Component detection method and system for power transmission line, storage medium and equipment

Abstract

The embodiment of the invention discloses a component detection method for a power transmission line. The method comprises the following steps: acquiring a component image of the power transmission line; inputting the component image into a target detection model, wherein the target detection model comprises a backbone network, a neck network and a head network; the backbone network comprises an image feature extraction module and a large kernel space convolution module, and the neck network comprises an asymmetrically filled feature fusion module; obtaining multi-scale image features of the component image according to an image feature extraction module, and obtaining detail features of the component image according to a large kernel space convolution module; inputting the multi-scale image features and the detail features into a neck network, and fusing the multi-scale image features and the detail features according to an asymmetrically filled feature fusion module of the neck network to obtain a component fusion feature map; and transmitting the component fusion feature map to a head network, and detecting the component fusion feature map by the head network to generate a final detection result. According to the invention, accurate detection of power transmission line components can be realized.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a component detection method, system, storage medium and device for power transmission lines. Background Art

[0002] In the field of power system operation and maintenance, monitoring the condition of key transmission line components (such as insulators, bolts, and vibration dampers) is a vital step in ensuring the safe and stable operation of the power grid. With the expansion of the power grid and the increasing complexity of the transmission environment, traditional manual inspection methods face problems such as low efficiency, high cost, and significant environmental impact on detection accuracy. This is especially true in complex terrain such as mountainous areas and waterways, as well as in harsh conditions such as rain, fog, and at night. It is difficult for manual inspections to comprehensively and promptly detect early defects (such as fine cracks and corrosion spots) in components.

[0003] Existing computer vision-based detection methods mostly use traditional convolutional neural network models, but their feature extraction capabilities for multi-scale targets (such as large-scale transmission tower structures and small-scale bolt defects) in complex scenarios are insufficient, making it difficult to take into account both global structures and local details at the same time. Summary of the Invention

[0004] Based on this, it is necessary to propose a component detection method for transmission lines to address the above problems.

[0005] A method for detecting components of a power transmission line, the method comprising the following steps:

[0006] Collect images of transmission line components;

[0007] Inputting the component image into an object detection model; the object detection model includes a backbone network, a neck network, and a head network; the backbone network includes an image feature extraction module and a large kernel spatial convolution module, and the neck network includes an asymmetric padding feature fusion module;

[0008] Acquire multi-scale image features of the component image according to the image feature extraction module, and acquire detail features of the component image according to the large kernel spatial convolution module;

[0009] Inputting the multi-scale image features and the detail features into the neck network, and fusing the multi-scale image features and the detail features according to the asymmetric filling feature fusion module of the neck network to obtain a component fusion feature map;

[0010] The component fusion feature map is transmitted to the head network, and the head network detects the component fusion feature map to generate a final detection result.

[0011] In the above solution, the image feature extraction module includes a convolution layer, a segmentation layer, a bottleneck submodule, and a splicing layer for feature aggregation; wherein the bottleneck submodule includes an integrated exponential moving average mechanism.

[0012] In the above solution, the step of obtaining the multi-scale image features of the component image according to the image feature extraction module specifically includes:

[0013] Inputting the component image into the first convolutional layer, obtaining the component image channel dimension features and the component image spatial dimension features, and transforming the two to obtain a transformed component image;

[0014] Inputting the converted component image into a segmentation layer to obtain a plurality of branch graphs;

[0015] Inputting the plurality of branch graphs into the bottleneck submodule in sequence to weight them, and obtaining a plurality of weighted branch graphs;

[0016] Inputting the weighted branch graphs into a splicing layer for aggregation to obtain a splicing component image;

[0017] The spliced ​​component image is input into the second convolutional layer to obtain corresponding multi-scale image features.

[0018] In the above scheme, the network structure of the large-kernel spatial convolution module includes: a large-kernel convolution sequence, an average pooling layer, a maximum pooling layer, a spatial attention layer and a convolution layer.

