Method and device for marking the degree of contamination of a power transmission component

Abstract

This application relates to a method and apparatus for labeling the degree of contamination of power transmission components. The method includes: acquiring inspection images of power transmission lines; sequentially performing noise suppression, illumination equalization correction, and standardization processing on the inspection images to obtain a standardized sample set to be labeled; inputting the sample set to be labeled into a trained hybrid detection model to output a preliminary detection result containing a first score of cognitive uncertainty; filtering and correcting the preliminary detection result based on the first score to obtain a contamination labeling result for the power transmission components; classifying the degree of contamination in the inspection images within the contamination labeling result to obtain a refined labeling result containing a second score of cognitive uncertainty; and filtering and correcting the refined labeling result based on the second score to obtain the final contamination labeling result for the power transmission components. This method can improve the accuracy of contamination labeling of power transmission components.

Description

Technical Field

[0001] This application relates to the technical field of power artificial intelligence, and in particular to a method and apparatus for labeling the degree of contamination of power transmission components. Background Technology

[0002] Transmission lines are exposed to the outdoor environment for extended periods, and components such as insulators and fittings are prone to accumulating dirt and grime. Accumulation of this dirt and grime can trigger flashover faults, causing widespread power outages. Therefore, accurately assessing the level of contamination in transmission components is a crucial aspect of line inspection.

[0003] In related technologies, existing pollution detection techniques for power transmission components have evolved from manual inspection and traditional image processing to deep learning. Although detection models based on convolutional neural networks (CNNs) are widely used, their local receptive field limits their ability to capture global contextual features, resulting in limited accuracy in scenarios with dispersed pollution distribution and blurred edges. Detection models based on Transformers can capture global information, but their ability to capture local details is relatively weak, and they require a large amount of training data, leading to inaccurate final annotation results. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, apparatus, computer equipment, computer-readable storage medium, and computer program product for labeling the degree of contamination of power transmission components, which can improve the accuracy of contamination labeling, in response to the above-mentioned technical problems.

[0005] Firstly, this application provides a method for labeling the degree of contamination of power transmission components, including:

[0006] Obtain inspection images of transmission lines, and sequentially perform noise suppression, illumination equalization correction, and standardization processing on the inspection images to obtain a standardized sample set to be labeled;

[0007] The sample set to be labeled is input into the trained hybrid detection model, and the preliminary detection result containing the first score of cognitive uncertainty is output.

[0008] Based on the first score, the preliminary detection results are filtered and corrected to obtain the pollution labeling results of the power transmission components;

[0009] The degree of filth in the inspection images in the filth labeling results is graded and labeled to obtain a refined labeling result that includes a second score of cognitive uncertainty;

[0010] Based on the second score, the fine annotation results are filtered and corrected to obtain the final pollution annotation results for the power transmission components.

[0011] In one embodiment, the hybrid detection model includes a feature extraction network, a hybrid encoder, and a preset refiner; the step of inputting the sample set to be labeled into the trained hybrid detection model and outputting a preliminary detection result containing a first score of cognitive uncertainty includes:

[0012] The sample set to be labeled is input into the feature extraction network to extract a multi-level feature set of the sample set to be labeled;

[0013] Determine the target features of a preset number of layers in the multi-level feature set, input the target features into the hybrid encoder, perform feature encoding on the target features, and obtain a hybrid encoded feature map;

[0014] The hybrid encoded feature map is input into a preset refiner, and the output includes a first score of cognitive uncertainty, the component category and boundary of the power transmission component, the first pollution category and first prediction box of the polluted area, and a preliminary detection result with a first confidence level.

[0015] In one embodiment, the step of sequentially performing noise suppression, illumination equalization correction, and standardization processing on the inspection image to obtain a standardized sample set to be labeled includes:

[0016] Gaussian filtering and median filtering are applied sequentially to the inspection image to obtain a denoised inspection image;

[0017] The denoised inspection image is then subjected to illumination equalization correction to obtain the corrected inspection image.

[0018] The corrected inspection image is divided into a preset number of sub-images according to a preset method. The sub-images are converted into a preset format and then scaled to a preset size to obtain a standardized sample set to be labeled.

[0019] In one embodiment, the step of filtering and correcting the preliminary detection results based on the first score to obtain the contamination labeling results of the power transmission components includes:

[0020] If, in the preliminary detection results, there is a detection result whose first score is greater than a preset uncertainty threshold, the corresponding detection result is deleted to obtain the preliminary detection results after deletion.

[0021] In the preliminary detection results after deletion, the detection results with a first confidence level greater than the preset confidence level are selected as the preliminary detection results after filtering;

[0022] The image region corresponding to the first prediction box of each detection result in the preliminary detection results after screening is determined, the target region in the image region that meets the preset extraction conditions is extracted, and a reflective mask is generated based on the target region;

[0023] Calculate the regional overlap between the reflective mask and the first prediction box, and delete the detection results with regional overlap greater than the first dynamic threshold from the preliminary detection results after filtering to obtain the preliminary detection results after filtering.

[0024] The initial test results after filtration are pre-corrected using pollution logic to obtain the pollution labeling results for the power transmission components.

