while maintaining a relatively low computational burden to improve the model discrimination ability.

[0104] Batch Norm (Batch Normalization) is a normalization technique commonly used in deep learning, which can standardize the feature distribution of each batch of samples during training, helping to alleviate the problems of gradient vanishing and gradient explosion, accelerate model convergence, and improve generalization performance.

[0105] S33, a convolution module is used to perform convolution operation on the first insulator multi-scale optimization feature map to generate a second utility pole insulator convolution feature map;

[0106] S34, a light multi-scale feature enhancement module is used to perform feature enhancement extraction on the second utility pole insulator convolution feature map to output a second insulator multi-scale optimization feature map;

[0107] S35, a convolution module is used to perform convolution operation on the second insulator multi-scale optimization feature map to generate a third utility pole insulator convolution feature map;

[0108] S35, a light multi-scale feature enhancement module is used to perform feature enhancement extraction on the third utility pole insulator convolution feature map to output a third insulator multi-scale optimization feature map.

[0109] It should be noted that F3 is input into the ConvBNSiLU module to generate the insulator feature map F4 (i.e. the second utility pole insulator convolution feature map); F4 is input into the LMRConv module to generate the insulator multi-scale optimization feature map F5 (i.e. the second insulator multi-scale optimization feature map); F5 is input into the ConvBNSiLU module to generate the insulator feature map F6 (i.e. the third utility pole insulator convolution feature map); F6 is input into the LMRConv module to generate the insulator multi-scale optimization feature map F7 (i.e. the third insulator multi-scale optimization feature map).

[0110] It is worth mentioning that the processing process of the second insulator multi-scale optimization feature map and the third insulator multi-scale optimization feature map is consistent with the processing process of the first insulator multi-scale optimization feature map described above, and the present application will not be described in more detail.

[0111] In the embodiment, an insulator image F1 with a size of 512*512*3 is input as a model; the F1 is input into the ConvBNSiLU module to generate an insulator feature map F2 with a size of 256*256*64; the F2 is input into the LMRConv module to generate an insulator multi-scale optimization feature map F3 with a size of 256*256*128; the F3 is input into the ConvBNSiLU module to generate a feature map F4 with a size of 128*128*256; the F4 is input into the LMRConv module to generate an insulator multi-scale optimization feature map F5 with a size of 128*128*256; the F5 is input into the ConvBNSiLU module to generate an insulator feature map F6 with a size of 64*64*256; and the F6 is input into the LMRConv module to generate an insulator multi-scale optimization feature map F7 with a size of 64*64*256.

[0112] In step 104, the target insulator multi-scale optimization feature map is detected by the adaptive feature aggregation unit to generate a power pole insulator defect detection result.

[0113] The adaptive feature aggregation unit includes a dynamic feature reorganization module, a convolution module and a target detection head.

[0114] It should be noted that the insulator multi-scale optimization feature map F3, the insulator multi-scale optimization feature map F5 and the insulator multi-scale optimization feature map F7 of different levels and different scales are input into the adaptive feature aggregation unit to output an image of labeled insulator defect positions, so as to realize the detection of insulator defects and output a prediction result. The adaptive feature aggregation unit has a structure as shown in Figure 5 The adaptive feature aggregation unit is composed of a target detection head, five dynamic feature reorganization modules with the same structure and two convolution modules with the same structure.

[0115] Further, the step 104 can include the following sub-steps:

[0116] In S41, the dynamic feature reorganization module is used to reorganize the third insulator multi-scale optimization feature map to output a first insulator reorganization feature map, and the first insulator reorganization feature map is up-sampled to generate a first insulator up-sampling feature map.

[0117] The dynamic feature reorganization module includes a 3*3 deformable convolution network layer, a reparameterization convolution layer and a 1*1 two-dimensional convolution layer.

[0118] It should be noted that F7 is input into the DFR module to generate an insulator feature map F8 (i.e., a first insulator reorganization feature map); the DFR module strengthens the contextual association and semantic expression ability of the micro target in the insulator defect through dynamic modeling and multi-scale feature fusion mechanism, and improves the detection accuracy and robustness. Among them, the application designs a dynamic feature reorganization module (Dynamic Feature Reorganizer Module, DFR), and the structure of the DFR module is as shown in Figure 6 As shown in the figure, the DFR module enhances the recognition ability of the micro target in the insulator defect detection through multi-scale feature fusion and dynamic attention mechanism. The module adaptively extracts complex boundary features using a deformable convolution and channel compression structure, and introduces a multi-scale attention mechanism to selectively strengthen the defect response area. Through a three-level semantic path, multi-scale context features are extracted, and different levels of semantic information are fused at the end to generate a fusion feature map with high semantic density and concentrated response, significantly improving the sensitivity and robustness of the model to micro defects.

[0119] Further, step S41 can include the following sub-steps:

[0120] S411, using a 3*3 deformable convolution network layer to perform deformable convolution on the third insulator multi-scale optimization feature map to generate an insulator deformable convolution feature map, and performing two-dimensional convolution on the insulator deformable convolution feature map through a 1*1 two-dimensional convolution layer to output a first insulator two-dimensional convolution feature map;

[0121] S412, sequentially performing dynamic channel attention calculation and batch normalization on the first insulator two-dimensional convolution feature map to generate an insulator normalization feature map, and performing reparameterization convolution on the insulator normalization feature map using a reparameterization convolution layer to output a first insulator reparameterization convolution feature map;

[0122] S413, performing maximum pooling on the insulator normalization feature map to output a first insulator maximum pooling feature map, and performing reparameterization convolution on the first insulator maximum pooling feature map using a reparameterization convolution layer to output a second insulator reparameterization convolution feature map;

[0123] S414, performing maximum pooling on the first insulator maximum pooling feature map to output a second insulator maximum pooling feature map, and performing reparameterization convolution on the second insulator maximum pooling feature map using a reparameterization convolution layer to output a third insulator reparameterization convolution feature map;

[0124] S415, performing maximum pooling on the second insulator maximum pooling feature map to output a third insulator maximum pooling feature map, and performing reparameterization convolution on the third insulator maximum pooling feature map using a reparameterization convolution layer to output a fourth insulator reparameterization convolution feature map;

[0125] S416. The reparameterized convolutional feature maps of the first insulator, the second insulator, the third insulator, and the fourth insulator are spliced ​​together to generate the spliced ​​feature map of the target insulator.

