Power transmission line insulator defect detection method and device

Abstract

The invention provides a power transmission line insulator defect detection method and apparatus. The method comprises the steps of obtaining a to-be-detected insulator inspection image and inputting the to-be-detected insulator inspection image into a backbone network to extract multi-scale features; the backbone network comprises a plurality of SFCF feature enhancement modules, the SFCF modules divide input features into first branch features and second branch features, multi-stage processing including convolution transformation and attention calibration is carried out on the second branch features, and after intermediate features of multi-stage processing are fused, the intermediate features are combined with the first branch features to output enhancement features; inputting the multi-scale features into a feature fusion and enhancement network to obtain multi-scale fusion features, and setting an SFCF feature enhancement module behind at least one fusion node to enhance the fusion features; and inputting the multi-scale fusion features into a detection head to predict a defect bounding box, a defect category and confidence, and screening according to a threshold value to generate defect detection information. According to the technical scheme, automatic detection of insulator defects is realized, and the defect identification accuracy and stability are improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent inspection of power equipment and computer vision technology, specifically relating to a method and device for detecting defects in power transmission line insulators. Background Technology

[0002] Insulators in power transmission lines provide electrical insulation and mechanical support. Long-term exposure to outdoor environments such as pollution, aging, discharge erosion, and mechanical impacts can easily lead to defects such as cracks, breakage, discharge marks, and corrosion. If these defects are not detected and addressed promptly, they may induce flashover or insulation breakdown, thereby affecting the safe and stable operation of the power system. Current inspection methods, such as manual inspection, are costly, inefficient, and susceptible to environmental and subjective factors. With the widespread adoption of drone inspections, automatic detection based on inspection images has become an important direction. However, insulators are often obstructed by trees, hardware, conductors, or crossarms, and complex background interference, affecting the stability of defect feature extraction and identification.

[0003] Insulator defects are often characterized by small size, irregular shape, indistinct texture / edge, and susceptibility to complex backgrounds and lighting changes. Fine-grained local structural information in insulator images is easily weakened, increasing the risk of missed detection.

[0004] Therefore, there is an urgent need for an insulator defect detection method that can effectively enhance the extraction and perception of local structural anomalies under complex background, occlusion, and weak texture conditions, improve the detection accuracy and reliability of local partial missing defects, and reduce the risk of missed detection under complex background, occlusion, and weak texture conditions. Summary of the Invention

[0005] This invention provides a method for detecting defects in power transmission line insulators. The method constructs a detection network comprising cascaded backbone feature extraction, multi-scale feature fusion, and detection head prediction. Multiple SFCF feature enhancement modules are introduced into the backbone network and feature fusion nodes to perform branching, convolutional transformation, attention calibration, and intermediate feature fusion on the features. The method outputs the defect location, category, and confidence level in the inspection image, thereby preserving and enhancing local defect features during multi-scale feature extraction and fusion, and improving the detection accuracy and reliability of some missing insulator defects.

[0006] In a first aspect, the present invention provides a method for detecting defects in power transmission line insulators, the method comprising: An inspection image of an insulator to be inspected in a transmission line is acquired. This image is then input into a cascaded backbone network. The image is processed along the cascaded path to output multi-scale features. The backbone network includes multiple SFCF feature enhancement modules. The multi-scale features are stage features output by the backbone network at at least two different network stages, and each stage feature is the output of the preceding cascaded modules. Each SFCF feature enhancement module executes the following process: the output features of the preceding cascaded modules are divided in the channel dimension into a first branch feature for preserving original information and a second branch feature for deep feature transformation; the second branch feature undergoes multi-level processing, which is implemented by at least one cascaded feature processing unit within a single SFCF feature enhancement module. Each level of processing sequentially includes convolutional transformation and attention-based calibration; intermediate feature information generated by the multi-level processing is fused, and the fused information is combined with the first branch feature to output an enhanced stage feature map. The multi-scale features output from the backbone network are input into the feature fusion and enhancement network. The features at different scales are fused through upsampling, downsampling, and feature concatenation to obtain multi-scale fused features. At least one SFCF feature enhancement module is set after at least one fusion node of the feature fusion and enhancement network to perform feature enhancement operations on the fused features. The multi-scale fusion features are input into the detection head to predict the bounding box position, defect category, and confidence level of the defect category of the insulator, thereby obtaining the insulator defect prediction result. The prediction result is then filtered according to a preset threshold to generate insulator defect detection information.

[0007] By adopting the above scheme, the present invention provides a method for detecting defects in transmission line insulators. This method introduces an SFCF feature enhancement module into the backbone network, dividing the input features into a first branch feature that retains the original information and a second branch feature for deep feature transformation, achieving parallel preservation of original structural information and deep representation information. By performing multi-level processing on the second branch features, including convolutional transformation and attention-based calibration, and fusing the intermediate feature information from the multi-level processing with the first branch features, continuous extraction and aggregation of features related to local structural anomalies are achieved. Furthermore, multi-scale feature fusion is completed in the feature fusion and enhancement network based on upsampling, downsampling, and feature splicing. An SFCF feature enhancement module is set after at least one fusion node to further enhance the fused features, achieving effective integration of cross-scale features and reinforcement of key fine-grained information. Finally, by decoupled regression and classification branches, defect bounding boxes, defect categories, and confidence scores are output and filtered according to thresholds, achieving stable output of detection results and significantly improving the detection capability and consistency of some missing insulator defects under complex backgrounds and weak texture conditions. The technical solution of the present invention can realize automatic detection of insulator defects, improving the accuracy and reliability of defect identification.