[0019] In the above solution, the obtaining of detail features of the component image according to the large-kernel spatial convolution module specifically includes:

[0020] According to the two large kernel convolution sequences, the features of the convolution kernels of different receptive fields of the component image are obtained and connected:

[0021]

[0022] in, Indicates the features after connection;

[0023] The connected features are pooled according to the average pooling layer and the maximum pooling layer:

[0024]

[0025] Among them, SA avg and SA max Represent the spatial descriptors after average pooling and maximum pooling respectively;

[0026] The pooled spatial descriptor is input into the spatial attention layer for interaction to obtain a spatial attention map:

[0027]

[0028] in, Represents the spatial attention map;

[0029] Apply the activation function to the spatial attention map to obtain the decoupled kernel features:

[0030]

[0031] in, represents the decoupled kernel feature;

[0032] Obtain the mask corresponding to the decoupling kernel feature, perform convolution on the decoupling kernel feature and the corresponding mask to obtain the detail features of the component image:

[0033]

[0034] Among them, S represents the detail features of the component image, represents the decoupled kernel feature, Indicates the corresponding mask.

[0035] In the above scheme, the asymmetric padding feature fusion module includes four groups of parallel asymmetric padding convolution operation units, each group of asymmetric padding convolution operation units is configured with a different padding parameter P, which is used to directionally expand the receptive field to capture edge gradient features; the asymmetric padding convolution operation units are sequentially connected to the normalization layer and the splicing layer.

[0036] In the above solution, the multi-scale image features and the detail features are input into the neck network, and the multi-scale image features and the detail features are fused according to the asymmetric filling feature fusion module of the neck network to obtain a component fusion feature map, which specifically includes:

[0037] The multi-scale image features and detail features are combined into an input component feature map X(h1, w1, c1), and the input component feature map is asymmetrically padded using different padding parameters P to generate a feature map with an extended receptive field:

[0038] Four sets of asymmetric padding convolution operations are performed in parallel on the feature map of the expanded receptive field:

[0039]

[0040] in, represents the convolution operator, represents a 1×3 kernel with C′ output channels, and the padding parameter P represents the pixel count in each spatial direction;

[0041] Get the dimension of the component feature map output after the asymmetric padding convolution operation:

[0042]

[0043] Among them, C2 represents the final output channel number of the asymmetric padded feature fusion module, and s represents the convolution stride;

[0044] Perform interleaved convolution on the component feature map dimensions to obtain the component fusion feature map:

[0045]

[0046] Among them, X (h′，w′，4c) Represents the component fusion feature map.

[0047] In the above scheme, the backbone network also includes a pooling module, an image feature extraction module, a convolution layer and a large-kernel spatial convolution module; the neck network also includes an upsampling layer, a splicing layer, a convolution layer and an asymmetric padding feature fusion module; the head network includes several detection layers.

[0048] In the above solution, the collecting of images of components of the transmission line specifically includes:

[0049] Capture images of insulators, vibration dampers, and bolts from multiple angles and at multiple times under conditions of strong light on sunny days, weak light on cloudy days, low light at night, rainy days, foggy days, snowy days, and in mountainous areas, plains, and water areas;

[0050] The components of the transmission line include insulators, conductors, poles and towers, hardware, lightning protection devices, vibration absorbers, bolts, bird nests and separators.

[0051] The present application also provides a readable storage medium storing a computer program. When the computer program is executed by a processor, the processor performs the following steps:

[0052] Collect images of transmission line components;

[0053] Inputting the component image into an object detection model; the object detection model includes a backbone network, a neck network, and a head network; the backbone network includes an image feature extraction module and a large kernel spatial convolution module, and the neck network includes an asymmetric padding feature fusion module;

[0054] Acquire multi-scale image features of the component image according to the image feature extraction module, and acquire detail features of the component image according to the large kernel spatial convolution module;

[0055] Inputting the multi-scale image features and the detail features into the neck network, and fusing the multi-scale image features and the detail features according to the asymmetric filling feature fusion module of the neck network to obtain a component fusion feature map;

[0056] The component fusion feature map is transmitted to the head network, and the head network detects the component fusion feature map to generate a final detection result.