[0126] S417. A 1×1 two-dimensional convolutional layer is used to perform two-dimensional convolution on the splicing feature map of the target insulator to generate the first insulator recombination feature map.

[0127] It should be noted that F7 is used as the input feature map of the DFR module. For ease of explanation, F7 is denoted as Y1 here. The input feature map Y1 is fed into the DCNv2 layer (Deformable Convolutional Network v2, 3×3 deformable convolutional network layer) for deformable convolution with a kernel size of 3×3, generating the insulator feature map Y2 (i.e., the deformable convolutional feature map of the insulator). This operation can dynamically adjust the receptive field according to the shape changes of the defect region, effectively improving the adaptability to irregular defect edges. Then, Y2 is fed into a 1×1 two-dimensional convolutional layer for convolution with a kernel size of 1×1, generating the insulator feature map Y3 (i.e., the first two-dimensional convolutional feature map of the insulator). This operation is used to compress the channel dimension and reduce the subsequent computational burden. Next, Y3 is fed into the SK Attention layer (Selective Kernel Attention, dynamic channel attention mechanism) for dynamic channel attention, generating the insulator feature map Y4. This module adaptively selects the optimal receptive field through multi-scale convolutional kernels, enhancing the response strength to small defect regions. Y4 is input into the Batch Norm layer for batch normalization to generate insulator feature map Y5 (i.e., insulator normalized feature map); this normalization operation improves the stability of the training process and optimizes the feature distribution.

[0128] Furthermore, Y5 is input into the MaxPooling layer for max pooling, generating insulator feature map Y6 (i.e., the first insulator max-pooled feature map). This operation expands the receptive field, which is beneficial for capturing contextual information of small targets. Next, Y6 is input into the MaxPooling layer for max pooling, generating insulator feature map Y7 (the second insulator max-pooled feature map). Y7 is then input into the MaxPooling layer for max pooling, generating insulator feature map Y8 (i.e., the third insulator max-pooled feature map).

[0129] Furthermore, Y5 is input into the RepVGG Block layer (Reparameterization VGGBlock) for structural reparameterization convolution, generating insulator feature map Y9 (the first insulator reparameterization convolution feature map). This operation improves expressive power while maintaining inference efficiency through reparameterization. Y6 is input into the RepVGGBlock layer for structural reparameterization convolution, generating insulator feature map Y10 (the second insulator reparameterization convolution feature map). This operation further captures mid-level semantic features, enhancing the ability to model local structural defects. Y7 is input into the RepVGG Block layer for structural reparameterization convolution, generating insulator feature map Y11 (i.e., the third insulator reparameterization convolution feature map). This operation models deeper semantic information, helping to improve the recognition accuracy of minor or ambiguous defects. Y8 is input into the RepVGG Block layer and subjected to structural reparameterization convolution operation to generate insulator feature map Y12 (i.e., the fourth insulator reparameterization convolution feature map); this operation extracts the deepest semantic features and provides global perception capability for feature fusion.

[0130] Then, Y9, Y10, Y11, and Y12 are input into the Concat layer to perform channel-dimensional splicing operations, generating the insulator feature map Y13 (i.e., the spliced ​​feature map of the target insulator). This operation integrates multi-level semantic information and improves the ability to perceive minute defects and weak texture areas.

[0131] Finally, Y13 is input into a 1×1 two-dimensional convolutional layer and convolutional operation with a 1×1 kernel is performed to generate the insulator feature map Y14 (i.e. the first insulator reconstructed feature map) finally output by the DFR module. This operation provides semantically rich and structurally compact feature representations for subsequent defect identification tasks.

[0132] Among them, DCNv2 (Deformable Convolutional Network v2) is an extension of standard convolution. By introducing learnable offsets and modulation mechanisms, it adaptively adjusts the convolution sampling position and weights, effectively improving the model's ability to model target deformation and geometric changes. It is particularly suitable for target detection tasks in complex backgrounds.

[0133] SKAttention (Selective Kernel Attention) is a dynamic channel attention mechanism that can automatically select convolutional kernel responses of different scales based on the contextual information of the input features, thereby achieving adaptive fusion of multi-scale information and enhancing the robustness and adaptability of the model to changes in the target scale.

[0134] RepVGGBlock is a reparameterized convolutional structure that introduces a multi-branch structure during the training phase to enhance feature representation capabilities, while merging the structure into a standard convolutional form during the inference phase through reparameterization. This balances training performance and deployment efficiency, making it suitable for applications in real-world scenarios where inference speed is critical.

[0135] S42. The upsampled feature map of the first insulator and the multi-scale optimized feature map of the second insulator are spliced ​​together to output the spliced ​​feature map of the first insulator. The dynamic feature recombination module is used to recombine the features of the spliced ​​feature map of the first insulator to output the recombined feature map of the second insulator.

[0136] S43. Upsample the recombined feature map of the second insulator to generate an upsampled feature map of the second insulator, and stitch the upsampled feature map of the second insulator with the multi-scale optimized feature map of the first insulator to output the stitched feature map of the second insulator.

[0137] S44. The dynamic feature recombination module is used to recombine the features of the second insulator splicing feature map to generate the third insulator recombination feature map, and the convolution module is used to perform convolution operation on the third insulator recombination feature map to output the first insulator convolution feature map.

[0138] S45. The convolutional feature map of the first insulator and the spliced ​​feature map of the first insulator are spliced ​​together to generate the spliced ​​feature map of the third insulator. The dynamic feature recombination module is used to recombine the features of the spliced ​​feature map of the third insulator and output the recombined feature map of the fourth insulator.

[0139] S46. Use a convolution module to perform convolution operation on the recombined feature map of the fourth insulator, output the convolution feature map of the second insulator, and splice the convolution feature map of the second insulator and the recombined feature map of the first insulator to output the spliced ​​feature map of the fourth insulator.

[0140] S47. Use the dynamic feature recombination module to recombine the splicing feature map of the fourth insulator to generate the recombined feature map of the fifth insulator.

[0141] S48. Input the reconstructed feature map of the fifth insulator, the reconstructed feature map of the fourth insulator, and the reconstructed feature map of the third insulator into the target detection head respectively, and output the defect detection results of the pole insulator.