[0025] In one embodiment, the first contamination category includes local contamination, global contamination, and no contamination; the step of performing contamination logic pre-correction on the preliminary detection results after filtering to obtain the contamination labeling results of the power transmission components includes:

[0026] From the preliminary test results after filtering, the test results of power transmission components with symmetrical structures and localized pollution were selected as the first set of test results;

[0027] Determine the axis of symmetry of the power transmission component in the first set of detection results, and based on the axis of symmetry, determine the mirror region corresponding to the target contamination region in the first set of detection results;

[0028] Calculate the feature similarity between the mirrored region and the target contaminated region;

[0029] Based on the first score, determine the dynamic upgrade threshold;

[0030] If the feature similarity is greater than the dynamic upgrade threshold, the dirt category corresponding to the predicted box will be modified to global dirt.

[0031] In one embodiment, the refined annotation result includes a second contamination category and a second predicted bounding box for the contaminated area, a second confidence level, and a second score; the filtering and correction of the annotation result based on the second score to obtain the final contamination annotation result for the power transmission component includes:

[0032] If, in the fine annotation results, there is a second score corresponding to a detection result that is greater than a preset uncertainty threshold, the corresponding detection result is deleted to obtain the fine detection result after deletion.

[0033] In the refined detection results after deletion, the detection results with a second confidence level greater than the preset confidence level are selected as the refined detection results after filtering;

[0034] Determine the image region corresponding to the second prediction box of each detection result in the filtered fine detection results, extract the first target region in the image region that meets the preset extraction conditions, and generate a reflective mask based on the first target region;

[0035] Calculate the regional overlap between the reflective mask and the second prediction box, and delete the detection results with regional overlap greater than the second dynamic threshold from the filtered fine detection results to obtain the filtered fine detection results.

[0036] The filtered fine test results are corrected for contamination logic to obtain the final contamination labeling results for the power transmission components.

[0037] Secondly, this application also provides a device for marking the degree of contamination of power transmission components, comprising:

[0038] The processing module is used to acquire inspection images of transmission lines, and sequentially perform noise suppression, illumination equalization correction and standardization processing on the inspection images to obtain a standardized sample set to be labeled.

[0039] The output module is used to input the sample set to be labeled into the trained hybrid detection model and output a preliminary detection result containing a first score of cognitive uncertainty;

[0040] The correction module is used to filter and correct the preliminary detection results based on the first score to obtain the pollution labeling results of the power transmission components;

[0041] The annotation module is used to classify and annotate the degree of dirtiness in the inspection images in the dirt annotation results, and obtain a fine annotation result that includes a second score containing cognitive uncertainty;

[0042] The correction module is further configured to filter and correct the fine labeling results based on the second score to obtain the final pollution labeling results for the power transmission components.

[0043] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0044] Obtain inspection images of transmission lines, and sequentially perform noise suppression, illumination equalization correction, and standardization processing on the inspection images to obtain a standardized sample set to be labeled;

[0045] The sample set to be labeled is input into the trained hybrid detection model, and the preliminary detection result containing the first score of cognitive uncertainty is output.

[0046] Based on the first score, the preliminary detection results are filtered and corrected to obtain the pollution labeling results of the power transmission components;

[0047] The degree of filth in the inspection images in the filth labeling results is graded and labeled to obtain a refined labeling result that includes a second score of cognitive uncertainty;

[0048] Based on the second score, the fine annotation results are filtered and corrected to obtain the final pollution annotation results for the power transmission components.

[0049] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0050] Obtain inspection images of transmission lines, and sequentially perform noise suppression, illumination equalization correction, and standardization processing on the inspection images to obtain a standardized sample set to be labeled;

[0051] The sample set to be labeled is input into the trained hybrid detection model, and the preliminary detection result containing the first score of cognitive uncertainty is output.

[0052] Based on the first score, the preliminary detection results are filtered and corrected to obtain the pollution labeling results of the power transmission components;

[0053] The degree of filth in the inspection images in the filth labeling results is graded and labeled to obtain a refined labeling result that includes a second score of cognitive uncertainty;

[0054] Based on the second score, the fine annotation results are filtered and corrected to obtain the final pollution annotation results for the power transmission components.

[0055] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0056] Obtain inspection images of transmission lines, and sequentially perform noise suppression, illumination equalization correction, and standardization processing on the inspection images to obtain a standardized sample set to be labeled;

[0057] The sample set to be labeled is input into the trained hybrid detection model, and the preliminary detection result containing the first score of cognitive uncertainty is output.

[0058] Based on the first score, the preliminary detection results are filtered and corrected to obtain the pollution labeling results of the power transmission components;

[0059] The degree of filth in the inspection images in the filth labeling results is graded and labeled to obtain a refined labeling result that includes a second score of cognitive uncertainty;

[0060] Based on the second score, the fine annotation results are filtered and corrected to obtain the final pollution annotation results for the power transmission components.

