Insulator defect detection method based on improved YOLO cross-scale fusion and related equipment

Abstract

The disclosure provides an insulator defect detection method based on improved YOLO cross-scale fusion and related equipment. The method comprises: acquiring an insulator image of a power transmission line, and inputting the insulator image into a pre-trained defect detection model; using the defect detection model, performing feature extraction processing on the insulator image to obtain an insulator feature map; using an attention feedback module in the defect detection model, determining an attention feedback feature according to the insulator feature map; obtaining a multi-scale insulator feature by fusing the attention feedback feature; and using the defect detection model, determining an insulator defect detection result according to the multi-scale insulator feature.

Description

Technical Field

[0001] This disclosure relates to the field of data processing technology, and in particular to an insulator defect detection method and related equipment based on improved YOLO cross-scale fusion. Background Technology

[0002] Insulators are fundamental components of power transmission lines, and the detection of insulator defects is crucial for the stable operation of these lines. However, manual inspections of insulators for visual checks and insulation tests suffer from low accuracy in identifying insulator defects.

[0003] Therefore, improving the accuracy of insulator defect identification in transmission lines has become an urgent technical problem to be solved. Summary of the Invention

[0004] In view of this, the purpose of this disclosure is to propose an insulator defect detection method and related equipment based on improved YOLO cross-scale fusion to solve or partially solve the above-mentioned technical problems.

[0005] To achieve the above objectives, the first aspect of this disclosure proposes an insulator defect detection method based on improved YOLO cross-scale fusion, the method comprising:

[0006] Acquire images of insulators of the transmission line and input the insulator images into a pre-trained defect detection model; Using the aforementioned defect detection model, feature extraction processing is performed on the insulator image to obtain an insulator feature map; The attention feedback module in the defect detection model is used to determine the attention feedback features based on the insulator feature map; Multi-scale insulator features are obtained by fusing the attention feedback features. Using the aforementioned defect detection model, the insulator defect detection results are determined based on the multi-scale insulator characteristics.

[0007] Based on the same inventive concept, a second aspect of this disclosure proposes an insulator defect detection device based on improved YOLO cross-scale fusion, comprising: An insulator image acquisition module is configured to acquire insulator images of transmission lines and input the insulator images into a pre-trained defect detection model; The feature extraction and processing module is configured to use the defect detection model to perform feature extraction processing on the insulator image to obtain an insulator feature map; The attention feedback feature determination module is configured to determine attention feedback features based on the insulator feature map using the attention feedback module in the defect detection model. The fusion processing module is configured to obtain multi-scale insulator features by fusing the attention feedback features; The defect detection result determination module is configured to use the defect detection model to determine the insulator defect detection result based on the multi-scale insulator characteristics.

[0008] Based on the same inventive concept, a third aspect of this disclosure proposes an electronic device including a memory, a processor, and a computer program stored in the memory and executable by the processor, wherein the processor implements the method described above when executing the computer program.

[0009] Based on the same inventive concept, a fourth aspect of this disclosure provides a non-transitory computer-readable storage medium that stores computer instructions for causing a computer to perform the methods described above.

[0010] As described above, this disclosure presents an insulator defect detection method based on improved YOLO cross-scale fusion. The method involves acquiring insulator images of transmission lines and inputting them into a pre-trained defect detection model. The model then extracts features from the insulator images to obtain insulator feature maps. An attention feedback module within the defect detection model determines attention feedback features based on the insulator feature maps. This module allocates limited computational resources to more important information in the current insulator defect detection task, reducing or even filtering out attention to irrelevant information. Multi-scale insulator features are obtained through fusion processing of the attention feedback features, enhancing the semantic information of lower-level features. Finally, the defect detection model determines the insulator defect detection result based on these multi-scale features. This approach enables the defect detection model to possess adaptability and robustness, meeting the requirements of insulator defect detection in different scenarios. It improves the accuracy of insulator defect identification in transmission lines, reduces the probability of missed and false detections, and provides strong protection for the safe operation of transmission lines. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in this disclosure or related technologies, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart of an embodiment of the insulator defect detection method based on improved YOLO cross-scale fusion according to this disclosure; Figure 2This is a schematic diagram of the attention feedback module according to an embodiment of the present disclosure; Figure 3 This is a schematic diagram of the cross-scale feature fusion module according to an embodiment of the present disclosure; Figure 4 This is a schematic diagram of the structure of the insulator defect detection device based on improved YOLO cross-scale fusion according to an embodiment of this disclosure; Figure 5 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present disclosure. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this disclosure clearer, the following detailed description is provided in conjunction with specific embodiments and the accompanying drawings.

[0014] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this disclosure should have the ordinary meaning understood by one of ordinary skill in the art to which this disclosure pertains. The terms "first," "second," and similar terms used in the embodiments of this disclosure do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0015] Based on the background description, with the continuous development of the economy, the electricity consumption demand of industrial production and urban residents is increasing day by day, and the power industry provides a solid guarantee for economic development. Along with the increase in power generation, the safety of the power system has become increasingly important. Transmission lines play a crucial role in the safety of the power system. Faults in transmission lines can cause localized power outages, disrupting normal production and life; in severe cases, they can lead to power system oscillations, disconnection, or even large-scale power outages, causing significant impacts on society. In the operation and maintenance of transmission lines, power line inspection is a core link, allowing for timely intervention in potential hazards and defects in equipment. Insulators are a key component of transmission lines, providing mechanical fixation, reinforcement, creepage distance, and electrical insulation. However, due to environmental factors (e.g., intense sunlight, heavy rain, typhoons, lightning, ice accumulation), climate, and service life, insulators are prone to problems such as string slippage, spontaneous explosion, and contamination, resulting in insulator defects. These defects can damage the structure of the transmission line, causing it to lose its insulation capacity, and consequently leading to widespread power outages or even catastrophic accidents, such as forest fires. Therefore, studying insulator fault detection algorithms to conduct power line insulator inspections and promptly identify potential insulator hazards and defects has significant theoretical and practical value.