[0142] It should be noted that F8 is input into the Upsample layer for upsampling, generating insulator feature map F9 (i.e., the first insulator upsampled feature map); F5 and F9 are input into the Concat layer for channel-dimensional concatenation, generating insulator feature map F10 (i.e., the first insulator concatenated feature map); F10 is input into the DFR layer, generating insulator feature map F11 (i.e., the second insulator reconstructed feature map); F11 is input into the Upsample layer for upsampling, generating insulator feature map F12 (i.e., the second insulator upsampled feature map); F3 and F12 are input into the Concat layer for channel-dimensional concatenation, generating insulator feature map F13 (i.e., the second insulator concatenated feature map); F13 is input into the DFR layer, generating insulator feature map F14 (i.e., the third insulator reconstructed feature map). (Figure 1) Input F14 into the ConvBNSiLU module (convolution module) to generate insulator feature map F15 (i.e., the first insulator convolution feature map); Input F10 and F15 into the Concat layer to perform channel-dimensional splicing operation to generate insulator feature map F16 (i.e., the third insulator spliced ​​feature map); Input F16 into the DFR layer to generate insulator feature map F17 (i.e., the fourth insulator recombined feature map); Input F17 into the ConvBNSiLU module to generate insulator feature map F18 (i.e., the second insulator convolution feature map); Input F18 and F8 into the Concat layer to perform channel-dimensional splicing operation to generate insulator feature map F19 (i.e., the fourth insulator spliced ​​feature map); Input F19 into the DFR layer to generate the final insulator feature map F20 (i.e., the fifth insulator recombined feature map).

[0143] Furthermore, F14, F17, and F29 are input into the target detection head, and the output is a prediction tensor containing detection information (i.e., the detection result of insulator defects on the utility pole). Each prediction tensor record corresponds to an insulator defect target, including its bounding box coordinates (i.e., prediction box), defect category, and confidence score. The bounding box is determined by the pixel coordinates of the upper left and lower right corners of the target.

[0144] It is worth mentioning that the processing procedures for the fifth insulator recombination feature map, the fourth insulator recombination feature map, the third insulator recombination feature map, and the second insulator recombination feature map are all consistent with the processing procedures for the first insulator recombination feature map mentioned above, and will not be elaborated further in this invention.

[0145] In this embodiment, F7 is input into the DFR module to generate an insulator feature map F8 with a size of 64×64×256; F8 is input into the Upsample module for upsampling to generate an insulator feature map F9 with a size of 128×128×256; F5 and F9 are concat-joined in the channel dimension to generate an insulator feature map F10 with a size of 128×128×512; F10 is input into the DFR module to generate an insulator feature map F10 with a size of 128×128×512. An 8×128×256 insulator feature map F11 is generated; F11 is input into the Upsample module for upsampling, generating an insulator feature map F12 of size 256×256×256; F3 and F12 are concat-joined to generate an insulator feature map F13 of size 256×256×384; F13 is input into the DFR module to generate an insulator feature map F14 of size 256×256×256; F14 is input into... In the ConvBNSiLU module, an insulator feature map F15 with a size of 128×128×128 is generated; F10 and F15 are then processed in the Concat module for channel dimension concatenation, generating an insulator feature map F16 with a size of 128×128×384; F16 is then input into the DFR module, generating an insulator feature map F17 with a size of 128×128×256; F17 is then input into the ConvBNSiLU module, generating a large... The insulator feature map F18 is 64×64×128. F18 and F8 are input into the Concat module for channel-dimensional concatenation, generating an insulator feature map F19 of size 64×64×384. F19 is input into the DFR module, generating an insulator feature map F20 of size 64×64×256. Finally, F14, F17, and F20 are input into the target detection head, outputting a prediction tensor containing insulator defect identification information. Each row of the prediction tensor records a corresponding insulator defect target, including its bounding box coordinates, defect category, and confidence score. The bounding box is determined by the pixel coordinates of the top-left and bottom-right corners of the target.

[0146] In this embodiment of the invention, a deep learning-based method for detecting defects in utility pole insulators is provided. The method involves acquiring images of utility pole insulators, enhancing these images, and outputting a target image of the utility pole insulator. The target image is then input into an insulator defect detection model, which includes a lightweight multi-scale feature extraction unit and an adaptive feature aggregation unit. The lightweight multi-scale feature extraction unit extracts features from the target image of the utility pole insulator, outputting an optimized multi-scale feature map of the target insulator. The adaptive feature aggregation unit then detects the optimized multi-scale feature map of the target insulator, generating a defect detection result. Based on this approach, the invention improves the quality of the original image through image enhancement preprocessing, laying a clear foundation for subsequent feature extraction. The lightweight multi-scale feature extraction unit can specifically capture key features of the insulator that are easily overlooked due to their small size and detailed structure, avoiding the loss of small target features. The adaptive feature aggregation unit actively suppresses the influence of complex background interference on feature judgment, improving the efficiency of effective feature aggregation and utilization. These three elements work synergistically to significantly improve the accuracy of utility pole insulator defect detection, thereby enhancing the final defect detection effect.

[0147] For better explanation, refer to Figure 7 The flowchart illustrates the steps of training the insulator defect detection model provided in Embodiment 2 of the present invention. This process may include the following steps:

[0148] Step 701: Obtain multiple images of pole insulators for model training, and filter and label these images to output the original dataset for insulator defect detection.

[0149] It should be noted that image samples of insulators in normal, cracked, damaged, and missing states are collected using drone inspection equipment. The data collection process must cover various weather conditions (such as sunny, cloudy, rainy, and foggy days) and different time periods (sunrise, noon, and evening). After image acquisition, invalid images with blurriness, severe occlusion, or high repetition are manually filtered out. Professional annotation tools are then used to accurately label the locations of insulators and their defect types in the images, ultimately constructing the original dataset D1 containing multiple states of insulator defect detection. Commonly used image annotation tools include LabelImg (LabelImage) and LabelME (Label Multi-Environment).