[0061] The aforementioned method, apparatus, computer equipment, computer-readable storage medium, and computer program product for labeling the contamination level of power transmission components first acquire inspection images of the power transmission line. These images are then subjected to noise suppression, illumination equalization correction, and standardization processing to obtain a standardized sample set to be labeled. This sample set is input into a trained hybrid detection model, which outputs a preliminary detection result containing a first score of cognitive uncertainty. Based on this first score, the preliminary detection result is filtered and corrected to obtain the contamination labeling result for the power transmission components. The contamination level of the inspection images in the contamination labeling result is then graded to obtain a refined labeling result containing a second score of cognitive uncertainty. Based on this second score, the refined labeling result is filtered and corrected to obtain the final contamination labeling result for the power transmission components. Thus, by using the cognitive uncertainty score as the core control signal for a dynamic threshold, intelligent and flexible scheduling of decision-making strictness is achieved, improving the balance between precision and recall, and further enhancing the accuracy of contamination labeling. Attached Figure Description

[0062] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0063] Figure 1 This is an application environment diagram of a method for labeling the degree of contamination of power transmission components in one embodiment;

[0064] Figure 2 This is a flowchart illustrating a method for labeling the degree of contamination of power transmission components in one embodiment;

[0065] Figure 3 This is a structural block diagram of a device for indicating the degree of contamination of power transmission components in one embodiment;

[0066] Figure 4 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0067] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0068] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0069] The method for labeling the degree of contamination of power transmission components provided in this application embodiment can be applied to, for example, Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or located in the cloud or on other network servers. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, drones, low-altitude aircraft, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, projection devices, etc. Portable wearable devices can include smartwatches, smart bracelets, head-mounted devices, etc. Head-mounted devices can be virtual reality (VR) devices, augmented reality (AR) devices, smart glasses, etc. Server 104 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0070] In one exemplary embodiment, such as Figure 2 As shown, a method for labeling the degree of contamination of power transmission components is provided, which can be applied to... Figure 1 Taking terminal 102 as an example, the explanation includes the following steps 202 to 210. Wherein:

[0071] Step 202: Obtain inspection images of the transmission line, and perform noise suppression, illumination equalization correction and standardization processing on the inspection images in sequence to obtain a standardized sample set to be labeled.

[0072] For example, the inspection images of the transmission line collected during the inspection process are received, and the inspection images are sequentially subjected to noise suppression, illumination equalization correction and standardization processing to obtain a standardized sample set to be labeled.

[0073] The inspection images include, but are not limited to, elements such as power transmission components in power transmission lines and dirt on power transmission components.

[0074] Step 204: Input the sample set to be labeled into the trained hybrid detection model and output the preliminary detection results containing the first score of cognitive uncertainty.

[0075] Among them, the hybrid detection model is a hybrid Transformer component detection model.

[0076] Optionally, the sample set to be labeled can be input into the trained hybrid detection model for detection processing to obtain preliminary detection results.

[0077] The preliminary test results include a first score used to characterize the cognitive uncertainty of the results.

[0078] Step 206: Based on the first score, the preliminary detection results are filtered and corrected to obtain the pollution labeling results of the power transmission components.

[0079] For example, the preliminary detection results are filtered and corrected based on a first score that characterizes the cognitive uncertainty of the preliminary detection results to obtain the pollution labeling results of the power transmission components.

[0080] Step 208: The degree of filth in the inspection images in the filth labeling results is graded and labeled to obtain a fine labeling result that includes a second score containing cognitive uncertainty.

[0081] Optionally, the dirt labeling results are input into the trained first hybrid detection model, and after processing, a refined labeling result containing a second score of cognitive uncertainty is obtained.

[0082] Among them, the first hybrid detection model is a finely tuned hybrid Transformer model with a similar structure to the hybrid detection model (optimized for the detection of dirty areas and the four-level classification task of "second dirty category: pure / local / global / non-dusty") to finely label the degree of dirtiness.

[0083] Step 210: Based on the second score, the fine annotation results are filtered and corrected to obtain the final pollution annotation results of the power transmission components.

[0084] For example, the fine annotation results are filtered and corrected based on a second score that characterizes the cognitive uncertainty of the fine annotation results to obtain the final pollution annotation results for the power transmission components.

[0085] In some embodiments, the final pollution labeling results of the power transmission components are displayed through a visual interactive interface, and the parameters are adjusted upon receiving adjustment instructions from the user.

[0086] In some embodiments, detection results with a first score and a second score greater than a preset uncertainty threshold, but which have been verified as correct or have been reasonably corrected, are marked as high-value samples and stored in a persistent sample library. At preset intervals, an incremental learning task is initiated to sample from the persistent sample library to perform targeted fine-tuning and optimization of the hybrid detection model and the first hybrid detection model.

[0087] In the aforementioned method for labeling the contamination level of power transmission components, inspection images of the transmission line are acquired. These images are then subjected to noise suppression, illumination equalization correction, and standardization processing to obtain a standardized sample set to be labeled. This sample set is input into a trained hybrid detection model, which outputs a preliminary detection result containing a first score of cognitive uncertainty. Based on this first score, the preliminary detection result is filtered and corrected to obtain the contamination labeling result for the power transmission component. The contamination level of the inspection images in the contamination labeling result is then graded to obtain a refined labeling result containing a second score of cognitive uncertainty. Based on this second score, the refined labeling result is filtered and corrected to obtain the final contamination labeling result for the power transmission component. Thus, by using the cognitive uncertainty score as the core control signal for a dynamic threshold, intelligent and flexible scheduling of decision-making strictness is achieved, improving the balance between precision and recall, and further enhancing the accuracy of contamination labeling.