[0016] In recent years, with the improvement of computer hardware and the rapid development of deep learning theory, deep learning has been widely applied in many fields, and has also made significant progress in insulator defect detection. Key advancements include: Amplified Receptive Field Network (ARFNet) can adjust the receptive field size of convolutional layers to adapt to the multi-scale characteristics of insulator defects, thus improving detection accuracy. A lightweight deep learning model was constructed, and its architecture was optimized for insulator fault features. A morphological method-deep learning-based insulator defect detection algorithm achieved high accuracy, but suffers from high computational complexity and cumbersome operation. An improved YOLOv4-based insulator self-explosion detection method, incorporating feature fusion and activation network modules, effectively improved insulator defect recognition. A semantic segmentation-based insulator defect detection method achieved good recognition accuracy and performance. Enhancement operations on the insulator dataset using a dehazing algorithm address the problem of decreased accuracy in foggy insulator defect recognition. Furthermore, attention mechanisms and feature reuse modules were integrated into the defect detection model, thereby improving the overall accuracy of insulator defect recognition. A cross-modal insulator defect detection algorithm is designed with cross-modal fusion and multi-scale enhancement modules. Results show that the proposed defect detection method can achieve good recognition accuracy. To address the problem of complex backgrounds and target overlap in insulator images, an insulator defect detection method embedding an Efficient Channel Attention Network (ECA-Net) attention mechanism is proposed, which can extract feature information from complex background images and small target features. An optimized YOLOv7 insulator defect detection method replaces the conventional convolutions in the backbone network with separable convolutions, improving defect detection accuracy and computational efficiency. An optimized YOLOv8 insulator fault type detection algorithm fuses a convolutional neural network (hostNet) with multi-scale asymmetric convolutions to establish a cross-stage partial 2f (C2f) module. Results show that this method can detect four types of insulator defects: normal, broken, damaged, and flashover. However, the small target detection accuracy of this method is low. A target detection method based on optimized YOLOv9 incorporates a generalized, efficient layer aggregation network and programmable gradient information, improving both accuracy and efficiency. A target detection method based on YOLOv10 employs a dual-label assignment strategy, a consistent matching metric, and a lightweight design to reduce computational complexity.The target detection method based on collaborative hybrid assignment training (Co-DETR) generates latent features through a backbone network and a Transformer encoder, thereby achieving high detection accuracy with a smaller model size. The Relation Detection Transformer (Relation-DETR) method proposes a novel positional relationship based on relative geometric features, achieving progressive attention refinement and improving the performance of insulator defect detection.

[0017] The above analysis shows that various insulator defect detection methods do not comprehensively consider the complex background of aerial images of insulators, such as background interference from power towers and buildings, nor do they account for the differences in string breakage defects across different scenarios. Therefore, designing a method that reduces computational complexity while maintaining adaptability to complex scenarios involving exposure, darkness, occlusion, and blur, thereby improving the accuracy of insulator string breakage defect detection, has become a key issue in insulator defect detection. In conclusion, to date, no relatively good insulator defect detection algorithm has been proposed for the attention-cross-scale fusion YOLOv7 method.

[0018] As mentioned above, improving the accuracy of insulator defect identification in transmission lines has become an important research issue.

[0019] Based on the above description, such as Figure 1 As shown, the insulator defect detection method based on improved YOLO cross-scale fusion proposed in this embodiment includes: Step 101: Obtain an image of the insulator of the transmission line and input the insulator image into a pre-trained defect detection model.

[0020] In practical implementation, insulators are key components in transmission lines, primarily used for insulating and mechanically supporting conductors. The process involves inputting insulator images into a pre-trained defect detection model, enabling the model to perform defect detection on insulators within the transmission line based on these images.

[0021] The defect detection model is obtained by pre-training an improved YOLOv7 model.

[0022] Step 102: Using the defect detection model, perform feature extraction processing on the insulator image to obtain an insulator feature map.

[0023] In practice, a defect detection model is used to extract features from the insulator image to obtain an insulator feature map. Specifically, the defect detection model is used to determine the insulator region from the insulator image, and then the corresponding insulator feature map is determined from the insulator image.

[0024] Step 103: Using the attention feedback module in the defect detection model, determine the attention feedback features based on the insulator feature map.

[0025] In practice, to address the issue of low detection accuracy of insulator defects in complex background images, an attention feedback model was constructed based on the YOLOv7 target detection algorithm. The attention feedback model can encode spatial features containing important channel information and pass them to the squeezing excitation network, thus solving the problem of spatial feature loss in the attention mechanism and enhancing the channel attention capability.

[0026] The attention feedback module is an improvement upon the squeeze-excitation network. It's a plug-and-play module that enables more powerful feature extraction. The attention feedback module allocates limited computational resources to more important information in the current object detection task, reducing or even filtering attention to irrelevant information and improving detection accuracy. The attention feedback module consists of an attention propagation path and an attention feedback path. The attention propagation path, constructed from the squeeze-excitation network, extracts the discriminative channel features of the network. The attention feedback path is the core of the attention feedback model; it extracts the spatial features enhanced by the squeeze-excitation network through convolution and activation, thereby improving how the network notices discriminative features.

[0027] The coordinate attention and convolutional block attention modules also introduce spatial attention and combine it with channel attention. These attention mechanisms can improve the accuracy of feature representation. However, the spatial attention is combined with channel attention in a series or parallel manner, while the attention feedback module feeds back spatial features carrying important channel information, so that the attention learns both spatial features and further enhances channel features.

[0028] Figure 2 This is a schematic diagram of the attention feedback module according to an embodiment of this disclosure. Figure 2 As shown, the insulator feature map is input into the attention feedback module, and a convolution operator is applied to the insulator feature map. The transformed feature map is obtained by performing transformation processing. Transform the feature map The feature map is used as the input feature map to the squeeze excitation network, and the squeeze excitation network is used to transform the feature map. Filtering is performed to obtain channel convolution features. Compress the channel convolution features ( ). ) to obtain channel compression characteristics Using the PSiLU activation function ( Determine the channel modulation weight vector based on the channel compression characteristics. Channel modulation weight vector Applying a broadcast method to the corresponding channel feature map Thus, the dimension recovery parameters output by the attention feedback module are obtained. Attention feedback features are determined based on dimensionality recovery parameters. .

[0029] Step 104: Multi-scale insulator features are obtained by fusing the attention feedback features.