[0150] In this embodiment, 3200 insulator images captured by a drone are processed. These 3200 images consist of 1200 images of insulators in normal condition, 1000 images of cracked condition, 600 images of damaged condition, and 400 images of missing condition. These images cover sunny, cloudy, and rainy / foggy weather conditions, and were captured at sunrise, noon, and dusk. Subsequently, invalid image samples with blurred, severely obscured, or highly repetitive images are removed through manual visual inspection to ensure data quality. Then, the labeling tool LabelImg is used to perform fine-grained annotation of the insulator regions and their defect types in the images, generating XML tag files containing target bounding boxes and category information. Finally, an insulator defect detection dataset D1 containing multiple conditions is constructed.

[0151] Step 702: Perform enhancement operations on the original insulator defect detection dataset to generate an insulator defect detection dataset.

[0152] It should be noted that data enhancement operations such as data rotation, flipping, cropping, contrast adjustment, and noise addition were performed on D1 to generate insulator defect detection dataset D2. Subsequently, D2 was randomly divided into training, validation, and test sets according to a certain ratio to obtain dataset D3 for model training, dataset D4 for model validation, and dataset D5 for model testing, respectively.

[0153] Among them, data rotation: rotating the image by a certain angle to increase the diversity of data.

[0154] Flip: Flip the image horizontally or vertically.

[0155] Cropping: Extracting a portion of an image.

[0156] Contrast Adjustment: Adjusts the contrast of the image to make bright areas brighter and dark areas darker.

[0157] Noise addition technique: Adding random noise to an image.

[0158] In this embodiment, the original insulator defect detection dataset D1 contains 3200 images. First, all images are uniformly cropped to 512×512 pixels. Then, various data augmentation techniques are used to expand the dataset, including data rotation, flipping, cropping, contrast adjustment, and noise addition, all of which can be implemented using Python. Specifically, rotation is performed using the `skimage.transform.rotate()` function; image flipping uses the `numpy.flip()` function; cropping is performed using the `PIL.Image.crop()` method; contrast adjustment is achieved using `PIL.ImageEnhance.Contrast()`; and noise addition is performed using a custom salt-and-pepper noise generation function. After these augmentation processes, dataset D1 is expanded into an insulator defect detection dataset D2 containing 12800 insulator images. Subsequently, D2 is randomly divided into training, validation, and test data in a 6:2:2 ratio, resulting in dataset D3 for training the model, dataset D4 for validating the model's performance, and dataset D5 for testing the model's performance. D3 contains 7680 insulator images, D4 contains 2560 insulator images, and D5 contains 2560 insulator images.

[0159] Step 703: Use a preset loss function to train the initial insulator defect detection model based on the insulator defect detection dataset, and determine the trained insulator defect detection model.

[0160] It should be noted that the initial insulator defect detection model is constructed, which consists of a lightweight multi-scale feature extraction unit and an adaptive feature aggregation unit, wherein:

[0161] Lightweight multi-scale feature extraction unit: Input the insulator image F1∈D2 into the ConvBNSiLU module to generate insulator feature map F2; input F2 into the LMRConv module to generate insulator feature map F3; input F3 into the ConvBNSiLU module to generate insulator feature map F4; input F3 into the LMRConv module to generate insulator feature map F5; input F5 into the ConvBNSiLU module to generate insulator feature map F6; input F6 into the LMRConv module to generate insulator feature map F7.

[0162] Adaptive Feature Aggregation Unit: Input F7 into the DFR module to generate insulator feature map F8; input F8 into the Upsample module for upsampling to generate insulator feature map F9; input F5 and F9 into the Concat module for channel-dimensional concatenation to generate insulator feature map F10; input F10 into the DFR module to generate insulator feature map F11; input F11 into the Upsample module for upsampling to generate insulator feature map F12; input F3 and F12 into the Concat module for channel-dimensional concatenation to generate insulator feature map F13; input F13 into... The insulator feature map F14 is generated by inputting F14 into the DFR module. F14 is then input into the ConvBNSiLU module to generate F15. F10 and F15 are input into the Concat module for channel-dimensional concatenation to generate F16. F16 is input into the DFR module to generate F17. F17 is input into the ConvBNSiLU module to generate F18. F18 and F8 are input into the Concat module for channel-dimensional concatenation to generate F19. F19 is input into the DFR module to generate F20. F14, F17, and F20 are then input into the target detection head (Head), outputting a prediction tensor containing detection information. Each line of the prediction tensor records a corresponding insulator defect target, including its bounding box coordinates, defect category, and confidence score. The bounding box is determined by the pixel coordinates of the target's upper left and lower right corners.

[0163] Furthermore, based on the constructed enhanced insulator defect detection dataset, the designed defect detection model is trained and optimized. Before training begins, the initial parameters of the network structure are preset, and relevant hyperparameters are configured, including the number of training epochs, batch size, optimization algorithm type, and initial learning rate.

[0164] During the training phase, the training and validation sets are input into the model in batches. After each batch of data undergoes forward propagation, the IAM-IoU loss function (i.e., the preset loss function) is used to measure the difference between the model's predictions and the labeled results. The model weights are continuously updated through backpropagation to improve the model's detection performance. Simultaneously, the validation set is used to dynamically evaluate the model's generalization ability during training, helping to determine whether overfitting has occurred and adjusting the training strategy accordingly.

[0165] Once the loss curve on the validation set stabilizes and converges, the training process ends, resulting in an insulator defect detection model with good learning performance. Subsequently, test set samples are input into the trained model for final evaluation, analyzing its accuracy and robustness in practical defect detection tasks to verify the model's reliability in real-world application scenarios.

[0166] It is worth mentioning that this invention proposes an instance-aware modulated IoU (IAM-IoU) loss function suitable for insulator defect detection tasks, aiming to alleviate the training bias problem caused by extreme imbalance between positive and negative samples. This loss function, based on regression only on positive samples, introduces an area reciprocal modulation and confidence weighting mechanism, thereby enhancing the focus on small-scale, high-confidence defect targets. By dynamically adjusting the gradient response weights, IAM-IoU accelerates model convergence while effectively enhancing the localization accuracy and detection sensitivity of insulator defect regions, providing a more targeted optimization method for high-performance defect identification.