[0088] In an exemplary embodiment, the hybrid detection model includes a feature extraction network, a hybrid encoder, and a preset refiner. The process involves inputting a sample set to be labeled into the trained hybrid detection model and outputting a preliminary detection result containing a first score of cognitive uncertainty. This includes: inputting the sample set to be labeled into the feature extraction network to extract a multi-level feature set of the sample set; determining target features at a preset number of layers in the multi-level feature set; inputting the target features into the hybrid encoder to perform feature encoding on the target features to obtain a hybrid encoded feature map; and inputting the hybrid encoded feature map into the preset refiner to output a preliminary detection result containing a first score of cognitive uncertainty, the component category and boundary of the transmission component, the first contamination category and first prediction box of the contaminated area, and a first confidence level.

[0089] The hybrid detection model includes a feature extraction network with HGNetv2 as its backbone, a hybrid encoder, and a preset refiner (such as a Dirichlet distribution refiner).

[0090] In practice, the sample set I to be labeled is input into the feature extraction network to extract the multi-level feature set {f} of the sample set to be labeled. m |m=1,2,...,M}, where f m Spatial resolution decreases while semantic information increases.

[0091] Determine the target features at a predetermined number of layers in a multi-level feature set (e.g., the deepest level three features {f}). -1 ,f -2 ,f -3 The target features are input into the hybrid encoder H to perform feature encoding, resulting in a hybrid encoded feature map.

[0092] In some embodiments, for the deepest feature f -1 First, perform depthwise convolution: F temp =DwConv(f -1 Then, immediately follow up with additive self-attention global interaction: F att =A-SA(F temp This process can be formalized as: s1=BN(Conv(F) temp )+f -1 ); s2=BN(A-SA(s1)+s1), after further processing by depthwise separable convolution and spatial feedforward network, the purified feature s3 at this scale is obtained.

[0093] Among them, F att =A-SA(F temp ) indicates that F temp The input is fed into an Additive Self-Attention (A-SA) layer for processing. Batch Normalization (BN) is used to normalize the data. Conv(F) temp ) indicates that the input pointwise convolutional layer (Conv) is processed by convolution.

[0094] Connect s3 and f -2 ,f -3 Perform Scale Invariant Product Fusion (SIPF). For example, upsample s3 and then combine it with f. -2 Perform SIPF:F mid =SIPF(f -2 ,UpSample(s3)). Similarly, F mid Upsampling and f -3 Fusion: F high =SIPF(f -3 UpSample(F) mid The SIPF operation internally includes dynamic weight generation and weighted multiplication. To enhance the information flow, reverse downsampling fusion is also performed to form a bidirectional feature pyramid. Finally, all fused features are concatenated along the channel dimension to obtain a hybrid encoded feature map S rich in semantics and details.

[0095] In one embodiment, f -3 The combined feature F is obtained by concatenating s3 and s3 along the channel dimension.cat (shape [B,C) h +C l [,H,W]). Where B is the batch size, C h f -3 Number of channels, C l H represents the number of s3 channels, H represents the height, and W represents the width.

[0096] Dynamic weighted graph generation: F cat The input is a lightweight conditional subnetwork. A specific implementation of this subnetwork is a 3×3 convolutional layer (with C output channels). mid C mid =8), followed by a ReLU activation function, then a 1×1 convolutional layer (output channel number 1), and finally a Sigmoid activation function to generate a single-channel, spatially adaptive fusion weight map W with a value range of [0,1] and a size of [B,1,H,W]. This process can be formulated as: W=Sigmoid(Conv 1x1 (ReLU(Conv 3x3 (F cat Conv 3x3 The padding is set to 1 to maintain the space size.

[0097] Among them, ReLU (Rectified Linear Unit) is an activation function used to introduce nonlinear transformations, enabling the network to learn and express complex nonlinear relationships in the input features, and avoiding the degradation into a linear transformation after multiple layers of convolution.

[0098] Weighted product fusion: using a dynamic weight graph W on f -3 It performs content-adaptive non-linear fusion with S3. The core fusion formula is: .in, This represents element-wise multiplication. The 1-W operation is obtained by calculating W element-wise. The key to this operation is that in detailed spatial locations such as contaminated boundaries (usually corresponding to high edge gradients), the weight W approaches 1, and the output F... fused Main inheritance f -3 The fine details are captured; in flat background areas, the W value approaches 0, and the output mainly relies on the strong semantic information of s3. Through this dynamic product weighting based on local image content, the spatial adaptive optimization fusion of detail information and semantic information is achieved, effectively protecting high-frequency features that are crucial for localization.

[0099] The hybrid encoded feature map S is input into a preset refiner, which includes a Transformer decoding layer and an evidence head. First, the features are decoded using the Transformer decoder, and then the evidence e for each category is calculated using the evidence head. k =Softplus(Decoder(S)). This yields the Dirichlet distribution parameter α. k =e k +1 and total strength S=Σα k Cognitive uncertainty can be calculated through subjective logic, for example, the uncertain mass u = K / S (K is the number of categories, such as insulators, vibration dampers, etc.). Simultaneously, coordinates are output through the bounding box regression head. Finally, the model outputs a set of component-level quadruple prediction results P = (component category and boundary cls of the transmission component). i The first soiling category and the first prediction box B of the soiled area i First confidence level C i The first fraction u i This includes the component categories and boundaries of transmission components, the first pollution category and first prediction box of the polluted area, the first confidence level, and the first score of cognitive uncertainty. i .