[0030] In practice, the cross-scale feature fusion module can make full use of the enhanced multi-level features in the backbone network to not only generate multi-scale insulator features with high-level semantics and precise location information, but also improve the defect detection model's ability to detect targets of different scales.

[0031] The cross-scale feature fusion module employs multiple cross-scale paths to fuse features, further enhancing the defect detection model's utilization of both detailed and global information. It uses a bottom-up approach to upsample high-level features, enhancing the semantic information of low-level features. Furthermore, it downsamples low-level features using a top-down approach to enrich the detailed information of high-level features. Moreover, since cross-scale connections can provide multiple inputs, the cross-scale feature fusion module can fuse richer feature information without increasing cost.

[0032] Step 105: Using the defect detection model, determine the insulator defect detection result based on the multi-scale insulator characteristics.

[0033] In practice, the defect detection model can be a YOLOv7 target detector, a cross-scale fusion feature network. The YOLOv7 target detector is used to perform defect detection based on multi-scale insulator features to obtain the insulator defect detection results.

[0034] By performing the above steps to identify defects in multi-scale insulator features, an insulator defect detection method based on attention-cross-scale fusion YOLOv7 was realized.

[0035] Through the above embodiments, insulator images of transmission lines are acquired and input into a pre-trained defect detection model. The defect detection model performs feature extraction on the insulator images to obtain insulator feature maps. Using the attention feedback module in the defect detection model, attention feedback features are determined based on the insulator feature maps. This module allocates limited computational resources to more important information in the current insulator defect detection task, reducing or even filtering out attention to irrelevant information. Multi-scale insulator features are obtained by fusing the attention feedback features, which enhances the semantic information of low-level features. The defect detection model then determines the insulator defect detection result based on these multi-scale features. This approach enables the defect detection model to be adaptive and robust, meeting the requirements of insulator defect detection in different scenarios. It improves the accuracy of insulator defect identification in transmission lines, reduces the probability of missed and false detections, and provides strong protection for the safe operation of transmission lines.

[0036] In some embodiments, step 103 includes: Step 1031: Using the attention feedback module in the defect detection model, the insulator feature map is filtered to obtain channel convolution features, and channel compression features are determined based on the channel convolution features.

[0037] In practice, complex background noise in insulator images can interfere with the training process of convolutional neural networks (CNNs), weakening their ability to focus on discriminative channel features. Fundamentally, if we can further focus on the differences in essential image attributes between complex background regions and target regions, we can improve detection performance. To further enhance the difference in essential image attributes between unaltered and tampered regions in insulator image forgery detection, an attention feedback module is proposed.

[0038] It is worth noting that, unlike the sequential or parallel combination of spatial and channel attention methods used to enhance attention capabilities, the attention feedback module considers using the PSiLU activation function on the feedback path. This allows it to acquire spatially related features to update the input of the squeezed excitation network. While the squeezing operation can lead to the compression and loss of spatial features, the attention feedback module obtains enhanced spatial features on the feedback path to update the input information of the squeezed excitation network, reducing spatial feature loss and enhancing channel attention capabilities.

[0039] The insulator feature map is transformed using the attention feedback module in the defect detection model. The channel convolution features are obtained by filtering the transformed feature map. Global average pooling is applied to the channel convolutional features to obtain the global average features. The median features are obtained by performing median pooling on the channel convolution features. Channel compression features are obtained by averaging the global average features and median features. .

[0040] Step 1032: Determine the channel modulation weight vector based on the channel compression characteristics using the activation function.

[0041] In practice, a squeezing excitation network is used to compress the channel characteristics. The channel modulation weight vector is obtained by performing an excitation operation. .

[0042] Step 1033: Multiply the channel modulation weight vector and the channel convolution feature to obtain the dimension recovery parameter, and determine the attention feedback feature based on the dimension recovery parameter.

[0043] In practical implementation, the channel modulation weight vector Applying a broadcast method to the corresponding channel feature map Thus, the dimension recovery parameters output by the attention feedback module are obtained. Specifically, the dimension recovery parameters are obtained by multiplying the channel modulation weight vector and the channel convolution features. ,in, Parameters for dimension recovery, For channel modulation weight vectors, These are channel convolution features.

[0044] like Figure 2 As shown, after the attention feedback module generates spatial features through 3×3 convolution and SiLU activation function, the response value is applied to the feature map. . The dimension is After PSiLU activation function The dimension is The formula for calculating the constructed feedback path is as follows: ,in, The attention feedback features generated for the attention feedback path reflect the response intensity of the input features at different spatial locations. and These are the input and output of the attention feedback module, respectively. It is the first Layer weights It's a linear mapping, changed through 3×3 convolution. Dimensions It is the SiLU activation function.

[0045] The spatial discrimination information extracted from the feedback path is fed back into the attention propagation path to process the input features. Element-wise modulation or residual fusion is performed to obtain updated features, which are then used as input to the squeeze excitation network to generate more accurate channel modulation weight vectors, thereby mitigating the spatial information loss caused by the squeeze operation and enhancing channel attention capabilities.

[0046] The above scheme utilizes the attention feedback module in the defect detection model to filter the insulator feature map, obtaining channel convolutional features, thus achieving preliminary screening and integration of the insulator feature map. Channel compression features are determined based on the channel convolutional features, enabling further refinement and dimensionality reduction. An activation function is used to determine the channel modulation weight vector based on the channel compression features, giving the defect detection model adaptive capabilities. This allows for flexible allocation of attention resources according to the characteristics and defect conditions of different insulator images, significantly improving the targeting and accuracy of insulator defect detection. Multiplying the channel modulation weight vector and the channel convolutional features yields dimensionality recovery parameters. These parameters are then used to determine the attention feedback features, achieving feature dimensionality transformation and information enhancement, thereby strengthening the defect detection model's ability to perceive insulator defects.

[0047] In some embodiments, step 1031 includes: Step 1031A: Using the attention feedback module in the defect detection model, the insulator feature map is transformed to obtain a transformed feature map, and the transformed feature map is filtered to obtain channel convolution features.

[0048] In practice, within the attention feedback module, for each attention propagation path, a convolution operator is first applied to the input insulator feature map. The transformed feature map is obtained by performing feature map transformation. .