[0008] In some embodiments of the present invention, the backbone network is a cascaded structure, which includes at least a first convolutional module, a second convolutional module, a first SFCF feature enhancement module, a third convolutional module, a second SFCF feature enhancement module, a first downsampling module, a third SFCF feature enhancement module, a second downsampling module, a fourth SFCF feature enhancement module, a spatial pyramid aggregation module, and an attention enhancement module.

[0009] In some embodiments of the present invention, dividing the output features of the preceding cascaded module into first branch features and second branch features along the channel dimension includes: performing channel compression and channel rearrangement on the input features and then dividing them along the channel dimension to obtain first branch features and second branch features.

[0010] In some embodiments of the present invention, the multi-level processing of the second branch feature includes: sequentially inputting the second branch feature into at least two cascaded feature processing units to update the second branch feature in each level of feature processing unit and output the corresponding intermediate feature information.

[0011] In some embodiments of the present invention, the attention-based calibration includes: after convolution transformation, calculating channel attention weights and spatial attention weights based on the features output by the convolution transformation, and performing the calibration by multiplying the channel attention weights and spatial attention weights element-wise with the input features at each level.

[0012] In some embodiments of the present invention, the calculation of the spatial attention weights includes: performing depthwise separable convolution processing on the input features, and performing anisotropic convolution processing along the horizontal and vertical directions respectively to generate spatial attention weights.

[0013] In some embodiments of the present invention, the fusion of intermediate feature information generated by the multi-level processing, and the combination of the fused information with the first branch feature, includes: The intermediate feature information at each level is concatenated and transformed by convolution to generate a fused feature; then, the fused feature is concatenated and / or weighted and superimposed with the first branch feature to output the enhanced feature.

[0014] In some embodiments of the present invention, the feature fusion and enhancement network includes: a top-down upsampling fusion path and a bottom-up downsampling fusion path, wherein the upsampling fusion path and the downsampling fusion path achieve the fusion of features at different scales through feature splicing, and output the multi-scale fusion features at least two scales.

[0015] In some embodiments of the present invention, filtering the prediction results according to a preset threshold includes: removing low-confidence candidate boxes based on a confidence threshold, and performing overlap suppression processing on the remaining candidate boxes; the overlap suppression processing retains the target candidate boxes based on the overlap threshold between candidate boxes, and generates the insulator defect detection information.

[0016] Compared with existing technologies, the advantages of this invention are as follows: By acquiring and preprocessing insulator inspection images before inputting them into a detection network, this invention achieves unified representation and processability of inspection image data; by setting multiple SFCF feature enhancement modules in the backbone network and dividing the input features into first-branch features and second-branch features, it achieves parallel preservation of original structural information and deep feature transformation information; by performing multi-level processing including convolutional transformation and attention calibration on the second-branch features, and fusing the intermediate feature information from the multi-level processing with the first-branch features, it achieves progressive extraction of features related to local structural anomalies. This invention employs a cross-scale aggregation approach. Through feature fusion and enhancement networks, multi-scale feature fusion is achieved based on upsampling, downsampling, and feature concatenation. An SFCF feature enhancement module is then added after at least one fusion node to further enhance the fused features, enabling effective integration of cross-scale information and reinforcement of fine-grained defect clues. Furthermore, the detection head uses decoupled regression and classification branches to output defect bounding boxes, defect categories, and confidence levels, respectively, and filters the results according to thresholds. This ensures stable and consistent detection output, significantly improving the recognition ability and robustness of some missing insulator defects under complex backgrounds, occlusion, and weak texture conditions. The technical solution of this invention enables automatic detection of insulator defects, improving the accuracy, stability, and engineering applicability of defect identification.

[0017] A second aspect of the present invention provides a defect detection system for power transmission line insulators, comprising: The image acquisition module is used to acquire inspection images of insulators to be inspected in the transmission line and input the inspection images into the backbone network of the cascaded structure. The backbone feature extraction module processes the inspection image along a cascaded path and outputs multi-scale features. The backbone network includes multiple SFCF feature enhancement modules. The multi-scale features are stage features output by the backbone network at at least two different network stages, and each stage feature is the output of the preceding cascaded modules. Each SFCF feature enhancement module executes the following process: dividing the output features of the preceding cascaded modules into a first branch feature for preserving original information and a second branch feature for deep feature transformation in the channel dimension; performing multi-level processing on the second branch feature, where each level of processing is implemented by at least one cascaded feature processing unit within a single SFCF feature enhancement module, wherein each level of processing sequentially includes convolutional transformation and attention-based calibration; fusing the intermediate feature information generated by the multi-level processing, combining the fused information with the first branch feature, and outputting an enhanced stage feature map. The feature fusion and enhancement module is used to input the multi-scale features output by the backbone network into the feature fusion and enhancement network, and to fuse features of different scales through upsampling, downsampling and feature concatenation to obtain multi-scale fused features; wherein, after at least one fusion node of the feature fusion and enhancement network, at least one SFCF feature enhancement module is set to perform feature enhancement operation on the fused features; The defect prediction module is used to input the multi-scale fused features into the detection head, predict the bounding box position, defect category and confidence level of the defect in the insulator, and obtain the insulator defect prediction result. The result filtering module is used to filter the prediction results of the defect prediction module according to a preset threshold and generate insulator defect detection information.