[0100] Softplus is an activation function used to ensure that the evidence value e output by the Softplus function is greater than or equal to 0; Decoder is an abbreviation for Transformer, used to decode and interact with the multi-scale and multi-level information contained in the hybrid encoded feature map S to generate a high-quality feature representation rich in semantic information, providing accurate input for the evidence head (filth classification) and the bounding box regression head (location localization).

[0101] The area within the first prediction box is the predicted dirty area.

[0102] In the above embodiments, a hybrid detection model comprising a feature extraction network, a hybrid encoder, and a preset thinner is employed to achieve integrated intelligent processing from image input to joint detection of parts and dirt: First, a multi-level feature set is extracted through the feature extraction network, laying the foundation for subsequent multi-scale analysis; second, target features of a preset number of layers are input into the hybrid encoder for feature encoding, effectively fusing deep semantic information with shallow detail information, enhancing the model's representation ability for complex scenes such as blurred dirt edges and varying scales; finally, the preset thinner simultaneously outputs the first cognitive uncertainty score, part category and boundary, dirt category and prediction box, and confidence level, achieving multi-task collaborative output of part detection and preliminary dirt classification. The cognitive uncertainty score quantifies the model's grasp of each prediction, providing an interpretable and reliable basis for subsequent intelligent screening and verification; the joint output of parts and dirt avoids computational redundancy caused by multiple independent detections, improves system processing efficiency, and provides necessary data support for dirt logic verification based on part symmetry (such as upgrading local dirt to global dirt).

[0103] In an exemplary embodiment, the inspection image is sequentially subjected to noise suppression, illumination equalization correction, and standardization processing to obtain a standardized sample set to be labeled. This includes: sequentially performing Gaussian filtering and median filtering on the inspection image to obtain a denoised inspection image; performing illumination equalization correction on the denoised inspection image to obtain a corrected inspection image; dividing the corrected inspection image into a preset number of sub-images according to a preset method; converting the sub-images to a preset format and scaling them to a preset size to obtain a standardized sample set to be labeled.

[0104] In practice, Gaussian filtering is used to process the inspection image to suppress Gaussian noise caused by jitter during acquisition, and median filtering is used to remove salt-and-pepper noise caused by the outdoor environment, resulting in a denoised inspection image. A contrast-limited adaptive histogram equalization algorithm is applied to perform illumination equalization correction on the denoised inspection image, resulting in a corrected inspection image.

[0105] For inspection images whose size exceeds the preset size in the corrected inspection images, the images are divided into a preset number of sub-images based on the sliding window cutting method, while ensuring that the power transmission components in the inspection images are not cut. The sub-images are converted into a preset format (e.g., PNG format) and scaled to a preset size (e.g., 1024*1024) using a bilinear interpolation algorithm to obtain a standardized sample set to be labeled.

[0106] In the above embodiments, by performing hierarchical noise suppression through Gaussian filtering and median filtering sequentially on the inspection images, Gaussian noise caused by jitter during acquisition and salt-and-pepper noise present in the outdoor environment can be specifically removed, effectively improving the image signal-to-noise ratio. Illumination equalization correction through contrast-limited adaptive histogram equalization can significantly improve the texture visibility of components in shadow and highly reflective areas, providing clearer visual input for subsequent feature extraction. By standardizing large-format images to a uniform size through sliding window cutting and bilinear interpolation scaling, the power transmission components are avoided from being segmented, and the requirements of deep learning models for fixed-size input are met. This significantly reduces the risk of subsequent model recognition being interfered with by environmental factors such as noise, uneven illumination, and size differences, improving the robustness and annotation accuracy of the system in complex outdoor inspection scenarios.

[0107] In an exemplary embodiment, based on a first score, the preliminary detection results are filtered and corrected to obtain the pollution labeling results of the power transmission components. This includes: deleting the corresponding detection results if the first score of a detection result in the preliminary detection results is greater than a preset uncertainty threshold, to obtain the deleted preliminary detection results; selecting detection results with a first confidence level greater than a preset confidence level from the deleted preliminary detection results as the filtered preliminary detection results; determining the image region corresponding to the first prediction box of each detection result in the filtered preliminary detection results, extracting the target region in the image region that meets the preset extraction conditions, and generating a reflective mask based on the target region; calculating the region overlap between the reflective mask and the first prediction box, deleting detection results with a region overlap greater than a first dynamic threshold from the filtered preliminary detection results, to obtain the filtered preliminary detection results; and performing pollution logic pre-correction on the filtered preliminary detection results to obtain the pollution labeling results of the power transmission components.

[0108] In practice, a preset uncertainty threshold (e.g., 0.75) is determined, all detection results in the preliminary detection results are traversed, and all detection results with a first score greater than the preset uncertainty threshold are deleted to obtain the preliminary detection results after deletion. Among the preliminary detection results after deletion, the detection results with a first confidence level greater than the preset confidence level are selected as the preliminary detection results after filtering.

[0109] The image region corresponding to the first prediction box of the detection result in the preliminary detection results after screening is extracted, the image region is converted to HSV color space, and the high brightness (V channel) and low saturation (S channel) region is extracted as the target region. After morphological closing operation, a reflective mask is generated.

[0110] Calculate the region overlap (IoU) between the reflective mask and the first prediction frame. Delete detection results with region overlap greater than the first dynamic threshold from the preliminary detection results after filtering. Determine that the detection results with region overlap IoU greater than the first dynamic threshold are reflective mislabeling results to obtain the preliminary detection results after filtering. Perform pollution logic pre-correction on the preliminary detection results after filtering to obtain the pollution labeling results of the power transmission components.