[0049] Then transform the feature map The feature map is used as the input feature map to the squeeze excitation network, and the squeeze excitation network is used to transform the feature map. Filtering is performed to obtain channel convolutional features. ,in, For the first Each channel convolutional feature For filters, To transform the feature map, For the first The parameters of each filter, the set of filters is represented as The set after convolution is represented as To utilize information from all channels, compression and activation operations are performed on the channel convolutional features to obtain channel compressed features and attention feedback features.

[0050] Step 1031B: Perform global average pooling on the channel convolutional features to obtain global average features, perform median pooling on the channel convolutional features to obtain median features, and perform mean processing on the global average features and the median features to obtain channel compressed features.

[0051] In practice, because convolution operations need to be performed on the convolutional features of each channel, spatial relationships become mixed up. Therefore, a squeezing operation is used to further compress the spatial features to obtain channel compressed features, thereby extracting this mixed feature information and enabling the network to directly learn channel relationships.

[0052] Channel compressed features are obtained by squeezing the channel convolutional features using a squeezing excitation network. Specifically, global average pooling is performed on the channel convolutional features to obtain global average features, median pooling is performed on the channel convolutional features to obtain median features, and the mean of the global average features and median features is applied to obtain channel compressed features.

[0053] The formula for calculating the extrusion operation is: ,in, This is a channel compression feature. For channel convolution features, It is a global average feature. This is a median feature. This represents the height of the insulator feature map in the spatial dimension. This represents the width of the insulator feature map in the spatial dimension. For the insulator characteristic map in the th individual channels, spatial locations eigenvalues, The insulator feature map is compressed and converted into The channel convolutional features of each channel in the spatial dimension are encoded into a global channel descriptor, so that the global average features in space can continue to be learned by the network but have the effect of channel characteristic response.

[0054] The above scheme utilizes the attention feedback module in the defect detection model to transform the insulator feature map, obtaining a transformed feature map. This transformed feature map is then filtered to obtain channel convolutional features, achieving preliminary screening and integration of the insulator feature map. Global average pooling is then applied to the channel convolutional features to obtain global average features, median pooling to obtain median features, and finally, averaging of the global average and median features yields channel compressed features. This further refines and reduces the dimensionality of the channel convolutional features.

[0055] In some embodiments, step 1032 includes: Step 1032A: Use the first hyperparameter to perform dimensionality reduction processing on the channel compression features to obtain intermediate variables.

[0056] In practice, the first hyperparameter is used. Channel compression features Dimensionality reduction is performed to obtain intermediate variables. ,in, As an intermediate variable, As the first hyperparameter, This is a channel compression feature.

[0057] Step 1032B: Use the first activation function to perform a nonlinear mapping on each element in the intermediate variable to obtain the mapping feature.

[0058] In practice, the first activation function is used to apply the intermediate variables. Each element in the matrix is ​​nonlinearly mapped to obtain the mapped features. ,in, For mapping features, This is the first activation function (PSiLU activation function). As the first hyperparameter, This is a channel compression feature.

[0059] To achieve better stability and efficient feature extraction capabilities, the first activation function is the PSiLU activation function. The PSiLU activation function has buffering and stability for feature extraction and is less likely to cause confusion due to defects in similar insulators.

[0060] The formula for calculating the PSiLU activation function is: ,in, For intermediate variables Perform a nonlinear mapping of the PSiLU activation function. As an intermediate variable, As the first learnable parameter, As the second learnable parameter, To process the maximum value, To obtain the minimum value, the first learnable parameter is... Second learnable parameter Used to modulate the response intensity of the positive and negative half-axis respectively. When When the PSiLU activation function is degenerated into the SiLU activation function, compatibility with the current network structure is ensured.

[0061] Step 1032C: Use the second hyperparameter to perform dimensionality restoration processing on the mapped features to obtain the channel weight representation.

[0062] In practice, the second hyperparameter is used to perform dimensionality restoration processing on the mapped features to obtain the channel weight representation. ,in, This is represented by channel weights. This is the second hyperparameter. This is the first activation function (PSiLU activation function). As the first hyperparameter, This is a channel compression feature.

[0063] Step 1032D: Determine the channel modulation weight vector based on the channel weight representation using the second activation function.

[0064] In practice, the second activation function is used to determine the channel modulation weight vector based on the channel weight representation. ,in, For channel modulation weight vectors, For the second activation function, This is the second hyperparameter. This is the first activation function (PSiLU activation function). As the first hyperparameter, This is a channel compression feature.

[0065] The channel modulation weight vector is determined based on the channel compression characteristics using the PSiLU activation function. ,in, For channel modulation weight vectors, The PSiLU activation function is used. This is a channel compression feature. For hyperparameters, To compress features from the channel To channel modulation weight vector The intermediate computation results of the two-layer mapping (MLP / Conv1×1 form), , To reduce the hyperparameters, The default value is 2. This is the SiLU activation function.

[0066] The PSiLU activation function is a simple self-gating mechanism that takes a global channel descriptor as input and produces a set of modulation weights for each channel. Therefore, the self-gating mechanism can control the transmission of information within the network and fully leverage the advantages of attention-based feedback models.

[0067] The above scheme uses a first hyperparameter to reduce the dimensionality of the channel compressed features to obtain intermediate variables. A first activation function is then used to perform a nonlinear mapping on each element of the intermediate variables to obtain mapped features. A second hyperparameter is used to restore the dimensionality of the mapped features to obtain the channel weight representation. The second activation function is then used to determine the channel modulation weight vector based on the channel weight representation. The first activation function is the PSiLU activation function, which provides buffering and stability to feature extraction, reducing the likelihood of confusion caused by defects in similar insulators.

[0068] In some embodiments, step 104 includes: Step 1041: Determine initial multi-scale features from the attention feedback features.

[0069] In practice, during the cross-scale feature fusion process, the attention feedback features extracted from the third to fifth stages of the backbone network after being enhanced by the attention feedback module are used as the initial multi-scale features.

[0070] The initial multi-scale features include: initial high-resolution features. Initial medium resolution features and initial low-resolution features .

[0071] Step 1042: The initial multi-scale features are fused to obtain intermediate multi-scale features.