[0018] A third aspect of the present invention provides a device for detecting defects in power transmission line insulators, characterized in that the device includes a computer device, the computer device including a processor and a memory, the processor storing computer instructions, and when the computer instructions are executed, the device implements the aforementioned insulator defect detection method.

[0019] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the text, or may be learned by practice of the invention. The objects and other advantages of the invention will become apparent from the description and the accompanying drawings.

[0020] Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable with the present invention will become clearer from the following detailed description. Attached Figure Description

[0021] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0022] In the attached diagram: Figure 1 This is a flowchart illustrating a method for detecting defects in power transmission line insulators according to an embodiment of the present invention.

[0023] Figure 2 This is a schematic diagram of the overall structure of a method for detecting defects in power transmission line insulators provided in an embodiment of the present invention.

[0024] Figure 3 This is a schematic diagram of the SFCF module structure of a method for detecting defects in power transmission line insulators provided in an embodiment of the present invention.

[0025] Figure 4 This is a schematic diagram of a defect detection system for power transmission line insulators provided in an embodiment of the present invention.

[0026] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0027] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0028] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terminology used in this invention is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.

[0029] Figure 1 This is a flowchart illustrating a method for detecting defects in power transmission line insulators according to an embodiment of the present invention.

[0030] Example 1, as Figure 1 As shown, the present invention provides a method for detecting defects in power transmission line insulators, the method comprising the following steps: S1. Obtain an inspection image of the insulators to be inspected in the transmission line, input the inspection image into the backbone network of the cascade structure, process the inspection image along the cascade path, and output multi-scale features; wherein, the backbone network includes multiple SFCF feature enhancement modules; wherein, the multi-scale features are stage features output by the backbone network at at least two different network stages, and each stage feature is the output after processing by the preceding cascade module; the process executed by each SFCF feature enhancement module is as follows: divide the output features of the preceding cascade module into a first branch feature for preserving the original information and a second branch feature for deep feature transformation in the channel dimension; perform multi-level processing on the second branch feature, wherein the multi-level processing is implemented by at least one feature processing unit cascaded within a single SFCF feature enhancement module, wherein each level of processing sequentially includes convolution transformation and calibration based on an attention mechanism; fuse the intermediate feature information generated by the multi-level processing, combine the fused information with the first branch feature, and output the enhanced stage feature map; S2. Input the multi-scale features output from the backbone network into the feature fusion and enhancement network, and fuse the features at different scales through upsampling, downsampling and feature concatenation to obtain multi-scale fused features; wherein, after at least one fusion node of the feature fusion and enhancement network, at least one SFCF feature enhancement module is set to perform feature enhancement operation on the fused features; S3. Input the multi-scale fusion features into the detection head to predict the bounding box position, defect category, and confidence level of the defect category of the insulator, obtain the insulator defect prediction result, filter the prediction result according to the preset threshold, and generate insulator defect detection information.

[0031] By adopting the above scheme, the present invention provides a method for detecting defects in transmission line insulators. This method introduces an SFCF feature enhancement module into the backbone network, dividing the input features into a first branch feature that retains the original information and a second branch feature for deep feature transformation, achieving parallel preservation of original structural information and deep representation information. By performing multi-level processing on the second branch features, including convolutional transformation and attention-based calibration, and fusing the intermediate feature information from the multi-level processing with the first branch features, continuous extraction and aggregation of features related to local structural anomalies are achieved. Furthermore, multi-scale feature fusion is completed in the feature fusion and enhancement network based on upsampling, downsampling, and feature splicing. An SFCF feature enhancement module is set after at least one fusion node to further enhance the fused features, achieving effective integration of cross-scale features and reinforcement of key fine-grained information. Finally, by decoupled regression and classification branches, defect bounding boxes, defect categories, and confidence scores are output and filtered according to thresholds, achieving stable output of detection results and significantly improving the detection capability and consistency of some missing insulator defects under complex backgrounds and weak texture conditions. The technical solution of the present invention can realize automatic detection of insulator defects, improving the accuracy and reliability of defect identification.

[0032] Specifically, in this embodiment, the overall structure of the algorithm model for implementing the method of the present invention is as follows: Figure 2 As shown, the algorithm model includes a backbone network, a neck network, and a head.