[0111] In some embodiments, the u of the first prediction box is determined. i Calculate the first dynamic threshold θ i =β-γ*u i Where β is the baseline threshold and γ is the sensitivity coefficient (e.g., 0.2).

[0112] In the above embodiments, by deleting detection results with a first score higher than a preset uncertainty threshold, predictions that the model self-assesses as "the least reliable" are excluded from the source. Secondly, by combining traditional confidence screening, dual quality control of uncertainty and confidence is achieved. Then, by generating a reflective mask and calculating the regional overlap with the prediction box, adaptive reflective mislabeling is performed in combination with a first dynamic threshold dynamically calculated based on the first score. This effectively filters out reflective mislabeling while avoiding the accidental deletion of high-confidence predictions.

[0113] In an exemplary embodiment, the first contamination category includes local contamination, global contamination, and no contamination. The preliminary detection results after filtering are pre-corrected for contamination logic to obtain contamination labeling results for the power transmission components. This includes: selecting detection results from the preliminary detection results where the power transmission component has a symmetrical structure and the contamination category is local contamination, as the first detection result set; determining the axis of symmetry of the power transmission components in the first detection result set, and based on the axis of symmetry, determining the mirror region corresponding to the target contamination region of the detection results in the first detection result set; calculating the feature similarity between the mirror region and the target contamination region; determining a dynamic upgrade threshold based on a first score; and modifying the contamination category corresponding to the predicted box to global contamination if the feature similarity is greater than the dynamic upgrade threshold.

[0114] In practice, for the preliminary test results after filtering, if the transmission components are symmetrical structures (such as insulators) and the pollution category is local pollution, the Hungarian matching algorithm is used to establish top-tail pairing and determine the axis of symmetry.

[0115] Sim calculates the similarity (DCT low-frequency coefficients) of the features of the contaminated area and its mirror image area on the other side of the axis of symmetry in the detection results. i .

[0116] In one embodiment, via φ i =α+λ*u iCalculate the dynamic upgrade threshold, which is the baseline similarity threshold (e.g., 0.7), where λ is the prudence coefficient. If Sim i >φ i If the polluted areas are highly similar on the symmetrical side, it may be a local manifestation of global pollution. Therefore, its label is upgraded to global pollution, and the pollution labeling results of the power transmission components are obtained.

[0117] In some embodiments, the soiling labeling results include an identifier indicating whether each detection result has been corrected and a corresponding first score.

[0118] In the above embodiments, intelligent semantic optimization based on component geometric priors is achieved by performing logical pre-correction on transmission components with symmetrical structures and labeled as locally contaminated in the preliminary detection results after filtering. First, detection results that meet the conditions of symmetrical structure and local contamination are selected, focusing on the scenarios most likely to be misjudged. Second, by establishing a symmetry axis and determining the mirror region, feature similarity between the original contaminated region and the mirror region is calculated using feature extraction methods such as discrete cosine transform, which can objectively quantify the degree of symmetry of the contamination distribution. Then, the upgrade threshold is dynamically calculated based on the first score of cognitive uncertainty, linking the upgrade decision with the model's own level of confidence. On the one hand, by utilizing the inherent mechanical symmetry prior knowledge of components such as insulators, continuous global contamination that is misjudged as isolated local contamination due to limited local vision can be effectively identified. On the other hand, through the dynamic threshold mechanism of uncertainty constraints, overconfident execution of erroneous upgrades is avoided when the original evidence of the model is weak, significantly improving the semantic consistency and reliability of the labeling results.

[0119] In an exemplary embodiment, the fine annotation result includes a second contamination category and a second prediction box, a second confidence level, and a second score for the contaminated area. Based on the second score, the annotation result is filtered and corrected to obtain the final contamination annotation result for the power transmission component. This includes: deleting the corresponding detection result if the second score of a detection result in the fine annotation result is greater than a preset uncertainty threshold, to obtain a deleted fine detection result; selecting detection results with a second confidence level greater than a preset confidence level from the deleted fine detection results as the filtered fine detection results; determining the image region corresponding to the second prediction box of each detection result in the filtered fine detection results, extracting a first target region in the image region that meets preset extraction conditions, and generating a reflective mask based on the first target region; calculating the regional overlap between the reflective mask and the second prediction box, deleting detection results with a regional overlap greater than a second dynamic threshold from the filtered fine detection results, to obtain a filtered fine detection result; and performing contamination logic correction on the filtered fine detection results to obtain the final contamination annotation result for the power transmission component.

[0120] In practice, if a detection result in the fine-marking results has a second score greater than a preset uncertainty threshold, the corresponding detection result is deleted to obtain a fine-marking result after deletion. From the fine-marking result after deletion, detection results with a second confidence level greater than a preset confidence level are selected as the fine-marking result after filtering. The image region corresponding to the second prediction box of each detection result in the fine-marking result after filtering is determined, and a first target region satisfying a preset extraction condition is extracted from the image region. A reflective mask is generated based on the first target region. The overlap between the reflective mask and the second prediction box is calculated, and detection results with an overlap greater than a second dynamic threshold are deleted from the fine-marking result after filtering to obtain the fine-marking result after filtering. The fine-marking result after filtering is then subjected to contamination logic correction to obtain the final contamination marking result for the power transmission component.