[0057] The present application also proposes a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes the following steps:

[0058] Collect images of transmission line components;

[0059] Inputting the component image into an object detection model; the object detection model includes a backbone network, a neck network, and a head network; the backbone network includes an image feature extraction module and a large kernel spatial convolution module, and the neck network includes an asymmetric padding feature fusion module;

[0060] Acquire multi-scale image features of the component image according to the image feature extraction module, and acquire detail features of the component image according to the large kernel spatial convolution module;

[0061] Inputting the multi-scale image features and the detail features into the neck network, and fusing the multi-scale image features and the detail features according to the asymmetric filling feature fusion module of the neck network to obtain a component fusion feature map;

[0062] The component fusion feature map is transmitted to the head network, and the head network detects the component fusion feature map to generate a final detection result.

[0063] The embodiment of the present invention has the following beneficial effects: first, the image feature extraction module is used to obtain multi-scale features of the component image, capturing the overall contour and structure of the component from a macroscopic perspective; second, the large-core spatial convolution module uses a larger convolution kernel receptive field to deeply explore the detailed features in the component image, such as texture information of subtle defects such as cracks and wear. Furthermore, the multi-scale features and detailed features are input into the neck network. With the help of the feature fusion module with asymmetric filling, the boundary information loss caused by traditional symmetric filling is avoided, and the precise alignment and complementary fusion of different scale and detail features are achieved, generating a component fusion feature map containing complete spatial location information and rich semantic information. Finally, the head network performs detection based on this fused feature map, combining the multi-scale context and detail representation in the features to achieve accurate positioning and category recognition of transmission line components, effectively improving the detection accuracy of small components and defective components in complex scenarios, reducing the missed detection and false detection rates, and ensuring the comprehensiveness and reliability of transmission line component detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0065] in:

[0066] Figure 1 1 is a flow chart of a component detection method for a power transmission line according to an embodiment;

[0067] Figure 2 Schematic diagram of the structure of an object detection model (ELA-YOLO) in one embodiment;

[0068] Figure 3 Schematic diagram of the structure of a large kernel spatial convolution module (LSK) according to an embodiment;

[0069] Figure 4 Schematic diagram of a comparison of detection effects in one embodiment.

[0070] Reference numerals

[0071] Backbone: backbone network; Neck: neck network; Head: head network; Detect: detection layer; Conv: convolution layer; Split: segmentation layer; Bottleneck: bottleneck submodule; BN: normalization layer; Concat: splicing layer; ELA-YOLO: target detection model; C2f_EMA: image feature extraction module; LSK: large kernel spatial convolution module; C2f_AP: asymmetric padding feature fusion module; SPPF: pooling module; Bottleneck_EMA: bottleneck submodule with integrated exponential moving average mechanism; APConv: asymmetric padding convolution layer; Upsample: upsampling; FC: fully connected layer; Larger Kernel: large kernel convolution sequence; Avg Max Pool average pooling layer, maximum pooling layer; SA: spatial attention layer. DETAILED DESCRIPTION

[0072] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0073] In the following description, numerous specific details are given to provide a more thorough understanding of the present invention; however, it will be apparent to those skilled in the art that the present invention may be practiced without one or more of these details; in other examples, some technical features known in the art are not described to avoid confusion with the present invention, and it should be understood that the present invention may be practiced in different forms and should not be construed as limited to the embodiments set forth herein; on the contrary, these embodiments are provided to make the disclosure thorough and complete and to fully convey the scope of the present invention to those skilled in the art.

[0074] The terminology used herein is intended only to describe specific embodiments and is not intended to limit the present invention. When used herein, the singular forms "a," "an," and "the" are intended to include the plural forms, unless the context clearly indicates otherwise. It should also be understood that the terms "comprising" and / or "including," when used in this specification, identify the presence of the recited features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups. When used herein, the term "and / or" includes any and all combinations of the relevant listed items.

[0075] In order to thoroughly understand the present invention, a detailed structure will be presented in the following description to illustrate the technical solution proposed by the present invention; optional embodiments of the present invention are described in detail below, but in addition to these detailed descriptions, the present invention may also have other implementation methods.