[0072] In practical implementation, to address the challenge of accurately detecting insulator defects in real-world power transmission line scenarios, separable quantum dilated convolution was used instead of conventional convolution. This separable quantum dilated convolution operation assigns corresponding quantum information to each pixel of the input tensor before each convolution, effectively improving the accuracy of multi-scale insulator feature extraction and localization regression in the defect detection model, thereby enhancing the detection and localization precision of power transmission line insulator defects.

[0073] Figure 3 This is a schematic diagram of the cross-scale feature fusion module according to an embodiment of this disclosure. Figure 3 As shown, for the initial high-resolution features Upsampling is performed to obtain high-resolution features of the target. For the first intermediate feature and initial medium resolution features Upsampling is performed to obtain the second intermediate feature. .

[0074] Step 1043: The intermediate multi-scale features are fused to obtain the target multi-scale features, and the target multi-scale features are used as multi-scale insulator features.

[0075] In specific implementation, such as Figure 3 As shown, for the initial low-resolution features The first intermediate feature is obtained by upsampling. For the second intermediate feature Perform upsampling and refine the initial medium-resolution features. High-resolution features of the target Downsampling is performed to obtain the target's medium-resolution features. For the first intermediate feature Perform upsampling and target medium-resolution features High-resolution features of the target Downsampling is performed to obtain low-resolution features of the target. .

[0076] The above scheme determines initial multi-scale features from attention feedback features. The initial multi-scale features are then fused to obtain intermediate multi-scale features. These intermediate multi-scale features are then fused to obtain target multi-scale features, which are then used as the multi-scale insulator features. This effectively improves the accuracy of the defect detection model in extracting and locating multi-scale insulator features, thereby enhancing the detection and location accuracy of transmission line insulator defects.

[0077] In some embodiments, step 1042 includes: Step 1042A: Determine the initial high-resolution features, initial medium-resolution features, and initial low-resolution features from the initial multi-scale features.

[0078] In practice, the initial high-resolution features are determined from the initial multi-scale features. Initial medium resolution features and initial low-resolution features .

[0079] Step 1042B: Perform convolution processing on the initial low-resolution features to obtain the first intermediate features.

[0080] In practice, the initial low-resolution features The first intermediate feature is obtained by performing convolution. The formula for calculating the first intermediate feature is: ,in, As the first intermediate feature, This is a separable quantum dilated convolution operation. These are the initial low-resolution features.

[0081] Step 1042C: Upsample the first intermediate feature to obtain a first upsampled feature; fuse the initial high-resolution feature, the initial medium-resolution feature, and the first upsampled feature to obtain a first fused feature; and convolve the first fused feature to obtain a second intermediate feature.

[0082] In specific implementation, the first intermediate feature The first upsampled feature is obtained by performing upsampling processing. For initial high-resolution features Initial medium resolution features and the first upsampled feature The first fusion feature is obtained by performing fusion processing. For the first fusion feature The second intermediate feature is obtained by performing convolution.

[0083] The formula for calculating the second intermediate feature is: ,in, As the second intermediate feature, This is a separable quantum dilated convolution operation. For initial high-resolution features, This represents the initial medium-resolution features. For upsampling processing, This is the first intermediate feature.

[0084] The first upsampled feature after upsampling is a high semantic feature, while the initial medium-resolution feature and the initial low-resolution feature are detail features. The high semantic feature and the detail feature are connected and fused to obtain an intermediate fused feature that combines semantic information and texture details.

[0085] The above scheme determines initial high-resolution, initial medium-resolution, and initial low-resolution features from the initial multi-scale features. The initial low-resolution features are convolved to obtain the first intermediate feature. The first intermediate feature is upsampled to obtain the first upsampled feature. The initial high-resolution, initial medium-resolution, and first upsampled features are fused to obtain the first fused feature. The first fused feature is then convolved to obtain the second intermediate feature. In this way, by upsampling the high-level features, the semantic information of the low-level features can be enhanced.

[0086] In some embodiments, step 1043 includes: Step 1043A: Upsample the second intermediate feature to obtain the second upsampled feature, fuse the initial high-resolution feature and the second upsampled feature to obtain the second fused feature, and convolve the second fused feature to obtain the target high-resolution feature.

[0087] In specific implementation, the second intermediate feature The second upsampled feature is obtained by performing upsampling processing. For initial high-resolution features Second upsampling features The second fusion feature is obtained by performing fusion processing. For the second fusion feature Convolution processing is performed to obtain high-resolution features of the target.

[0088] The formula for calculating the high-resolution features of the target is: ,in, For high-resolution features of the target, This is a separable quantum dilated convolution operation. For initial high-resolution features, For upsampling processing, This is the second intermediate feature.

[0089] Step 1043B: Downsample the target high-resolution features to obtain a first downsampled feature; fuse the initial medium-resolution features, the second intermediate features, and the first downsampled features to obtain a third fused feature; and convolve the third fused feature to obtain the target medium-resolution features.

[0090] In practice, high-resolution features of the target The first downsampled feature is obtained by performing downsampling processing. For initial medium resolution features Second intermediate features and the first downsampling feature The third fusion feature is obtained by performing fusion processing. Regarding the third fusion feature Convolution processing is performed to obtain the medium-resolution features of the target.

[0091] The formula for calculating the mid-resolution features of the target is: ,in, For the target's medium-resolution features This is a separable quantum dilated convolution operation. This represents the initial medium-resolution features. As the second intermediate feature, For downsampling processing, High-resolution features for the target.

[0092] Step 1043C: Downsample the target's medium-resolution features to obtain a second downsampled feature; fuse the initial high-resolution features, the first intermediate features, the second downsampled features, and the first downsampled features to obtain a fourth fused feature; and convolve the fourth fused feature to obtain the target's low-resolution features.

[0093] In practice, the medium-resolution features of the target are... The second downsampling feature is obtained by performing downsampling processing. For initial high-resolution features First intermediate feature Second downsampling features and the first downsampling feature The fourth fusion feature is obtained by performing fusion processing. Regarding the fourth fusion feature Convolution processing is performed to obtain low-resolution features of the target.

[0094] The formula for calculating the low-resolution features of the target is: ,in, For low-resolution features of the target. This is a separable quantum dilated convolution operation. For initial high-resolution features, As the first intermediate feature, For downsampling processing, For the target's medium-resolution features High-resolution features for the target.