[0033] The backbone network of this invention comprises, in sequence: a Conv1 convolutional module, a Conv2 convolutional module, an SFCF1 feature enhancement module, a Conv3 convolutional module, an SFCF2 feature enhancement module, an SCDown1 downsampling module, an SFCF3 feature enhancement module, an SCDown2 downsampling module, an SFCF4 feature enhancement module, an SPPF block, and a PSA module. The Conv1 convolution module is used to perform initial convolution feature extraction on the input insulator inspection image and complete the first feature mapping to extract basic edge and texture information in the image; The Conv2 convolution module is used to perform further convolution operations on the output features of Conv1 and to perform downsampling processing to reduce the feature map resolution and expand the receptive field. The SFCF1 feature enhancement module is used to perform structured feature splitting, recursive enhancement, and cross-branch fusion processing on the features output by Conv2, so as to enhance the expression of local structural anomalies at a higher spatial resolution. The Conv3 convolution module is used to perform convolution operations on the features processed by SFCF1 and to perform further downsampling in order to extract higher-level semantic features. The SFCF2 feature enhancement module is used to perform structured enhancement processing on the features output by Conv3, so as to continuously strengthen the modeling ability of structural incomplete features caused by partial missing parts of the insulator at the mesoscale feature level; (the output P3 of this module is fed into the feature enhancement and fusion network) The SCDown1 downsampling module is used to downsample the feature map while preserving the channel feature relationships, so as to achieve scale conversion and provide a suitable spatial resolution for subsequent deep feature modeling; The SFCF3 feature enhancement module is used to perform structured feature enhancement on the features downsampled by SCDown1. Through recursive feature modeling and cross-branch fusion, it further highlights local structural anomalies. (The output P4 of this module is fed into the feature enhancement and fusion network.) The SCDown2 downsampling module is used to re-downsample the features processed by SFCF3 to obtain a deeper level of feature representation. The SFCF4 feature enhancement module is used to perform structured enhancement processing on the deep features after SCDown2 downsampling, so that the structural anomalies corresponding to local missing features are continuously preserved in the high-level semantic modeling process. SPPF blocks are used to perform fast spatial pyramid aggregation of features within different receptive fields while maintaining the spatial resolution of the feature map, thereby enhancing the multi-scale contextual representation of features. The PSA module performs attention weighting on the SPPF-aggregated features, adaptively adjusting the feature response intensity to highlight important feature information related to the defect region. (The output P5 of this module is fed into the feature enhancement and fusion network.) The neck network used for feature enhancement and fusion includes: Upsample1 upsampling module, Concat1 feature concatenation module, SFCF5 feature enhancement module, Upsample2 upsampling module, Concat2 feature concatenation module, SFCF6 feature enhancement module, Conv4 module, Concat3 feature concatenation module, SFCF7 feature enhancement module, SCDown3 downsampling module, Concat4 feature concatenation module, and C2fCIB module. Among these, The Upsample1 upsampling module is used to upsample the features output from the backbone network PSA module to make their spatial resolution consistent with that of the mid-layer features, so as to facilitate feature fusion. The Concat1 feature concatenation module is used to concatenate the features output by Upsample1 with the features output by the SFCF3 feature enhancement module of the backbone network in the channel dimension to form a mesoscale fused feature. The SFCF5 feature enhancement module is used to perform structured feature splitting, recursive enhancement, and cross-branch fusion processing on mesoscale fused features to enhance the ability to express local structural anomalies. The Upsample2 upsampling module is used to further upsample the features processed by SFCF5 to keep their spatial resolution consistent with that of the shallow features. The Concat2 feature concatenation module is used to concatenate the features output by Upsample2 with the features output by the SFCF2 feature enhancement module of the backbone network along the channel dimension to form a high-resolution fused feature. The SFCF6 feature enhancement module is used to perform structured enhancement processing on high-resolution fused features, so that structural anomalies caused by local missing features are fully modeled at the shallow scale; (the output of this module is sent to the detection head) The Conv4 module is used to perform convolution operations on the fused features processed by SFCF6 to compress the number of channels and integrate feature information, providing a stable representation for subsequent feature backpropagation; The Concat3 feature concatenation module is used to concatenate the features output by Conv4 with features from the mesoscale path to further integrate feature information from different scales. The SFCF7 feature enhancement module is used to perform structured feature enhancement processing on the spliced ​​and fused features. Through recursive feature modeling and cross-branch fusion, it further enhances the expression of structural anomalies caused by partial insulator defects. (The output of this module is sent to the detection head.) The SCDown3 downsampling module is used to downsample the features processed by SFCF7 in order to achieve feature scale back and participate in deep feature fusion. The Concat4 feature concatenation module is used to concatenate the features output by SCDown3 with the deep path features along the channel dimension to form a deep-scale fused feature. The C2fCIB module performs multi-branch convolution and feature fusion processing on deep-scale fused features to generate the final multi-scale feature output of the neck network. (The output of this module is fed into the detection head.) The detection head decouples regression and classification prediction from multi-scale features from the neck network, and combines distributed bounding box decoding and candidate selection mechanisms to achieve high-precision detection of the location and category of target defects. Specifically, the following steps are involved: (1) Multi-scale feature input The detection head first receives multi-scale feature maps output from the neck network. Different scale feature maps correspond to different spatial resolutions and are used to detect target defects of different sizes.