[0076] like Figure 1 As shown, in one embodiment, a method for detecting components of a power transmission line is provided. The method for detecting components of a power transmission line includes steps S101 to S105, which are described in detail as follows:

[0077] S101, collecting images of components of a transmission line;

[0078] Capturing component images through image acquisition can provide a foundation for subsequent analysis. Preferably, capturing high-resolution or multi-angle component images improves the integrity of component features and enhances the model's adaptability to complex environments (such as lighting changes and occlusion).

[0079] In some embodiments, collecting images of components of a power transmission line specifically includes:

[0080] Capture images of insulators, vibration dampers, and bolts from multiple angles and at multiple times under conditions of strong light on sunny days, weak light on cloudy days, low light at night, rainy days, foggy days, snowy days, and in mountainous areas, plains, and water areas;

[0081] The components of a transmission line include insulators, conductors, poles, hardware, lightning protection devices, vibration dampers, bolts, bird nests and separators.

[0082] Specifically, drones equipped with high-definition cameras, infrared thermal imagers and other equipment can be used to collect images from multiple angles and multiple time periods under different lighting conditions (such as strong light on sunny days, weak light on cloudy days, and low light at night), different weather conditions (such as rainy days, foggy days, and snowy days), and key components of transmission lines in different regions and types (such as insulators, shock absorbers, bolts, etc.). This can fully consider terrain factors and reasonably plan drone flight paths and shooting parameters for transmission lines in different terrains such as mountainous areas, plains, and waters to ensure that the acquired data is comprehensive and representative. In this way, a rich variety of raw data can be collected, providing a solid data foundation for subsequent model training.

[0083] S102. Input the component image into the target detection model; the target detection model includes a backbone network, a neck network, and a head network; the backbone network includes an image feature extraction module and a large kernel spatial convolution module, and the neck network includes an asymmetric padding feature fusion module;

[0084] The modular design enables division of labor and collaboration. The backbone network is mainly used to extract basic features; the neck network is mainly used to fuse and enhance features; and the head network is used to generate the final detection results: bounding boxes and category labels.

[0085] The image feature extraction module uses multi-scale feature extraction (such as convolution layers at different levels) to capture global and local features of components, adapting to the detection needs of components of different sizes. The large-core spatial convolution module uses larger-than-normal convolution kernels (such as 7×7 and 9×9) to expand the receptive field and enhance the ability to capture component details (such as cracks and wear). The feature fusion module with asymmetric padding uses asymmetric padding (such as single-sided padding) during feature fusion, which can preserve the spatial position information of the feature map, reduce the boundary information blurring caused by traditional symmetric padding, and improve the detection accuracy of small components or edge features.

[0086] like Figures 2 to 3 As shown, in some embodiments, the image feature extraction module includes a convolution layer, a segmentation layer, a bottleneck submodule, and a splicing layer for feature aggregation; wherein the bottleneck submodule includes an integrated exponential moving average mechanism.

[0087] Furthermore, the network structure of the large-kernel spatial convolution module includes: a large-kernel convolution sequence, an average pooling layer, a maximum pooling layer, a spatial attention layer, and a convolution layer.

[0088] Furthermore, the asymmetric padding feature fusion module includes four groups of parallel asymmetric padding convolution operation units, each group of asymmetric padding convolution operation units is configured with a different padding parameter P, which is used to directionally expand the receptive field to capture edge gradient features; the asymmetric padding convolution operation units are sequentially connected to the normalization layer and the splicing layer.

[0089] S103, obtaining multi-scale image features of the component image using the image feature extraction module, and obtaining detail features of the component image using the large kernel spatial convolution module;

[0090] Specifically, multi-scale features simultaneously retain the spatial details of low-level features and the semantic information of high-level features, enabling effective detection of components of varying sizes. Detailed features, which extract texture and edge information, can help identify subtle surface defects and improve fault prediction capabilities.