[0095] Step 1043D: The target high-resolution features, the target medium-resolution features, and the target low-resolution features are used as target multi-scale features.

[0096] In practical implementation, the target's high-resolution features Medium resolution features of the target and target low-resolution features As a target multi-scale feature.

[0097] Through a dual-path mechanism of bottom-up upsampling fusion and top-down downsampling to supplement details, the cross-scale feature fusion module can achieve full fusion of high-level semantic information and low-level detailed information without significantly increasing computational cost, thereby obtaining richer multi-scale feature representations and effectively improving the model's classification and localization performance in insulator defect detection tasks.

[0098] The above scheme involves upsampling the second intermediate feature to obtain a second upsampled feature, fusing the initial high-resolution feature and the second upsampled feature to obtain a second fused feature, and then convolving the second fused feature to obtain the target high-resolution feature. The target high-resolution feature is then downsampled to obtain a first downsampled feature. The initial medium-resolution feature, the second intermediate feature, and the first downsampled feature are then fused to obtain a third fused feature, and this third fused feature is then convolved to obtain the target medium-resolution feature. The target medium-resolution feature is then downsampled to obtain a second downsampled feature. The initial high-resolution feature, the first intermediate feature, the second downsampled feature, and the first downsampled feature are then fused to obtain a fourth fused feature, and this fourth fused feature is then convolved to obtain the target low-resolution feature. These target high-resolution, target medium-resolution, and target low-resolution features are used as the target multi-scale features. This effectively fuses the semantic information of high-level features and the detailed information of low-level features, resulting in richer feature information in the target multi-scale features.

[0099] Through the above embodiments, insulator images of transmission lines are acquired and input into a pre-trained defect detection model. The defect detection model performs feature extraction on the insulator images to obtain insulator feature maps. Using the attention feedback module in the defect detection model, attention feedback features are determined based on the insulator feature maps. This module allocates limited computational resources to more important information in the current insulator defect detection task, reducing or even filtering out attention to irrelevant information. Multi-scale insulator features are obtained by fusing the attention feedback features, which enhances the semantic information of low-level features. The defect detection model then determines the insulator defect detection result based on these multi-scale features. This approach enables the defect detection model to be adaptive and robust, meeting the requirements of insulator defect detection in different scenarios. It improves the accuracy of insulator defect identification in transmission lines, reduces the probability of missed and false detections, and provides strong protection for the safe operation of transmission lines.

[0100] Existing insulator defect detection methods do not comprehensively consider the complex background of aerial images of insulators, such as interference from power towers and buildings, and also fail to account for the differences in string breakage defects across different scenarios. To address these issues, this disclosure proposes an attention feedback module, resolving the problem of spatial feature loss in attention mechanisms and enhancing channel attention capabilities. This disclosure also designs a PSiLU activation function to reduce the confusion between normal insulator features and insulator defect features. Furthermore, this disclosure constructs a novel feature fusion module that effectively fuses semantic information from high-level features and detailed information from low-level features, obtaining richer multi-scale feature information and effectively solving the aforementioned problems related to insulator defect detection.

[0101] Technical effects of the embodiments disclosed herein: (1) This disclosure proposes an attention feedback module, which can allocate limited computing resources to more important information in the current object detection task, reducing or even filtering out attention to irrelevant information. The attention feedback module includes: ① an attention propagation path, which is composed of a squeeze-excitation network and is used to extract the discriminative channel features of the network. ② an attention feedback path, which extracts the spatial features enhanced by the squeeze-excitation network through convolution and activation, thereby improving the way the network pays attention to discriminative features.

[0102] (2) The present invention proposes the PSiLU activation function, which makes feature extraction buffered and stable, and less likely to cause confusion of defects in similar insulators.

[0103] (3) The embodiments of this disclosure employ separable quantum dilated convolution. The separable quantum dilated convolution operation assigns corresponding quantum information to each pixel of the input tensor before each convolution, effectively improving the model's ability to extract multi-scale features and localization regression accuracy.

[0104] (4) This embodiment of the present disclosure constructs a cross-scale feature fusion module, which can make full use of the enhanced multi-level features in the backbone network. The cross-scale feature fusion module adopts a bottom-up path to upsample high-level features and enhance the semantic information of low-level features.

[0105] (5) The insulator defect detection method of attention-cross-scale fusion YOLOv7 used in this embodiment has adaptive capability and robustness. It can optimize the defect detection method according to the insulator conditions in different scenarios. The insulator defect detection method meets the requirements of insulator defect detection in different scenarios.

[0106] (6) The insulator defect detection method of attention-cross-scale fusion YOLOv7 in this disclosure can achieve good insulator defect identification effect, and also has good exploratory value for the development of related theories and technologies. This disclosure can be widely applied in the field of insulator defect detection in transmission lines and substations.

[0107] It should be noted that the method of this disclosure embodiment can be executed by a single device, such as a computer or server. The method of this embodiment can also be applied to a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method of this disclosure embodiment, and the multiple devices will interact with each other to complete the method described.

[0108] It should be noted that the above description describes some embodiments of this disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in a different order than that shown in the above embodiments and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0109] Based on the same inventive concept, corresponding to any of the above embodiments, this disclosure also provides an insulator defect detection device based on improved YOLO cross-scale fusion.

[0110] refer to Figure 4 The insulator defect detection device based on improved YOLO cross-scale fusion includes: The insulator image acquisition module 301 is configured to acquire insulator images of transmission lines and input the insulator images into a pre-trained defect detection model; Feature extraction processing module 302 is configured to use the defect detection model to perform feature extraction processing on the insulator image to obtain an insulator feature map; The attention feedback feature determination module 303 is configured to determine attention feedback features based on the insulator feature map using the attention feedback module in the defect detection model. The fusion processing module 304 is configured to obtain multi-scale insulator features by fusing the attention feedback features; The defect detection result determination module 305 is configured to use the defect detection model to determine the insulator defect detection result based on the multi-scale insulator characteristics.