[0034] (2) Decoupling of prediction branches For each scale of input features, the detection head uses a decoupled structure to construct regression prediction branches and classification prediction branches respectively: The regression prediction branch is used to model the location and shape information of the target; The classification prediction branch is used to model the category to which the target belongs and its confidence information.

[0035] (3) Bounding box regression prediction process In the regression prediction branch, the detection head outputs regression parameters associated with the bounding box for each spatial location, which describe the position and size of the target in the image.

[0036] (4) Category confidence prediction process In the classification prediction branch, the detector head outputs the confidence value corresponding to each category for each spatial location, which is used to indicate the probability that the location belongs to a certain target category.

[0037] (5) Decoding the prediction results and candidate selection The regression prediction results and classification prediction results output by the detection head are first decoded to map the network output into candidate bounding boxes and their corresponding class confidence scores in the actual image coordinate system.

[0038] (6) Redundancy detection result suppression and output Among the candidate detection results, there may be multiple highly overlapping predicted boxes. The detection head uses non-maximum suppression (NMS) or an equivalent screening strategy to suppress overlapping candidate boxes, retaining only the detection results with higher confidence and better spatial location.

[0039] In this embodiment, as Figure 3 As shown, the SFCF mainly consists of three components: 1) The main structure for implementing splitting, transformation, and merging; 2) A recursive enhancement structure consisting of multiple EffusionNeck modules that integrate FocusCalibNeck; 3) Cross-branch feature fusion path.

[0040] in, 1) The main structure for implementing splitting, transforming, and merging (Split–Transform–Merge) includes: Given input feature mapping The SFCF module first compresses and reassembles the channels using a 1×1 convolution.

[0041]

[0042] Where c= C_out e , This represents the channel compression ratio.

[0043] Then, tensor Divided into two parts along the channel dimension:

[0044] in, It is retained as a shortcut branch, and Sent to The structure is refined step by step in a cascaded EffusionNeck block.

[0045] make Indicates the first Given an input to an EffusionNeck, its feature refinement process can be recursively defined as:

[0046] in, Indicates the first The transformations performed by each EffusionNeck module. Intermediate output. This was then used for cross-branch feature fusion.

[0047] 2) EffusionNeck module: Each EffusionNeck block consists of two consecutive... It consists of convolutional layers, followed by a FocusCalibNeck attention module, and the input features are added to the output features through residual connections.

[0048] i) For the input The calculation process of this module is as follows:

[0049] in, This represents the SiLU nonlinear activation function, with each convolutional layer followed by a batch normalization operation. Subsequently, the resulting features are calibrated using the FocusCalibNeck attention module. The implementation and its final output are:

[0050] ii) FocusCalibNeck module section: FocusCalibNeck is a dual-branch attention module used to jointly model multi-scale channel attention and spatial attention. For a given input feature: The channel attention branch simultaneously aggregates global and local statistical information:

[0051] in, This indicates global average pooling. Subsequently, the two descriptors are broadcast to the size... And stitch them together along the channel dimension:

[0052] Next, through a pair of bottleneck structures with nonlinear activation Convolution generates channel attention maps:

[0053] in, This represents the Sigmoid function.

[0054] The spatial attention branch takes channel-reweighted features as input: , in, This indicates element-wise multiplication.

[0055] Subsequently, first through a depth-separable Convolution (DWConv) followed by SiLU activation is used to capture local structural patterns; then, respectively... and Anisotropic convolutions are used to emphasize defect patterns in the horizontal and vertical directions, respectively.

[0056]

[0057]

[0058] Finally, the spatial attention map is defined as:

[0059] The final calibration feature obtained is as follows:

[0060] 3) Cross-Branch Fusion: In addition to local refinement within each EffusionNeck, SFCF introduces a cross-branch feature fusion path to aggregate information from all intermediate refined features.

[0061] set up This is the result of concatenating the features from the intermediate layer. Then, through a... Convolutional computation for cross-branch fusion features:

[0062] Finally, this fused feature is added back to the last level of refinement features. superior:

[0063] This allows information from the early refinement stages to be passed to the final output branch. Finally, the features of all branches are concatenated and processed through a... Convolution maps to the desired output dimension:

[0064] In some embodiments of the present invention, the preprocessing includes: performing size normalization and pixel normalization on the image to generate preprocessed input image data.

[0065] Specifically, in this embodiment, images of transmission line insulators are acquired using a drone inspection system. The original images are then normalized in size to meet the network input requirements. Simultaneously, pixel values ​​are normalized to reduce the impact of different lighting conditions on the model's inference results. The preprocessed images serve as the network's input data.

[0066] In some embodiments of the present invention, dividing the input features into first branch features and second branch features along the channel dimension includes: performing channel compression and or channel rearrangement processing on the input features and then dividing them along the channel dimension.

[0067] In some embodiments of the present invention, the multi-level processing of the second branch feature includes: sequentially inputting the second branch feature into at least one feature processing unit to update the second branch feature in each feature processing unit and output the corresponding intermediate feature information.