[0167] To more effectively enhance regression performance for small targets and high-confidence defect samples, this invention constructs the IAM-IoU layer loss function in the following form:

[0168] ;

[0169] in, The loss value corresponding to the preset loss function; This represents the number of positive samples in the current batch. Let be the intersection-union ratio (IoU) between the predicted bounding box and the ground truth bounding box of the i-th positive sample. This is a focus modulation factor used to enhance the model's attention to difficult samples with low IoU. This is the weight term corresponding to the i-th positive sample, applied to each positive sample instance to dynamically adjust its gradient contribution to the total loss. This weight is calculated jointly based on the IoU value between the predicted and ground truth bounding boxes and the target area. When the target area is small or the prediction confidence is high, this weight is adjusted accordingly. The larger gradient weights allow small targets and prediction samples with higher confidence to receive greater gradient weights during parameter updates, thus enhancing their optimization priority during model training. Compared to traditional localization losses that rely solely on IoU calculations, IAM-IoU significantly enhances the model's ability to perceive small-scale defects and targets with blurred boundaries. Let be the area of ​​the ground truth bounding box of the i-th positive sample; To prevent division by zero of small constants; To control the degree to which the model enhances small targets; The confidence score of the i-th predicted bounding box (output of the sigmoid layer); This is used to adjust the sensitivity of the confidence weight.

[0170] Furthermore, by assigning significant weights to different positive samples, the IAM-IoU layer indirectly suppresses the dominant influence of a large number of negative samples on the gradient without directly introducing negative sample regression, effectively alleviating the training bias problem caused by the extreme imbalance in the number of positive and negative samples. In the early stages of training, the model can quickly focus on feature learning from effective positive samples, improving convergence speed; in the later stages, it further optimizes the response strength to difficult-to-detect defects, achieving more accurate localization performance.

[0171] The IAM-IoU loss function introduces a saliency modulation mechanism that integrates target area and prediction confidence without increasing model structure complexity and computational overhead. It constructs a dynamic weighting strategy for key samples, which effectively improves the model's ability to distinguish small-scale and weakly textured defect targets, while enhancing the overall detection stability and robustness.

[0172] Finally, the trained insulator defect detection model is deployed to the transmission line defect inspection system to process insulator inspection images in real time, ultimately outputting the defect type, bounding box coordinates, and confidence score. This provides inspection personnel with accurate early warning information and a basis for maintenance decisions.

[0173] For example, a defect detection experiment was conducted on insulator assemblies operating within a transmission line area, with the insulator body being the detection target. The corresponding insulator images acquired by the equipment are shown below. Figure 8 As shown, the insulator image is used as input to the insulator defect detection model, and the output detection result is as follows. Figure 9 As shown, the output detection results include: 1. Insulator defect category: Damaged. 2. Bounding box coordinates: (115, 170, 400, 345). 3. Confidence score: 0.92.

[0174] For comparison of technical effectiveness, existing technologies can be referenced. In recent years, deep learning-based image recognition technology has made significant progress in intelligent inspection of power facilities, especially in the automatic defect identification using UAV images, showing broad prospects. However, existing methods still face three technical bottlenecks: First, insulator targets are small in size and have intricate structures, making them susceptible to background interference, resulting in insufficient detection accuracy; second, the number of defect samples in actual acquired images is far less than the number of normal samples, and the distribution of positive and negative samples is extremely unbalanced, making traditional training methods prone to model bias; third, existing detection models have high computational requirements, and performance bottlenecks exist when deployed on resource-constrained edge devices (such as UAVs and front-end gateways), making it difficult to achieve lightweight deployment and real-time identification.

[0175] Based on the above, current insulator defect detection methods have the following shortcomings:

[0176] (1) Existing detection methods mostly rely on manual inspection or high-cost equipment, which have problems such as long detection cycles, limited coverage, and slow response, making it difficult to meet the real-time and automation requirements of complex power grid environments. Some deep learning methods have high computational complexity and are not suitable for deployment on edge devices, which limits the flexibility of engineering applications.

[0177] (2) Insulators are typical small-sized targets that are easily affected by occlusion, blurring and background interference. Existing models are insufficient in terms of small target feature extraction and expression, making it difficult to effectively identify minor defects such as cracks and damage, resulting in low detection accuracy.

[0178] (3) The proportion of defective samples in the inspection images is small, and the positive and negative samples are severely unbalanced. Traditional training strategies are prone to causing the model to overfit to the majority class, and the recognition results are biased towards the normal state, making it difficult to guarantee the recognition recall rate and model generalization ability of defective insulator samples.

[0179] Therefore, there is an urgent need for an insulator defect detection technology that balances detection accuracy and computational efficiency, has good small target recognition capabilities, and effectively mitigates the impact of positive and negative sample imbalance, in order to meet the needs of efficient inspection and intelligent operation and maintenance of power grid equipment in complex environments and promote the level of power system safety management to a higher level.

[0180] To address the aforementioned issues, this invention proposes a deep learning-based method for detecting defects in power pole insulators. This method aims to solve key problems in current insulator defect detection technologies, such as high model redundancy, insufficient ability to identify small-sized defects, and uneven distribution of positive and negative samples. The goal is to improve the efficiency and accuracy of detecting typical defects in power systems, such as cracks, damage, and missing parts, and to meet the intelligent inspection needs in complex scenarios.

[0181] This invention addresses the challenges of insulator defect detection, such as small target size, weak features, and numerous interferences. It proposes a lightweight deep learning model and designs and constructs three key modules: a lightweight multi-branch residual module (LMRM), a multi-scale detail-fusion reconstruction module (DFR), and an instance-aware localization loss function (IAM-IoU). These modules collaboratively optimize the overall detection performance from three aspects: model lightweighting, feature fusion, and training mechanism.

[0182] To address the issues of existing insulator defect detection models being complex in structure and consuming large computational resources, making them unsuitable for real-time deployment by UAVs or edge devices, this paper designs an LMRM module. This module employs a multi-scale parallel convolutional structure, integrating 1×1, 3×3, and 5×5 convolutional kernels. It dynamically adjusts feature weights using a channel attention mechanism and compresses redundant channels through residual connections, effectively improving the perception capability of local defect regions. This achieves a balance between lightweight feature extraction and efficient information utilization, significantly enhancing the model's applicability and inference speed in resource-constrained environments.