[0111] In some embodiments, the attention feedback feature determination module 303 includes: The channel compression feature determination unit is configured to use the attention feedback module in the defect detection model to filter the insulator feature map to obtain channel convolution features, and determine channel compression features based on the channel convolution features. The channel modulation weight vector determination unit is configured to determine the channel modulation weight vector based on the channel compression features using an activation function. The attention feedback feature determination unit is configured to multiply the channel modulation weight vector and the channel convolution feature to obtain the dimension recovery parameter, and determine the attention feedback feature based on the dimension recovery parameter.

[0112] In some embodiments, the channel compression feature determination unit includes: The channel convolution feature determination subunit is configured to use the attention feedback module in the defect detection model to transform the insulator feature map to obtain a transformed feature map, and then to filter the transformed feature map to obtain channel convolution features. The channel compression feature determination subunit is configured to perform global average pooling on the channel convolutional features to obtain global average features, perform median pooling on the channel convolutional features to obtain median features, and perform mean processing on the global average features and the median features to obtain channel compression features.

[0113] In some embodiments, the channel modulation weight vector determination unit includes: The intermediate variable determination subunit is configured to use the first hyperparameter to perform dimensionality reduction processing on the channel compression features to obtain intermediate variables; The mapping feature determination subunit is configured to obtain a mapping feature by performing a non-linear mapping on each element of the intermediate variable using a first activation function. The channel weight representation is determined by a sub-unit, which is configured to use a second hyperparameter to perform dimensionality recovery processing on the mapped features to obtain the channel weight representation. The channel modulation weight vector determination subunit is configured to determine the channel modulation weight vector based on the channel weight representation using a second activation function.

[0114] In some embodiments, the multi-scale insulator feature determination module 304 includes: An initial multi-scale feature determination unit is configured to determine initial multi-scale features from the attention feedback features; The intermediate multi-scale feature determination unit is configured to perform a fusion process on the initial multi-scale features to obtain intermediate multi-scale features; The target multi-scale feature determination unit is configured to perform fusion processing on the intermediate multi-scale features to obtain target multi-scale features, and to use the target multi-scale features as multi-scale insulator features.

[0115] In some embodiments, the intermediate multi-scale feature determination unit includes: An initial multi-scale feature determination subunit is configured to determine initial high-resolution features, initial medium-resolution features, and initial low-resolution features from the initial multi-scale features; The first intermediate feature determination subunit is configured to perform convolution processing on the initial low-resolution features to obtain the first intermediate features; The second intermediate feature determination subunit is configured to perform upsampling processing on the first intermediate feature to obtain a first upsampled feature, perform fusion processing on the initial high-resolution feature, the initial medium-resolution feature and the first upsampled feature to obtain a first fused feature, and perform convolution processing on the first fused feature to obtain a second intermediate feature.

[0116] In some embodiments, the target multi-scale features include: The target high-resolution feature determination subunit is configured to perform upsampling processing on the second intermediate feature to obtain a second upsampled feature, perform fusion processing on the initial high-resolution feature and the second upsampled feature to obtain a second fused feature, and perform convolution processing on the second fused feature to obtain the target high-resolution feature; The target medium-resolution feature determination subunit is configured to perform downsampling processing on the target high-resolution features to obtain a first downsampled feature, perform fusion processing on the initial medium-resolution feature, the second intermediate feature and the first downsampled feature to obtain a third fused feature, and perform convolution processing on the third fused feature to obtain the target medium-resolution feature; The target low-resolution feature determination subunit is configured to perform downsampling processing on the target medium-resolution features to obtain a second downsampled feature, perform fusion processing on the initial high-resolution features, the first intermediate features, the second downsampled features and the first downsampled features to obtain a fourth fused feature, and perform convolution processing on the fourth fused feature to obtain the target low-resolution features; The target multi-scale feature determination subunit is configured to use the target high-resolution feature, the target medium-resolution feature, and the target low-resolution feature as target multi-scale features.

[0117] For ease of description, the above apparatus is described in terms of its functions, divided into various modules. Of course, in implementing this disclosure, the functions of each module can be implemented in one or more software and / or hardware.

[0118] The apparatus of the above embodiments is used to implement the corresponding insulator defect detection method based on improved YOLO cross-scale fusion in any of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0119] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this disclosure also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the insulator defect detection method based on improved YOLO cross-scale fusion as described in any of the above embodiments.

[0120] Figure 5 This embodiment illustrates a more specific hardware structure of an electronic device. The device may include a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, memory 1020, input / output interface 1030, and communication interface 1040 are interconnected internally via the bus 1050.

[0121] The processor 1010 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0122] The memory 1020 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.

[0123] The input / output interface 1030 is used to connect input / output modules to realize information input and output. Input / output modules can be configured as components within the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touchscreens, microphones, various sensors, etc., while output devices may include displays, speakers, vibrators, indicator lights, etc.

[0124] The communication interface 1040 is used to connect the communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB (Universal Serial Bus), network cable, etc.) or wireless means (such as mobile network, WIFI (Wireless Fidelity), Bluetooth, etc.).

[0125] Bus 1050 includes a pathway for transmitting information between various components of the device, such as processor 1010, memory 1020, input / output interface 1030, and communication interface 1040.

[0126] It should be noted that although the above-described device only shows the processor 1010, memory 1020, input / output interface 1030, communication interface 1040, and bus 1050, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this specification, and not necessarily all the components shown in the figures.

[0127] The electronic devices described above are used to implement the corresponding insulator defect detection method based on improved YOLO cross-scale fusion in any of the foregoing embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0128] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this disclosure also provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to execute the insulator defect detection method based on improved YOLO cross-scale fusion as described in any of the above embodiments.

[0129] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0130] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to execute the insulator defect detection method based on improved YOLO cross-scale fusion as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0131] Based on the same inventive concept, corresponding to any of the above embodiments, this application also provides a computer program product, including computer program instructions. When the computer program instructions are run on a computer, the computer executes the insulator defect detection method based on improved YOLO cross-scale fusion as described in any of the above embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0132] It is understood that before using the technical solutions of the various embodiments in this disclosure, users will be informed of the type, scope of use, and usage scenarios of the personal information involved in an appropriate manner, and user authorization will be obtained.

[0133] For example, upon receiving a user's active request, a prompt message is sent to the user to explicitly inform them that the requested operation will require the acquisition and use of the user's personal information. This allows the user to independently choose, based on the prompt message, whether to provide personal information to the software or hardware such as electronic devices, applications, servers, or storage media performing the operations of this disclosed technical solution.