[0091] In some embodiments, obtaining multi-scale image features of a component image according to an image feature extraction module specifically includes:

[0092] Input the component image into the first convolutional layer, obtain the component image channel dimension features and the component image spatial dimension features, and transform the two to obtain the transformed component image;

[0093] Input the converted component image into the segmentation layer to obtain several branch graphs;

[0094] Inputting several branch graphs into the bottleneck submodule in sequence to weight them, and obtaining several weighted branch graphs;

[0095] Input the weighted branch graphs into the splicing layer for aggregation to obtain the splicing component image;

[0096] The spliced ​​component image is input into the second convolutional layer to obtain the corresponding multi-scale image features.

[0097] In some embodiments, obtaining detailed features of a component image using a large-kernel spatial convolution module specifically includes:

[0098] According to the two large kernel convolution sequences, the features of the convolution kernels with different receptive fields of the component image are obtained and connected:

[0099]

[0100] in, Indicates the features after connection;

[0101] The connected features are pooled according to the average pooling layer and the maximum pooling layer:

[0102]

[0103] Among them, SAavg and SA max Represent the spatial descriptors after average pooling and maximum pooling respectively;

[0104] The pooled spatial descriptor is input into the spatial attention layer for interaction to obtain the spatial attention map:

[0105]

[0106] in, Represents the spatial attention map;

[0107] Apply activation function to the spatial attention map to obtain the decoupled kernel features:

[0108]

[0109] in, represents the decoupled kernel feature;

[0110] Obtain the mask corresponding to the decoupling kernel feature, perform convolution on the decoupling kernel feature and the corresponding mask to obtain the detailed features of the component image:

[0111]

[0112] Among them, S represents the detail features of the component image, represents the decoupled kernel feature, Indicates the corresponding mask.

[0113] S104: Input the multi-scale image features and detail features into the neck network, and fuse the multi-scale image features and detail features according to the asymmetric filling feature fusion module of the neck network to obtain a component fusion feature map;

[0114] The asymmetric filling fusion module can avoid information loss during feature alignment, thereby organically combining features of different levels and dimensions to generate a more comprehensive and semantically rich feature representation. The fused feature map has both strong semantic information and precise spatial information, which can improve detection accuracy.

[0115] In some embodiments, the step of inputting the multi-scale image features and the detail features into the neck network, fusing the multi-scale image features and the detail features according to the asymmetric filling feature fusion module of the neck network, and obtaining the component fusion feature map specifically includes:

[0116] The multi-scale image features and detail features are combined into the input component feature map X(h1,,w1,c1), and the input component feature map is asymmetrically padded by different padding parameters P to generate a feature map with an extended receptive field:

[0117] Four sets of asymmetric padding convolution operations are performed in parallel on the feature maps of the expanded receptive field:

[0118]

[0119] in, represents the convolution operator, represents a 1×3 kernel with C′ output channels, and the padding parameter P represents the pixel count in each spatial direction;

[0120] Get the dimension of the component feature map output after the asymmetric padding convolution operation:

[0121]

[0122] Among them, C2 represents the final output channel number of the asymmetric padded feature fusion module, and s represents the convolution stride;

[0123] Perform interleaved convolution on the component feature map dimensions to obtain the component fusion feature map:

[0124]

[0125] Among them, X′ (h′，w′，4c) Represents the component fusion feature map.

[0126] S105: Transmit the component fusion feature map to the head network, and the head network detects the component fusion feature map to generate a final detection result.

[0127] Based on the fused feature map, the head network outputs the location, category, confidence and abnormal status of the component through the classifier and regressor.

[0128] like Figure 2 As shown, in some embodiments, the backbone network also includes a pooling module, an image feature extraction module, a convolution layer and a large-kernel spatial convolution module; the neck network also includes an upsampling layer, a splicing layer, a convolution layer and a feature fusion module with asymmetric padding; the head network includes several detection layers.

[0129] Furthermore, we select an appropriate loss function (such as the cross-entropy loss function) and optimizer (such as the Adam optimizer) and train the improved object detection model using the training set data. During training, we employ techniques such as batch normalization and regularization to prevent overfitting and ensure that the object detection model has good generalization capabilities. Furthermore, we dynamically adjust the object detection model's hyperparameters, such as the learning rate and batch size, based on the performance of the validation set data.