[0183] To address the issue of insufficient feature fusion and significant information loss due to the small size and blurred edge information of small targets in insulator defect detection, which in turn affects detection accuracy, this paper designs a DFR module. It adopts a feature reconstruction structure that combines low-level detail guidance with high-level semantic fusion, and combines upsampling and cross-layer residual enhancement mechanisms to strengthen the alignment and information interaction of multi-scale features. This improves the model's ability to identify fine-grained defects such as micro-cracks and edge damage, and significantly improves the robustness and accuracy of small target detection in complex environments.

[0184] To address the problem of extreme imbalance between positive and negative samples in insulator defect detection tasks, which leads to overfitting of the majority class and insufficient defect recognition ability of the minority class during model training, this paper proposes an IAM-IoU loss function. Based on regression only on positive samples, it introduces the inverse area weight and confidence modulation mechanism, enabling the model to provide a larger gradient response to small targets and high-confidence defect samples during training. This effectively alleviates training bias, improves model convergence speed and localization accuracy, and enhances the detection effect of defect samples.

[0185] Specifically, please refer to Figure 10 First, images from multiple scenes are collected and the locations and types of defects in insulators are labeled to construct an insulator defect detection dataset. Then, the insulator images are input into a lightweight multi-scale feature extraction unit to extract multi-scale optimized feature maps of the insulators. Next, the multi-level optimized insulator feature maps are input into an adaptive feature aggregation unit to complete the detection of defect regions. Subsequently, the model is optimized and trained using the training set. Finally, the model is deployed in an actual inspection scenario to output detection result images containing defect locations and categories.

[0186] In summary, this invention designs a lightweight multi-receptive convolution module (LMRConv) to enhance the multi-scale feature extraction capability of insulator defects while controlling the number of parameters. This module first performs channel compression using 1×1 convolutions, then splits into two parallel branches: the first branch uses consecutive 3×3 depthwise separable convolutions (DWConv) to extract local detail features, while the second branch uses 5×5 depthwise separable convolutions to enhance the receptive field and capture larger-scale texture information. Each branch incorporates ReLU6 activation, BatchNorm normalization, and ECA channel attention mechanisms to highlight significant channel responses in the defect region. The output features of each branch are fused element-wise and then concatenated with the input residual. Feature fusion and dimensionality restoration are then performed sequentially using 3×3 and 1×1 convolutions, ultimately outputting a refined feature map in a unified semantic space.

[0187] Simultaneously, this invention proposes a Dynamic Feature Reorganizer (DFR) module to address the challenge of perceiving small-target insulator defects in complex scenarios, integrating deformable convolution, channel attention, and multi-layer semantic enhancement paths. First, basic features are extracted using DCNv2 to improve the model's adaptability to irregular defect boundaries and geometric deformations. Then, the SK attention module dynamically selects key channel information within multi-scale receptive fields to suppress redundant background interference. A RepVGG Block is introduced to construct multi-scale semantic paths, combined with multi-level max pooling to extract semantic features at different scales, and finally concatenated and fused at the end to form a highly expressive feature map in a unified semantic space, thereby significantly enhancing multi-layer semantic understanding and structural perception of small-target defect regions.

[0188] Furthermore, this invention designs a saliency modulation regression loss function, IAM-IoU (Instance-Aware Modulated IoU), to alleviate the problem of extremely uneven distribution of positive and negative samples and severe training bias in insulator defect detection. Based on standard IoU regression, this loss function introduces an inverse area weight and a confidence-based modulation factor to construct an instance saliency weight term, enabling the model to automatically enhance its gradient response to small targets and high-confidence prediction samples during training. IAM-IoU provides a clear optimization direction when the predicted boxes are completely misaligned, and finely penalizes position and scale errors when partial overlap exists, thereby achieving more stable boundary fitting. This mechanism establishes a sample weighting strategy guided by both "area + confidence" without introducing additional network structures or complex computations, effectively controlling model complexity while improving attention to difficult-to-detect targets and overall detection robustness.

[0189] In this embodiment of the invention, a deep learning-based method for detecting defects in utility pole insulators is proposed, aiming to address the problems of inaccurate identification of small-sized defects, low feature extraction efficiency, and imbalanced training samples in existing technologies. By optimizing the feature modeling method and detection mechanism, the method significantly improves the ability to identify defects such as cracks, breaks, and missing parts in insulators, enhancing the model's adaptability and deployment efficiency in complex inspection environments. Specifically, this invention designs a lightweight multi-scale convolution module, LMRConv (Lightweight Multi-Receptive Convolution Module), specifically designed for the efficient extraction of fine-grained features in insulator defect detection tasks. This module adopts a dual-branch structure, extracting texture and morphological information at different scales through parallel 3×3 and 5×5 depth-separable convolutional paths, enhancing the model's ability to perceive the morphology of defects such as micro-cracks and breaks. Within each branch, the ReLU6 activation function and ECA channel attention mechanism are combined to achieve adaptive weighting of salient regions, thereby highlighting the response of defect areas. This module enhances semantic consistency through element-wise addition and fusion, and introduces 3×3 and 1×1 convolutions to achieve feature fusion and channel compression, ultimately outputting a highly expressive unified feature map. While maintaining a lightweight computational structure, this module effectively improves the model's feature modeling capabilities under complex backgrounds and varied defect morphologies. Simultaneously, this invention proposes a Dynamic Feature Reorganizer (DFR) module to address the challenges of perceiving small targets and losing weak textures in insulator defect detection. It constructs a feature processing path that integrates multi-scale attention and deformable receptive fields. This module achieves geometrically adaptive feature extraction based on 3×3 deformable convolutions, introduces a Selective Kernel attention mechanism to dynamically weight responses from different channels, and strings multiple RepVGG Blocks in the main path, combined with multi-level max pooling for semantic expansion. By concatenating multiple feature layers, a multi-semantic representation with both shallow and deep features is constructed, effectively preserving local defect information and global structural semantics. This module enhances the model's ability to capture small defect regions and the robustness of its overall representation, making it particularly suitable for the fine-grained detection requirements under complex working conditions. Furthermore, this invention designs an IAM-IoU loss function (Instance-Aware Modulated IoU Loss) to alleviate the training bias problem caused by the extreme imbalance between positive and negative samples in insulator defect detection. Based on IoU regression, this loss function introduces saliency modulation weights and a focus adjustment mechanism to enhance the gradient contribution of small targets and high-confidence positive samples. Area-based scaling is used to improve the model's response sensitivity to small-sized defect regions, while the confidence modulation term dynamically adjusts the learning focus, guiding the training process to focus on high-quality prediction boxes.This mechanism effectively improves the accuracy of small target recognition and the stability and robustness of the overall model training without requiring additional structures or parameters.