[0068] In some embodiments of the present invention, the attention-based calibration includes: calculating channel attention weights and spatial attention weights based on input features at each level, and performing the calibration by multiplying the channel attention weights and spatial attention weights element-wise with the input features at each level.

[0069] In some embodiments of the present invention, the calculation of the spatial attention weights includes: performing depthwise separable convolution processing on the input features, and performing anisotropic convolution processing along the horizontal and vertical directions respectively to generate spatial attention weights.

[0070] In some embodiments of the present invention, the fusion of intermediate feature information generated by the multi-level processing and the combination of the fused information with the first branch feature includes: concatenating the intermediate feature information at each level and generating a fused feature through convolution transformation; and concatenating and / or weighting the fused feature with the first branch feature and outputting the enhanced feature.

[0071] Specifically, in this embodiment, the preprocessed image is input into the backbone network. During the multi-scale feature extraction process in the backbone network, the Conv convolution module and the SCDown downsampling module are used to progressively reduce the feature map resolution and expand the receptive field; the SFCF feature enhancement module performs structured splitting, recursive enhancement, and cross-branch fusion processing on features at multiple scales; the SPPF block is used to aggregate multi-scale contextual information; and the PSA module is used to perform attention weighting on features, thereby highlighting important feature responses related to insulator defects. Finally, the backbone network outputs multi-scale feature maps for use by the subsequent neck network.

[0072] In some embodiments of the present invention, the feature fusion and enhancement network includes: a top-down upsampling fusion path and a bottom-up downsampling fusion path, wherein the upsampling fusion path and the downsampling fusion path achieve the fusion of features at different scales through feature splicing, and output the multi-scale fusion features at least two scales.

[0073] Furthermore, the multi-scale features output from the backbone network are fused with the input features and enhanced by the network, including the Upsample1 upsampling module, the Concat1 feature concatenation module, the SFCF5 feature enhancement module, the Upsample2 upsampling module, the Concat2 feature concatenation module, the SFCF6 feature enhancement module, the Conv4 module, the Concat3 feature concatenation module, the SFCF7 feature enhancement module, the SCDown3 downsampling module, the Concat4 feature concatenation module, and the C2fCIB module.

[0074] The upsampling module is used for scale recovery of high-level features, the feature splicing module is used to fuse feature information from different levels, and the SFCF feature enhancement module is used to further enhance the expressive ability of local structural anomalies at the fusion node. After the neck network processing described above to achieve feature fusion and enhancement, fused features suitable for defect detection at different scales are obtained.

[0075] The multi-scale features output by the neck network are input into the YOLOv10 detection head.

[0076] The detection head performs regression and classification predictions at various scales, including: The regression branch is used to predict the bounding box location of candidate targets and outputs the bounding box parameters using a distributed regression approach; The classification branch is used to predict the defect category to which the candidate target belongs and its corresponding confidence level.

[0077] By decoupling regression and classification tasks, mutual interference between different prediction tasks is reduced, and detection stability is improved.

[0078] In some embodiments of the present invention, filtering the prediction results according to a preset threshold includes: removing low-confidence candidate boxes based on a confidence threshold, and performing overlap suppression processing on the remaining candidate boxes; the overlap suppression processing retains the target candidate boxes based on the overlap threshold between candidate boxes, and generates the insulator defect detection information.

[0079] Specifically, in this implementation, the regression prediction results output by the detection head are decoded and converted into candidate bounding boxes in the actual image coordinate system; at the same time, the candidate detection results are initially screened based on the confidence level output by the classification branch, and candidate results with confidence levels higher than a preset threshold are retained.

[0080] Subsequently, overlap suppression processing is performed on candidate detection results with highly overlapping spatial locations, retaining only the optimal detection result to obtain the final insulator defect detection output.

[0081] The output includes the bounding box location, defect category, and corresponding confidence information of insulator partial missing defects, which can be used for defect location, alarm, and subsequent operation and maintenance decision support in the transmission line inspection system.

[0082] Compared with the prior art, the beneficial effects of the present invention are as follows: 1) Structured feature enhancement modules are introduced at several key locations in the backbone network and the feature fusion and enhancement network, enabling features to be not only extracted but also repeatedly enhanced and corrected during the hierarchical transmission process. By retaining the original feature branches and performing multi-stage enhancement processing on another branch, local structural anomalies can be continuously monitored and amplified at different levels, thereby avoiding the loss of key information due to feature compression. Therefore, when facing fine-grained structural defects such as partial defects, this invention provides more stable detection results and a significantly reduced false negative rate, thus achieving more stable detection and fewer false negatives for partial defects in insulators.

[0083] 2) An attention mechanism that simultaneously focuses on channel information and spatial structure information is introduced during feature enhancement, further strengthening the modeling ability for local structure and directional features. This enables the network to not only focus on salient regions but also respond to subtle but structurally significant anomalous patterns. Consequently, even under conditions of weak texture and low contrast, defect-related features can still be effectively highlighted, improving overall detection robustness.