[0134] As an optional but not limited implementation, in response to a user's active request, sending a prompt message to the user can be done via a pop-up window, where the prompt message can be presented in text format. Furthermore, the pop-up window can also include a selection control allowing the user to choose "agree" or "disagree" to provide personal information to the electronic device.

[0135] It is understood that the above notification and user authorization process are merely illustrative and do not constitute a limitation on the implementation of this disclosure. Other methods that comply with relevant laws and regulations may also be applied to the implementation of this disclosure.

[0136] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this disclosure is limited to these examples; within the framework of this disclosure, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of this disclosure as described above, which are not provided in detail for the sake of brevity.

[0137] Additionally, to simplify the description and discussion, and to avoid obscuring the embodiments of this disclosure, the provided drawings may or may not show well-known power / ground connections to integrated circuit (IC) chips and other components. Furthermore, the apparatus may be shown in block diagram form to avoid obscuring the embodiments of this disclosure, and this also takes into account the fact that the details of implementation of these block diagram apparatuses are highly dependent on the platform on which the embodiments of this disclosure will be implemented (i.e., these details should be fully understood by those skilled in the art). While specific details (e.g., circuits) have been set forth to describe exemplary embodiments of this disclosure, it will be apparent to those skilled in the art that the embodiments of this disclosure can be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0138] Although this disclosure has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.

[0139] This disclosure is intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this disclosure. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the protection scope of this disclosure.

Claims

1. A method for detecting insulator defects based on improved YOLO cross-scale fusion, characterized in that, The method includes: Acquire images of insulators of the transmission line and input the insulator images into a pre-trained defect detection model; Using the aforementioned defect detection model, feature extraction processing is performed on the insulator image to obtain an insulator feature map; The attention feedback module in the defect detection model is used to determine the attention feedback features based on the insulator feature map; Multi-scale insulator features are obtained by fusing the attention feedback features. Using the aforementioned defect detection model, the insulator defect detection results are determined based on the multi-scale insulator characteristics.

2. The method according to claim 1, characterized in that, The step of using the attention feedback module in the defect detection model to determine the attention feedback features based on the insulator feature map includes: Using the attention feedback module in the defect detection model, the insulator feature map is filtered to obtain channel convolution features, and channel compression features are determined based on the channel convolution features. The channel modulation weight vector is determined using an activation function based on the channel compression characteristics; The dimension recovery parameters are obtained by multiplying the channel modulation weight vector and the channel convolution features, and the attention feedback features are determined based on the dimension recovery parameters.

3. The method according to claim 2, characterized in that, The step of using the attention feedback module in the defect detection model to filter the insulator feature map to obtain channel convolution features, and determining channel compression features based on the channel convolution features, includes: The insulator feature map is transformed using the attention feedback module in the defect detection model to obtain a transformed feature map, and the transformed feature map is then filtered to obtain channel convolution features. The channel convolutional features are subjected to global average pooling to obtain global average features, the channel convolutional features are subjected to median pooling to obtain median features, and the global average features and the median features are subjected to mean processing to obtain channel compressed features.

4. The method according to claim 2, characterized in that, The step of determining the channel modulation weight vector based on the channel compression features using an activation function includes: The channel compression features are reduced in dimensionality using the first hyperparameter to obtain intermediate variables; The mapping feature is obtained by performing a nonlinear mapping on each element of the intermediate variable using the first activation function. The channel weight representation is obtained by performing dimensionality restoration processing on the mapped features using the second hyperparameter; The channel modulation weight vector is determined using the second activation function based on the channel weight representation.

5. The method according to claim 1, characterized in that, The process of fusing the attention feedback features to obtain multi-scale insulator features includes: Initial multi-scale features are determined from the attention feedback features; The initial multi-scale features are fused to obtain intermediate multi-scale features; The intermediate multi-scale features are fused to obtain the target multi-scale features, and the target multi-scale features are used as multi-scale insulator features.

6. The method according to claim 5, characterized in that, The process of fusing the initial multi-scale features to obtain intermediate multi-scale features includes: From the initial multi-scale features, determine the initial high-resolution features, initial medium-resolution features, and initial low-resolution features; The initial low-resolution features are convolved to obtain the first intermediate features; The first intermediate feature is upsampled to obtain a first upsampled feature. The initial high-resolution feature, the initial medium-resolution feature, and the first upsampled feature are fused to obtain a first fused feature. The first fused feature is then convolved to obtain a second intermediate feature.

7. The method according to claim 6, characterized in that, The process of fusing the intermediate multi-scale features to obtain the target multi-scale features includes: The second intermediate feature is upsampled to obtain the second upsampled feature. The initial high-resolution feature and the second upsampled feature are fused to obtain the second fused feature. The second fused feature is then convolved to obtain the target high-resolution feature. The target high-resolution features are downsampled to obtain a first downsampled feature. The initial medium-resolution features, the second intermediate features, and the first downsampled feature are fused to obtain a third fused feature. The third fused feature is then convolved to obtain the target medium-resolution features. The target's medium-resolution features are downsampled to obtain a second downsampled feature. The initial high-resolution features, the first intermediate features, the second downsampled features, and the first downsampled features are fused to obtain a fourth fused feature. The fourth fused feature is then convolved to obtain the target's low-resolution features. The target's high-resolution features, medium-resolution features, and low-resolution features are used as the target's multi-scale features.

8. An insulator defect detection device based on improved YOLO cross-scale fusion, characterized in that, include: An insulator image acquisition module is configured to acquire insulator images of transmission lines and input the insulator images into a pre-trained defect detection model; The feature extraction and processing module is configured to use the defect detection model to perform feature extraction processing on the insulator image to obtain an insulator feature map; The attention feedback feature determination module is configured to determine attention feedback features based on the insulator feature map using the attention feedback module in the defect detection model. The fusion processing module is configured to obtain multi-scale insulator features by fusing the attention feedback features; The defect detection result determination module is configured to use the defect detection model to determine the insulator defect detection result based on the multi-scale insulator characteristics.

9. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the program, implements the method as claimed in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions for causing a computer to perform the method according to any one of claims 1 to 7.

Images