[0121] In one embodiment, the process of performing contamination logic correction on the filtered fine detection results to obtain the final contamination labeling result of the transmission component is the same as the process of performing contamination logic pre-correction on the filtered preliminary detection results to obtain the contamination labeling result of the transmission component.

[0122] The calculation process for the second dynamic threshold is θ'=β'-γ'*u'. j The calculation process for the first dynamic upgrade threshold in the process of correcting the contamination logic of the filtered fine detection results is φ'=α'+λ'*u'. j Among them, u' j As the second fraction, β', γ', α', and λ' can all be set by the user according to the actual situation, and this application embodiment does not limit this.

[0123] In the above embodiments, a second purification of the dirt labeling results is achieved on the basis of the first round of filtering. Especially for difficult samples such as those with strong reflective interference and blurred dirt edges, the mislabeling rate can be further reduced. At the same time, the decision strictness is dynamically adjusted by using the second score, which effectively protects the dirt labels with high confidence from being deleted by mistake, greatly improving the labeling robustness of the system in complex outdoor environments and the credibility of the final output results.

[0124] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0125] Based on the same inventive concept, this application also provides a device for marking the degree of contamination of power transmission components to implement the above-described method for marking the degree of contamination of power transmission components. The solution provided by this device is similar to the solution described in the above-described method. Therefore, the specific limitations of one or more embodiments of the device for marking the degree of contamination of power transmission components provided below can be found in the limitations of the method for marking the degree of contamination of power transmission components described above, and will not be repeated here.

[0126] In one exemplary embodiment, such as Figure 3 As shown, a device for marking the degree of contamination of power transmission components is provided, comprising: a processing module 301, an output module 302, a correction module 303, and a marking module 304, wherein:

[0127] The processing module is used to acquire inspection images of transmission lines, and sequentially perform noise suppression, illumination equalization correction and standardization processing on the inspection images to obtain a standardized sample set to be labeled.

[0128] The output module is used to input the sample set to be labeled into the trained hybrid detection model and output a preliminary detection result containing a first score of cognitive uncertainty.

[0129] The correction module is used to filter and correct the preliminary detection results based on the first score to obtain the pollution labeling results of the power transmission components.

[0130] The annotation module is used to classify and annotate the degree of dirtiness in the inspection images in the dirt annotation results, and obtain a fine annotation result that includes a second score containing cognitive uncertainty.

[0131] The correction module is further configured to filter and correct the fine labeling results based on the second score to obtain the final pollution labeling results for the power transmission components.

[0132] Each module in the aforementioned device for indicating the degree of contamination of power transmission components can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0133] In one exemplary embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 4 As shown, the computer device includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. When executed by the processor, the computer program implements a method for labeling the degree of contamination of power transmission components. The display unit is used to form a visually visible image and can be a display screen, projection device, or virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0134] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0135] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0136] Obtain inspection images of transmission lines, and sequentially perform noise suppression, illumination equalization correction, and standardization processing on the inspection images to obtain a standardized sample set to be labeled;

[0137] The sample set to be labeled is input into the trained hybrid detection model, and the preliminary detection result containing the first score of cognitive uncertainty is output.

[0138] Based on the first score, the preliminary detection results are filtered and corrected to obtain the pollution labeling results of the power transmission components;

[0139] The degree of filth in the inspection images in the filth labeling results is graded and labeled to obtain a refined labeling result that includes a second score of cognitive uncertainty;

[0140] Based on the second score, the fine annotation results are filtered and corrected to obtain the final pollution annotation results for the power transmission components.

[0141] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0142] Obtain inspection images of transmission lines, and sequentially perform noise suppression, illumination equalization correction, and standardization processing on the inspection images to obtain a standardized sample set to be labeled;

[0143] The sample set to be labeled is input into the trained hybrid detection model, and the preliminary detection result containing the first score of cognitive uncertainty is output.

[0144] Based on the first score, the preliminary detection results are filtered and corrected to obtain the pollution labeling results of the power transmission components;

[0145] The degree of filth in the inspection images in the filth labeling results is graded and labeled to obtain a refined labeling result that includes a second score of cognitive uncertainty;

[0146] Based on the second score, the fine annotation results are filtered and corrected to obtain the final pollution annotation results for the power transmission components.

[0147] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, performs the following steps:

[0148] Obtain inspection images of transmission lines, and sequentially perform noise suppression, illumination equalization correction, and standardization processing on the inspection images to obtain a standardized sample set to be labeled;

[0149] The sample set to be labeled is input into the trained hybrid detection model, and the preliminary detection result containing the first score of cognitive uncertainty is output.

[0150] Based on the first score, the preliminary detection results are filtered and corrected to obtain the pollution labeling results of the power transmission components;

[0151] The degree of filth in the inspection images in the filth labeling results is graded and labeled to obtain a refined labeling result that includes a second score of cognitive uncertainty;

[0152] Based on the second score, the fine annotation results are filtered and corrected to obtain the final pollution annotation results for the power transmission components.