[0084] 3) By introducing a cross-branch fusion mechanism in the feature enhancement module, feature information generated at different enhancement stages can be centrally utilized and fed back to the final output branch, thereby explicitly preserving structural cues at different levels in the output features. This approach avoids the dilution of key structural information during the fusion process, making the utilization of multi-scale features more thorough and effective.

[0085] 4) By rationally designing the feature enhancement and fusion structure, targeted enhancement of key features is achieved while maintaining controllable overall network depth and complexity. This approach improves feature representation quality without significantly increasing computational resource consumption, achieving a good balance between detection accuracy and computational efficiency, making it more suitable for practical applications such as UAV inspection.

[0086] In summary, this invention introduces multi-stage structured enhancement and targeted attention mechanisms during feature extraction and fusion, enabling the network to stably perceive local structural anomalies corresponding to partial insulator defects under complex backgrounds and weak texture conditions. This not only improves detection accuracy and reliability but also takes into account computational efficiency, demonstrating significant engineering application value.

[0087] Figure 4 This is a flowchart illustrating a defect detection system for power transmission line insulators according to an embodiment of the present invention.

[0088] Example 2, as Figure 4 As shown, the present invention also provides a defect detection system for power transmission line insulators, comprising: an image acquisition module S11, a backbone feature extraction module S12, a feature fusion enhancement module S13, a defect prediction module S14, and a result screening module S15.

[0089] The image acquisition module is used to acquire inspection images of insulators to be inspected in the transmission line and input the inspection images into the backbone network of the cascaded structure. A backbone feature extraction module is used to process the inspection image along a cascaded path and output multi-scale features. The backbone network includes multiple SFCF feature enhancement modules. The multi-scale features are stage features output by the backbone network at at least two different network stages, and each stage feature is the output of the preceding cascaded modules. The process executed by each SFCF feature enhancement module is as follows: the output features of the preceding cascaded modules are divided in the channel dimension into a first branch feature for preserving original information and a second branch feature for deep feature transformation; the second branch feature undergoes multi-level processing, which is implemented by at least one cascaded feature processing unit within a single SFCF feature enhancement module, wherein each level of processing sequentially includes convolutional transformation and attention-based calibration; intermediate feature information generated by the multi-level processing is fused, and the fused information is combined with the first branch feature to output an enhanced stage feature map. The feature fusion and enhancement module is used to input the multi-scale features output by the backbone network into the feature fusion and enhancement network, and to fuse features of different scales through upsampling, downsampling and feature concatenation to obtain multi-scale fused features; wherein, after at least one fusion node of the feature fusion and enhancement network, at least one SFCF feature enhancement module is set to perform feature enhancement operation on the fused features; The defect prediction module is used to input multi-scale fused features into the detection head to predict the bounding box location, defect category, and confidence level of the defect category of the insulator, and obtain the insulator defect prediction result. The result filtering module is used to filter the prediction results of the defect prediction module according to a preset threshold and generate insulator defect detection information.

[0090] Example 3: The present invention also provides a transmission line insulator defect detection device, the device including a computer device, the computer device including a processor and a memory, the processor storing computer instructions, and when the computer instructions are executed, the device implements the insulator defect detection method.

[0091] Example 4, as Figure 5 As shown, the present invention also provides an electronic device 100 for implementing an insulator defect detection method.

[0092] The electronic device 100 includes a memory 101, at least one processor 102, a computer program 103 stored in the memory 101 and executable on at least one processor 102, and at least one communication bus 104.

[0093] The memory 101 can be used to store the computer program 103. The processor 102 implements the steps of the insulator defect detection method described in the first aspect of the present invention by running or executing the computer program stored in the memory 101 and calling the data stored in the memory 101.

[0094] The memory 101 may primarily include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created based on the use of the electronic device 100 (such as audio data), etc. In addition, the memory 101 may include non-volatile memory, such as hard disk, RAM, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other non-volatile solid-state storage device.

[0095] At least one processor 102 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Processor 102 may be a microprocessor or any conventional processor. Processor 102 is the control center of electronic device 100, connecting various parts of electronic device 100 via various interfaces and lines.

[0096] The memory 101 in the electronic device 100 stores multiple instructions to implement a method for detecting defects in power transmission line insulators, and the processor 102 can execute multiple instructions to achieve the following: An inspection image of an insulator to be inspected in a transmission line is acquired. This image is then input into a cascaded backbone network. The image is processed along the cascaded path to output multi-scale features. The backbone network includes multiple SFCF feature enhancement modules. The multi-scale features are stage features output by the backbone network at at least two different network stages, and each stage feature is the output of the preceding cascaded modules. Each SFCF feature enhancement module executes the following process: the output features of the preceding cascaded modules are divided in the channel dimension into a first branch feature for preserving original information and a second branch feature for deep feature transformation; the second branch feature undergoes multi-level processing, which is implemented by at least one cascaded feature processing unit within a single SFCF feature enhancement module. Each level of processing sequentially includes convolutional transformation and attention-based calibration; intermediate feature information generated by the multi-level processing is fused, and the fused information is combined with the first branch feature to output an enhanced stage feature map. The multi-scale features output from the backbone network are input into the feature fusion and enhancement network. The features at different scales are fused through upsampling, downsampling, and feature concatenation to obtain multi-scale fused features. At least one SFCF feature enhancement module is set after at least one fusion node of the feature fusion and enhancement network to perform feature enhancement operations on the fused features. The multi-scale fusion features are input into the detection head to predict the bounding box position, defect category, and confidence level of the defect category of the insulator, thereby obtaining the insulator defect prediction result. The prediction result is then filtered according to a preset threshold to generate insulator defect detection information.