[0190] Please see Figure 11 , Figure 11 This is a structural block diagram of a deep learning-based defect detection device for utility pole insulators, provided in Embodiment 3 of the present invention.

[0191] This invention provides a deep learning-based defect detection device for utility pole insulators, comprising:

[0192] The acquisition module 1101 is used to acquire images of pole insulators, perform enhancement operations on the pole insulator images, and output the target pole insulator image.

[0193] Input module 1102 is used to input the target pole insulator image into the insulator defect detection model, which includes a lightweight multi-scale feature extraction unit and an adaptive feature aggregation unit.

[0194] Feature extraction module 1103 is used to extract features from the target pole insulator image using a lightweight multi-scale feature extraction unit and output the target insulator multi-scale optimized feature map;

[0195] The detection module 1104 is used to detect the multi-scale optimized feature map of the target insulator through the adaptive feature aggregation unit and generate the defect detection results of the pole insulator.

[0196] In one optional device embodiment, it further includes:

[0197] The first module is used to acquire multiple images of utility pole insulators for model training, and to filter and label these images to output the original dataset for insulator defect detection.

[0198] The second module is used to perform enhancement operations on the original insulator defect detection dataset to generate an insulator defect detection dataset.

[0199] The third module is used to train the initial insulator defect detection model based on the insulator defect detection dataset using a preset loss function, and to determine the trained insulator defect detection model.

[0200] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the above-described device and module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0201] This invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program; when the computer program is executed by the processor, the processor performs the steps of the deep learning-based method for detecting defects in utility pole insulators as described in any of the above embodiments.

[0202] This invention also provides a computer-readable storage medium storing a computer program / instructions thereon, which, when executed by a processor, implements the steps of the deep learning-based method for detecting defects in utility pole insulators as described in any of the above embodiments.

[0203] This invention also provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the deep learning-based method for detecting defects in utility pole insulators as described in any of the above embodiments.

[0204] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting defects in utility pole insulators based on deep learning, characterized in that, include: Acquire an image of a utility pole insulator, perform enhancement operations on the image, and output a target utility pole insulator image; The image of the target utility pole insulator is input into the insulator defect detection model, which includes a lightweight multi-scale feature extraction unit and an adaptive feature aggregation unit. The lightweight multi-scale feature extraction unit is used to extract features from the target pole insulator image and output the target insulator multi-scale optimized feature map. The adaptive feature aggregation unit is used to detect the multi-scale optimized feature map of the target insulator, generating defect detection results for the pole insulator.

2. The method for detecting defects in utility pole insulators based on deep learning according to claim 1, characterized in that, The lightweight multi-scale feature extraction unit includes a convolution module and a lightweight multi-scale feature enhancement module; the target insulator multi-scale optimized feature map includes a first insulator multi-scale optimized feature map, a second insulator multi-scale optimized feature map, and a third insulator multi-scale optimized feature map; The step of using the lightweight multi-scale feature extraction unit to extract features from the target pole insulator image and outputting a multi-scale optimized feature map of the target insulator includes: A convolution module is used to perform convolution operations on the target pole insulator image to generate a first pole insulator convolution feature map; The convolutional feature map of the first pole insulator is enhanced and extracted using a lightweight multi-scale feature enhancement module, and the multi-scale optimized feature map of the first insulator is output. A convolution module is used to perform convolution operations on the multi-scale optimized feature map of the first insulator to generate a convolution feature map of the second pole insulator. The convolutional feature map of the second pole insulator is enhanced and extracted using a lightweight multi-scale feature enhancement module, and the multi-scale optimized feature map of the second insulator is output. A convolution module is used to perform convolution operations on the multi-scale optimized feature map of the second insulator to generate a convolution feature map of the third pole insulator. The convolutional feature map of the third pole insulator is enhanced and extracted using a lightweight multi-scale feature enhancement module, and the multi-scale optimized feature map of the third insulator is output.

3. The method for detecting defects in utility pole insulators based on deep learning according to claim 2, characterized in that, The lightweight multi-scale feature enhancement module includes a 1×1 two-dimensional convolutional layer, a 3×3 two-dimensional convolutional layer, a 3×3 depthwise convolutional layer, a 5×5 depthwise convolutional layer, and a ReLU6 activation function layer. The feature enhancement and extraction of the convolutional feature map of the first utility pole insulator using the lightweight multi-scale feature enhancement module, outputting a multi-scale optimized feature map of the first insulator, includes: A 1×1 two-dimensional convolutional layer is used to perform a two-dimensional convolution operation on the convolutional feature map of the first power pole insulator, and the two-dimensional convolutional feature map of the first power pole insulator is output. The two-dimensional convolutional feature map of the first pole insulator is input into a 3×3 depth convolutional layer and a 5×5 depth convolutional layer respectively for depth convolution, and the depth convolutional feature map of the first pole insulator and the depth convolutional feature map of the second pole insulator are output. The first pole insulator depth convolution feature map is used as the input of the ReLU6 activation function layer to output the first pole insulator transformation feature map. The first pole insulator transformation feature map and the second pole insulator depth convolution feature map are added element by element to output the first pole insulator addition feature map. A 3×3 depth convolution layer is used to perform depth convolution on the transformed feature map of the first pole insulator to output the depth convolution feature map of the third pole insulator. The first pole insulator summation feature map is used as the input to the ReLU6 activation function layer to generate the second pole insulator transformation feature map; Batch normalization and channel attention calculation are performed on the depth convolution feature map of the third pole insulator and the transformation feature map of the second pole insulator respectively, and the channel feature maps of the first pole insulator and the second pole insulator are output. The first pole insulator channel feature map and the second pole insulator channel feature map are added element by element to output the second pole insulator added feature map. The second pole insulator added feature map is then convolved in two dimensions through a 3×3 two-dimensional convolutional layer to output the second pole insulator two-dimensional convolutional feature map. The two-dimensional convolutional feature map of the second pole insulator, the summed feature map of the second pole insulator, and the summed feature map of the first pole insulator are spliced ​​together to output a spliced ​​feature map of the pole insulator. The spliced ​​feature map of the pole insulator is then subjected to two-dimensional convolution through a 1×1 two-dimensional convolutional layer to generate a multi-scale optimized feature map of the first insulator.