[0153] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0154] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0155] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0156] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for labeling the degree of contamination of power transmission components, characterized in that, The method includes: Obtain inspection images of transmission lines, and sequentially perform noise suppression, illumination equalization correction, and standardization processing on the inspection images to obtain a standardized sample set to be labeled; The sample set to be labeled is input into the trained hybrid detection model, and the preliminary detection result containing the first score of cognitive uncertainty is output. Based on the first score, the preliminary detection results are filtered and corrected to obtain the pollution labeling results of the power transmission components; The degree of filth in the inspection images in the filth labeling results is graded and labeled to obtain a refined labeling result that includes a second score containing cognitive uncertainty; Based on the second score, the fine annotation results are filtered and corrected to obtain the final pollution annotation results for the power transmission components.

2. The method according to claim 1, characterized in that, The hybrid detection model includes a feature extraction network, a hybrid encoder, and a preset refiner; the step of inputting the sample set to be labeled into the trained hybrid detection model and outputting a preliminary detection result containing a first score of cognitive uncertainty includes: The sample set to be labeled is input into the feature extraction network to extract a multi-level feature set of the sample set to be labeled; Determine the target features of a preset number of layers in the multi-level feature set, input the target features into the hybrid encoder, perform feature encoding on the target features, and obtain a hybrid encoded feature map; The hybrid encoded feature map is input into a preset refiner, and the output includes a first score of cognitive uncertainty, the component category and boundary of the power transmission component, the first pollution category and first prediction box of the polluted area, and a preliminary detection result with a first confidence level.

3. The method according to claim 1, characterized in that, The inspection images are sequentially subjected to noise suppression, illumination equalization correction, and standardization processing to obtain a standardized sample set to be labeled, including: Gaussian filtering and median filtering are applied sequentially to the inspection image to obtain a denoised inspection image; The denoised inspection image is then subjected to illumination equalization correction to obtain the corrected inspection image. The corrected inspection image is divided into a preset number of sub-images according to a preset method. The sub-images are converted into a preset format and then scaled to a preset size to obtain a standardized sample set to be labeled.

4. The method according to claim 2, characterized in that, The process of filtering and correcting the preliminary detection results based on the first score to obtain the pollution labeling results for the power transmission components includes: If, in the preliminary detection results, there is a detection result whose first score is greater than a preset uncertainty threshold, the corresponding detection result is deleted to obtain the preliminary detection results after deletion. In the preliminary detection results after deletion, the detection results with a first confidence level greater than the preset confidence level are selected as the preliminary detection results after filtering; The image region corresponding to the first prediction box of each detection result in the preliminary detection results after screening is determined, the target region in the image region that meets the preset extraction conditions is extracted, and a reflective mask is generated based on the target region; Calculate the regional overlap between the reflective mask and the first prediction box, and delete the detection results with regional overlap greater than the first dynamic threshold from the preliminary detection results after filtering to obtain the preliminary detection results after filtering. The initial test results after filtration are pre-corrected using pollution logic to obtain the pollution labeling results for the power transmission components.

5. The method according to claim 4, characterized in that, The first pollution category includes local pollution, global pollution, and no pollution; the pollution logic pre-correction of the preliminary detection results after filtering to obtain the pollution labeling results of the power transmission components includes: From the preliminary test results after filtering, the test results of power transmission components with symmetrical structures and localized pollution were selected as the first set of test results; Determine the axis of symmetry of the power transmission component in the first set of detection results, and based on the axis of symmetry, determine the mirror region corresponding to the target contamination region in the first set of detection results; Calculate the feature similarity between the mirrored region and the target contaminated region; Based on the first score, determine the dynamic upgrade threshold; If the feature similarity is greater than the dynamic upgrade threshold, the dirt category corresponding to the predicted box will be modified to global dirt.

6. The method according to claim 1, characterized in that, The refined annotation results include a second pollution category and a second prediction box for the polluted area, a second confidence level, and a second score; based on the second score, the annotation results are filtered and corrected to obtain the final pollution annotation results for the transmission components, including: If, in the fine annotation results, there is a second score corresponding to a detection result that is greater than a preset uncertainty threshold, the corresponding detection result is deleted to obtain the fine detection result after deletion. In the refined detection results after deletion, the detection results with a second confidence level greater than the preset confidence level are selected as the refined detection results after filtering; Determine the image region corresponding to the second prediction box of each detection result in the filtered fine detection results, extract the first target region in the image region that meets the preset extraction conditions, and generate a reflective mask based on the first target region; Calculate the regional overlap between the reflective mask and the second prediction box, and delete the detection results with regional overlap greater than the second dynamic threshold from the filtered fine detection results to obtain the filtered fine detection results. The filtered fine test results are corrected for contamination logic to obtain the final contamination labeling results for the power transmission components.

7. A device for indicating the degree of contamination of power transmission components, characterized in that, The device includes: The processing module is used to acquire inspection images of transmission lines, and sequentially perform noise suppression, illumination equalization correction and standardization processing on the inspection images to obtain a standardized sample set to be labeled. The output module is used to input the sample set to be labeled into the trained hybrid detection model and output a preliminary detection result containing a first score of cognitive uncertainty; The correction module is used to filter and correct the preliminary detection results based on the first score to obtain the pollution labeling results of the power transmission components; The annotation module is used to classify and annotate the degree of dirtiness in the inspection images in the dirt annotation results, and obtain a fine annotation result that includes a second score containing cognitive uncertainty; The correction module is further configured to filter and correct the fine labeling results based on the second score to obtain the final pollution labeling results for the power transmission components.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images