[0130] For example, the initial value of the learning rate is set to 0.001, the batch size is 16, and the number of training rounds is 100. The learning rate of the target detection model is adjusted according to the average accuracy of the validation set data every 10 rounds. If the average accuracy no longer improves, the learning rate of the target detection model is multiplied by 0.1 to continuously optimize the model performance and improve the detection accuracy.

[0131] Preferably, the trained object detection model is pruned using the torch.nn.utils.prune library, setting the pruning ratio to 30%. This removes unimportant channels from the convolutional layers, thereby reducing the number of model parameters and computational complexity. Next, the model is converted to an 8-bit integer quantization model using PyTorch's quantization tool, torch.quantization, to further reduce computational complexity. The lightweight model is deployed on a device with limited computing resources, such as the NVIDIA Jetson AGX Xavier development board, and connected to a drone's image transmission system to achieve real-time detection of actual transmission line components. The detection results are then transmitted to a monitoring center via the network.

[0132] Experimental validation on the PTL-AI Furna dataset shows that ELA-YOLO outperforms existing detection methods in metrics such as mean average precision (mAP). While maintaining high detection accuracy, it reduces model parameters by 40% and computational complexity by 35%. This effectively meets the precision and lightweight requirements for detecting key transmission line components in complex scenarios, providing efficient and reliable technical support for intelligent transmission line operation and maintenance. The method has high practical application value and broad prospects for widespread adoption.

[0133] like Figure 4 The figure shows the significant practical advantages of the proposed method over competing methods. Transmission lines contain small components such as bolts and small-sized hardware. YOLOv5s and YOLOv8s may miss or ambiguously identify some of these small components. Compared to baseline methods, ELA-YOLO (object detection model), by accurately capturing the features of small objects, can clearly identify and label them. This achieves a higher recall rate in small object detection scenarios (such as tiny line components against complex backgrounds), effectively avoiding potential line hazards caused by undetected small component faults and ensuring power transmission safety. The bounding boxes generated by YOLOv5s and YOLOv8s in the image sometimes exhibit offset and inaccuracy (such as framing non-target areas or not fully enclosing the target). The bounding boxes generated by ELA-YOLO (object detection model) more closely match the actual component contours and provide precise positioning. In actual maintenance, these precise bounding boxes help personnel quickly locate the position and status of components, especially when identifying component defects (such as the location of damaged insulators), providing a more reliable basis for troubleshooting.

[0134] In summary, existing power transmission line detection technologies based on deep learning mostly use the traditional YOLO series or other general target detection algorithms, which have obvious defects when processing small targets, low-contrast images and partially occluded targets. The ELA-YOLO (target detection model) proposed in this solution deeply improves the backbone network and neck network of YOLOv8, and innovatively integrates a variety of new mechanisms. The backbone network embeds an image feature extraction module, and realizes dynamic weighted fusion through an efficient multi-scale attention mechanism. Compared with the single feature extraction method of traditional algorithms, it can more accurately capture target features of different scales; the neck network introduces an asymmetric padded feature fusion module, and uses asymmetric padded convolution to enhance edge gradient response, breaking through the limitations of traditional convolution structures in small target detection; a large kernel spatial convolution module is embedded in the middle, which can adaptively adjust the receptive field and enhance context perception. These improvements have greatly improved the detection accuracy of power transmission line components in complex scenarios, especially in the detection of small targets (such as bolts) and sub-pixel defects.

[0135] The present application also proposes a component detection system for a power transmission line, the system comprising: a power transmission line component image acquisition unit and a model detection unit;

[0136] A transmission line component image acquisition unit, used to acquire images of transmission line components;

[0137] Model detection unit, with built-in target detection model, which includes backbone network, neck network and head network;

[0138] The backbone network is configured with an image feature extraction module and a large-core spatial convolution module. The image feature extraction module is used to obtain multi-scale image features of component images, and the large-core spatial convolution module is used to obtain detailed features of component images.