[0097] Example 5: If the modules / units integrated in the electronic device 100 are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, and read-only memory (ROM).

[0098] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0099] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0100] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0101] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0102] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0103] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for detecting defects in power transmission line insulators, characterized in that, The method includes: An inspection image of an insulator to be inspected in a transmission line is acquired. This image is then input into a cascaded backbone network. The image is processed along the cascaded path to output multi-scale features. These multi-scale features are stage features output by the backbone network at at least two different network stages, and each stage feature is the output of the preceding cascaded modules. The backbone network includes multiple SFCF feature enhancement modules. Each SFCF feature enhancement module executes the following process: the output features of the preceding cascaded modules are divided in the channel dimension into a first branch feature for preserving original information and a second branch feature for deep feature transformation; the second branch feature undergoes multi-level processing, which is implemented by at least one cascaded feature processing unit within a single SFCF feature enhancement module. Each level of processing sequentially includes convolutional transformation and attention-based calibration; intermediate feature information generated by the multi-level processing is fused, and the fused information is combined with the first branch feature to output an enhanced stage feature map. The multi-scale features output from the backbone network are input into the feature fusion and enhancement network. The features at different scales are fused through upsampling, downsampling, and feature concatenation to obtain multi-scale fused features. At least one SFCF feature enhancement module is set after at least one fusion node of the feature fusion and enhancement network to perform feature enhancement operations on the fused features. The multi-scale fusion features are input into the detection head to predict the bounding box position, defect category, and confidence level of the defect category of the insulator, thereby obtaining the insulator defect prediction result. The prediction result is then filtered according to a preset threshold to generate insulator defect detection information.

2. The method for detecting defects in transmission line insulators according to claim 1, characterized in that, The backbone network is a cascaded structure, which includes at least a first convolutional module, a second convolutional module, a first SFCF feature enhancement module, a third convolutional module, a second SFCF feature enhancement module, a first downsampling module, a third SFCF feature enhancement module, a second downsampling module, a fourth SFCF feature enhancement module, a spatial pyramid aggregation module, and an attention enhancement module.

3. The method for detecting defects in transmission line insulators according to claim 1, characterized in that, The step of dividing the output features of the preceding cascaded module into first branch features and second branch features in the channel dimension includes: After performing channel compression and channel rearrangement on the input features, they are divided along the channel dimension to obtain the first branch features and the second branch features.

4. The method for detecting defects in transmission line insulators according to claim 1, characterized in that, The multi-level processing of the second branch features includes: The second branch features are sequentially input into at least two cascaded feature processing units to update the second branch features and output the corresponding intermediate feature information in each feature processing unit.

5. The method for detecting defects in transmission line insulators according to claim 4, characterized in that, The attention-based calibration includes: After the convolution transformation, based on the features output by the convolution transformation, channel attention weights and spatial attention weights are calculated, and the calibration is performed by multiplying the channel attention weights and spatial attention weights element by element with the input features at each level.

6. The method for detecting defects in transmission line insulators according to claim 5, characterized in that, The calculation of the spatial attention weights includes: The input features are subjected to depthwise separable convolution, and anisotropic convolution is performed along the horizontal and vertical directions respectively to generate spatial attention weights.

7. The method for detecting defects in transmission line insulators according to claim 1, characterized in that, The process of fusing intermediate feature information generated by the multi-level processing and combining the fused information with the first branch feature includes: The intermediate feature information at each level is concatenated and transformed by convolution to generate a fused feature; then, the fused feature is concatenated and / or weighted and superimposed with the first branch feature to output the enhanced feature.

8. The method for detecting defects in transmission line insulators according to claim 1, characterized in that, The feature fusion and enhancement network includes: The system employs a top-down upsampling fusion path and a bottom-up downsampling fusion path. The upsampling fusion path and the downsampling fusion path fuse features at different scales through feature concatenation, and output the multi-scale fusion features at least two scales.

9. The method for detecting defects in transmission line insulators according to claim 1, characterized in that, The prediction results are filtered based on preset thresholds, including: Low-confidence candidate boxes are eliminated based on a confidence threshold, and overlap suppression processing is performed on the remaining candidate boxes. The overlap suppression processing retains the target candidate boxes based on the overlap threshold between candidate boxes, thereby generating the insulator defect detection information.

10. A defect detection device for power transmission line insulators, characterized in that, The device includes a computer device, which includes a processor and a memory. The processor stores computer instructions. When the computer instructions are executed, the device implements the method for detecting defects in power transmission line insulators as described in any one of claims 1 to 9.

Images