4. The method for detecting defects in utility pole insulators based on deep learning according to claim 2, characterized in that, The adaptive feature aggregation unit includes a dynamic feature recombination module, a convolution module, and a target detection head; The step of inputting the multi-scale optimized feature map of the target insulator into the adaptive feature aggregation unit for detection, and generating defect detection results for the pole insulator, includes: The dynamic feature recombination module is used to reconstruct the multi-scale optimized feature map of the third insulator, outputting the reconstructed feature map of the first insulator, and upsampling the reconstructed feature map of the first insulator to generate the upsampled feature map of the first insulator. The upsampled feature map of the first insulator and the multi-scale optimized feature map of the second insulator are spliced ​​together to output the spliced ​​feature map of the first insulator. The dynamic feature recombining module is then used to recombine the features of the spliced ​​feature map of the first insulator to output the recombined feature map of the second insulator. The reconstructed feature map of the second insulator is upsampled to generate an upsampled feature map of the second insulator. The upsampled feature map of the second insulator and the multi-scale optimized feature map of the first insulator are then spliced ​​together to output a spliced ​​feature map of the second insulator. The second insulator splicing feature map is recombined using a dynamic feature recombination module to generate a third insulator recombined feature map, and the third insulator recombined feature map is convolved using a convolution module to output the first insulator convolution feature map. The first insulator convolution feature map and the first insulator splicing feature map are spliced ​​together to generate the third insulator splicing feature map. The third insulator splicing feature map is then recombined using a dynamic feature recombining module to output the fourth insulator recombined feature map. A convolution module is used to perform convolution operation on the recombined feature map of the fourth insulator, outputting a convolution feature map of the second insulator. The convolution feature map of the second insulator and the recombined feature map of the first insulator are then spliced ​​together to output a spliced ​​feature map of the fourth insulator. The fourth insulator splicing feature map is recombined using a dynamic feature recombination module to generate the fifth insulator recombined feature map. The fifth insulator recombination feature map, the fourth insulator recombination feature map, and the third insulator recombination feature map are respectively input into the target detection head, and the defect detection results of the pole insulator are output.

5. The method for detecting defects in utility pole insulators based on deep learning according to claim 4, characterized in that, The dynamic feature reconstruction module includes a 3×3 deformable convolutional network layer, a reparameterized convolutional layer, and a 1×1 two-dimensional convolutional layer; the dynamic feature reconstruction module is used to reconstruct the multi-scale optimized feature map of the third insulator and output the reconstructed feature map of the first insulator, including: A 3×3 deformable convolutional network layer is used to perform deformable convolution on the multi-scale optimized feature map of the third insulator to generate a deformable convolutional feature map of the insulator. Then, a 1×1 two-dimensional convolutional layer is used to perform two-dimensional convolution on the deformable convolutional feature map of the insulator to output a two-dimensional convolutional feature map of the first insulator. Dynamic channel attention calculation and batch normalization are performed sequentially on the two-dimensional convolutional feature map of the first insulator to generate a normalized feature map of the insulator. Then, a reparameterized convolutional layer is used to perform reparameterized convolution on the normalized feature map of the insulator to output the reparameterized convolutional feature map of the first insulator. Max pooling is performed on the normalized feature map of the insulator to output the first max pooling feature map of the insulator. Then, a reparameterized convolutional layer is used to perform reparameterized convolution on the first max pooling feature map of the insulator to output the second reparameterized convolutional feature map of the insulator. Max pooling is performed on the first insulator max pooling feature map to output the second insulator max pooling feature map, and a reparameterized convolutional layer is used to perform reparameterized convolution on the second insulator max pooling feature map to output the third insulator reparameterized convolutional feature map. Max pooling is performed on the second insulator max pooling feature map to output the third insulator max pooling feature map, and a reparameterized convolutional layer is used to perform reparameterized convolution on the third insulator max pooling feature map to output the fourth insulator reparameterized convolutional feature map. The reparameterized convolutional feature maps of the first insulator, the second insulator, the third insulator, and the fourth insulator are spliced ​​together to generate a spliced ​​feature map of the target insulator. A 1×1 two-dimensional convolutional layer is used to perform two-dimensional convolution on the splicing feature map of the target insulator to generate the first insulator recombination feature map.

6. The method for detecting defects in utility pole insulators based on deep learning according to claim 1, characterized in that, The model training process for the insulator defect detection model is as follows: Multiple images of utility pole insulators for model training are obtained, and these images are filtered and labeled to output the original dataset for insulator defect detection. An enhancement operation is performed on the original insulator defect detection dataset to generate an insulator defect detection dataset; The initial insulator defect detection model is trained using a preset loss function based on the insulator defect detection dataset to determine the trained insulator defect detection model.

7. A deep learning-based defect detection device for utility pole insulators, characterized in that, include: The acquisition module is used to acquire images of utility pole insulators, perform enhancement operations on the images of utility pole insulators, and output target images of utility pole insulators. The input module is used to input the target pole insulator image into the insulator defect detection model, which includes a lightweight multi-scale feature extraction unit and an adaptive feature aggregation unit. The feature extraction module is used to extract features from the target pole insulator image using the lightweight multi-scale feature extraction unit and output a multi-scale optimized feature map of the target insulator. The detection module is used to detect the multi-scale optimized feature map of the target insulator through the adaptive feature aggregation unit, and generate the defect detection result of the pole insulator.

8. A computer device, characterized in that, The device includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the deep learning-based method for detecting defects in pole insulators as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed, it implements the deep learning-based method for detecting defects in utility pole insulators as described in any one of claims 1-6.

10. A computer program product, characterized in that, The computer program product includes a computer program stored on a non-transitory computer-readable storage medium, the computer program including program instructions, wherein when the program instructions are executed by a computer, the computer performs the deep learning-based method for detecting defects in pole insulators as described in any one of claims 1-6.

Images