[0139] The neck network is equipped with an asymmetric padded feature fusion module, which is used to receive the multi-scale image features and detail features output by the backbone network and fuse them to obtain the component fusion feature map;

[0140] The head network is used to receive the component fusion feature map output by the neck network, detect the component fusion feature map and generate the final detection result.

[0141] The present application also provides a readable storage medium storing a computer program. When the computer program is executed by a processor, the processor performs the following steps:

[0142] Collect images of transmission line components;

[0143] Feed the part image into the object detection model;

[0144] The target detection model includes a backbone network, a neck network, and a head network;

[0145] The backbone network includes an image feature extraction module and a large-kernel spatial convolution module, and the neck network includes an asymmetric padding feature fusion module;

[0146] The multi-scale image features of the component image are obtained by the image feature extraction module, and the detail features of the component image are obtained by the large kernel spatial convolution module;

[0147] The multi-scale image features and detail features are input into the neck network, and the multi-scale image features and detail features are fused according to the feature fusion module of the asymmetric filling of the neck network to obtain the component fusion feature map;

[0148] The component fusion feature map is transmitted to the head network, and the head network detects the component fusion feature map to generate the final detection result.

[0149] The present application also proposes a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer program is executed by the processor to perform the following steps:

[0150] Collect images of transmission line components;

[0151] Feed the part image into the object detection model;

[0152] The target detection model includes a backbone network, a neck network, and a head network;

[0153] The backbone network includes an image feature extraction module and a large-kernel spatial convolution module, and the neck network includes an asymmetric padding feature fusion module;

[0154] The multi-scale image features of the component image are obtained by the image feature extraction module, and the detail features of the component image are obtained by the large kernel spatial convolution module;

[0155] The multi-scale image features and detail features are input into the neck network, and the multi-scale image features and detail features are fused according to the feature fusion module of the asymmetric filling of the neck network to obtain the component fusion feature map;

[0156] The component fusion feature map is transmitted to the head network, and the head network detects the component fusion feature map to generate the final detection result.

[0157] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program. The program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0158] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0159] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person skilled in the art would be able to make numerous modifications and improvements without departing from the spirit of the present application, all of which fall within the scope of protection of the present application. The above disclosures are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Therefore, equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. A method for detecting components of a power transmission line, characterized in that: Methods include: Collect images of transmission line components; Inputting the component image into an object detection model, the object detection model includes a backbone network, a neck network, and a head network; the backbone network includes an image feature extraction module and a large kernel spatial convolution module, and the neck network includes an asymmetric padding feature fusion module; Acquire multi-scale image features of the component image according to the image feature extraction module, and acquire detail features of the component image according to the large kernel spatial convolution module; Inputting the multi-scale image features and the detail features into the neck network, and fusing the multi-scale image features and the detail features according to the asymmetric filling feature fusion module of the neck network to obtain a component fusion feature map; The component fusion feature map is transmitted to the head network, and the head network detects the component fusion feature map to generate a final detection result.

2. The component detection method for a power transmission line according to claim 1, characterized in that: The image feature extraction module includes a convolution layer, a segmentation layer, a bottleneck submodule, and a splicing layer for feature aggregation; wherein the bottleneck submodule includes an integrated exponential moving average mechanism.

3. The component detection method for a power transmission line according to claim 2, characterized in that: The step of obtaining the multi-scale image features of the component image according to the image feature extraction module specifically includes: Inputting the component image into the first convolutional layer, obtaining the component image channel dimension features and the component image spatial dimension features, and transforming the two to obtain a transformed component image; Inputting the converted component image into a segmentation layer to obtain a plurality of branch graphs; Inputting the plurality of branch graphs into the bottleneck submodule in sequence to weight them, and obtaining a plurality of weighted branch graphs; Inputting the weighted branch graphs into a splicing layer for aggregation to obtain a splicing component image; The spliced ​​component image is input into the second convolutional layer to obtain corresponding multi-scale image features.

4. The component detection method for a power transmission line according to claim 1, characterized in that: The network structure of the large-kernel spatial convolution module includes: a large-kernel convolution sequence, an average pooling layer, a maximum pooling layer, a spatial attention layer and a convolution layer.

5. The component detection method for a power transmission line according to claim 4, characterized in that: The obtaining of detail features of the component image according to the large-kernel spatial convolution module specifically includes: According to the two large kernel convolution sequences, the features of the convolution kernels of different receptive fields of the component image are obtained and connected: in, Indicates the features after connection; The connected features are pooled according to the average pooling layer and the maximum pooling layer: Among them, SA avg and SA max Represent the spatial descriptors after average pooling and maximum pooling respectively; The pooled spatial descriptor is input into the spatial attention layer for interaction to obtain a spatial attention map: in, Represents the spatial attention map; Apply the activation function to the spatial attention map to obtain the decoupled kernel features: in, represents the decoupled kernel feature; Obtain the mask corresponding to the decoupling kernel feature, perform convolution on the decoupling kernel feature and the corresponding mask to obtain the detail features of the component image: Among them, S represents the detail features of the component image, represents the decoupled kernel feature, Indicates the corresponding mask.

6. The component detection method for a power transmission line according to claim 1, characterized in that: The asymmetric padding feature fusion module includes four groups of parallel asymmetric padding convolution operation units, each group of asymmetric padding convolution operation units is configured with a different padding parameter P, which is used to directionally expand the receptive field to capture edge gradient features; the asymmetric padding convolution operation units are sequentially connected to the normalization layer and the splicing layer.

7. The component detection method for power transmission lines according to claim 6, characterized in that: The step of inputting the multi-scale image features and the detail features into the neck network, fusing the multi-scale image features and the detail features according to the asymmetric filling feature fusion module of the neck network to obtain a component fusion feature map specifically includes: The multi-scale image features and detail features are combined into an input component feature map X(h1, w1, c1), and the input component feature map is asymmetrically padded using different padding parameters P to generate a feature map with an extended receptive field: Four sets of asymmetric padding convolution operations are performed in parallel on the feature map of the expanded receptive field: in, represents the convolution operator, represents a 1×3 kernel with C′ output channels, The padding parameter P represents the pixel count in each spatial direction; Get the dimension of the component feature map output after the asymmetric padding convolution operation: Among them, C2 represents the final output channel number of the feature fusion module with asymmetric padding, and s represents Convolution stride; Perform interleaved convolution on the component feature map dimensions to obtain the component fusion feature map: Among them, X′ (h′,w′,4c) Represents the component fusion feature map.

8. The component detection method for a power transmission line according to claim 1, characterized in that: The backbone network also includes a pooling module, an image feature extraction module, a convolution layer and a large-kernel spatial convolution module; the neck network also includes an upsampling layer, a splicing layer, a convolution layer and a feature fusion module with asymmetric padding; the head network includes several detection layers.

9. The component detection method for a power transmission line according to claim 1, characterized in that: The collecting of component images of the transmission line specifically includes: Capture images of insulators, vibration absorbers, and bolts from multiple angles and at multiple times under conditions of strong light on sunny days, weak light on cloudy days, low light at night, rainy days, foggy days, snowy days, and in mountainous areas, plains, and water areas. The components of the transmission line include insulators, conductors, poles and towers, hardware, lightning protection devices, vibration absorbers, bolts, bird nests and separators.

10. A component detection system for a power transmission line, characterized in that: The system comprises: a transmission line component image acquisition unit and a model detection unit; The power transmission line component image acquisition unit is used to acquire images of power transmission line components; The model detection unit has a built-in target detection model, and the target detection model includes a backbone network, a neck network and a head network; The backbone network is configured with an image feature extraction module and a large-kernel spatial convolution module, wherein the image feature extraction module is used to obtain multi-scale image features of the component image, and the large-kernel spatial convolution module is used to obtain detail features of the component image; The neck network is configured with an asymmetric filled feature fusion module for receiving multi-scale image features and detail features output by the backbone network and fusing them to obtain a component fusion feature map; The head network is used to receive the component fusion feature map output by the neck network, detect the component fusion feature map and generate a final detection result.

11. A readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 9.

12. A computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 9.

Images