Transmission line power components and abnormal target detection method and equipment

Abstract

The present invention provides a method and device for detecting power components and abnormal targets in power transmission lines. The method comprises: Step 1, a feature extraction phase, replacing the ResNet101 backbone feature extraction network with MobileNetV3; Step 2, a decoding phase, inputting the features output by the bottleneck3, bottleneck5, and bottleneck7 structures in the MobileNetV3 feature extraction network into a feature enhancement structure, and fusing the output features of the feature enhancement structure with the features of the upsampling layer via residual attention fusion; Step 3, a training phase, encoding the target's true value frame as the object's center point coordinates and object size information, and projecting the object's center point onto a heat map using an elliptical Gaussian scattering kernel. The present invention improves the matching degree between the predicted frame and the true value frame.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of machine vision technology, and in particular to a method and device for detecting power components and abnormal targets in a transmission line. Background Art

[0002] Currently, deep learning-based methods are widely used in the field of power transmission line component and abnormal target detection. Methods that combine drone aerial photography technology with deep learning can significantly improve work efficiency and protect the lives of power inspection personnel. Common detection models in the industrial field require a selective search algorithm to select all possible prediction boxes for each category on the feature map to accurately match the target's true value box. While these anchor-based object detection methods have high detection accuracy, they require post-processing to remove redundant prediction frames, resulting in long detection inference times and a very high computational load. This also increases the overall complexity of the model and prevents full end-to-end network training. The Centernet network first predicts the center point of the target and then directly regresses the target's width and height information at the predicted center point to generate a predicted box surrounding the target. This anchor-free detection algorithm can greatly shorten the model's inference speed. When power companies use drones equipped with power components and abnormal target detection algorithms to conduct power inspections on transmission lines, they often need to consider the network model's lightweight, high-precision, and fast characteristics to adapt to the limited storage capacity of intelligent inspection equipment. In addition, the CenterNet algorithm has difficulty in achieving rapid and accurate target detection due to the large number of power components and abnormalities in high-voltage overhead transmission lines, and the model has a large number of parameters and computational complexity. Therefore, the development of a method and device for detecting power components and abnormal targets on transmission lines that can effectively overcome the shortcomings of the above-mentioned related technologies has become a technical problem that needs to be urgently solved in the industry. Summary of the Invention

[0003] In view of the above problems existing in the prior art, an embodiment of the present invention provides a method and device for detecting power components and abnormal targets in a transmission line.

[0004] In the first aspect, an embodiment of the present invention provides a method for detecting power components and abnormal targets in a transmission line, including: step 1, feature extraction stage, replacing the ResNet101 backbone feature extraction network with MobileNetV3; step 2, decoding stage, inputting the features output by the bottleneck3, bottleneck5 and bottleneck7 structures in the MobileNetV3 feature extraction network into the feature enhancement structure, and fusing the output features of the feature enhancement structure with the upsampling layer features through residual attention fusion; step 3, training stage, encoding the true value frame of the target into the center point coordinates and size information of the object, and projecting the center point of the object onto the heat map through an elliptical Gaussian scattering kernel.

[0005] Based on the content of the above method embodiment, the method for detecting power components and abnormal targets in transmission lines provided in the embodiment of the present invention, step 1 specifically includes: step 1.1, inputting a 512×512 three-channel power inspection image into the MobileNetV3 network, and the feature map size obtained after 3×3 convolution and seven bottleneck structures is 16×16, and the number of channels is 320; step 1.2, the bottleneck in step 1.1 is different from the residual block of ResNet. The bottleneck in step 1.1 adopts an inverted residual structure. The input feature matrix is ​​dimensionally upgraded by 1×1 convolution to increase the channel size, processed by 3×3 depth convolution, and reduced in dimension by 1×1 convolution. MobileNetV3 introduces the SE channel attention module based on the MobileNetV2 structure, and the SE module constructs the mutual dependence relationship between channels.

[0006] Based on the content of the above method embodiment, the power transmission line power component and abnormal target detection method provided in the embodiment of the present invention, step 2 specifically includes: step 2.1, inputting the bottleneck3, bottleneck5, and bottleneck7 layer features of the feature extraction network into the feature enhancement structure, and outputting features with attention channel level and features with multi-scale receptive field respectively, so as to improve the network model's detection ability for small targets and multi-scale targets; step 2.2, dividing the features output by the bottleneck3 layer in the MobileNetV3 structure into two paths, one of which transforms the channel and size of the features and adds them to the bottleneck5 output features to obtain intermediate-level features, and the other path is a direct connection of the features output by bottleneck3 to an efficient channel attention. force mechanism ECA to obtain attention to channel features; add a skip branch between the bottleneck5 layer and the efficient channel attention ECA; add the features output by bottleneck7 and the intermediate-level features, and extract multi-scale features through the void space pyramid structure; step 2.3, to solve the problem that CenterNet cannot fully fit two similar features through summation operation in the decoding network feature recovery stage, a residual attention feature fusion module RAFF is designed; the three features in step 2.1 are output after the feature enhancement structure designed in step 2.2, and the output features of the skip branch, deconvolution layer 1 and deconvolution layer 2 are recorded as B{B1, B2, B3}; use the RAFF module to fit the two feature sets A and B.

[0007] Based on the content of the above method embodiment, the method for detecting power components and abnormal targets in transmission lines provided in the embodiment of the present invention, step 3 specifically includes: introducing an elliptical Gaussian kernel to generate a heat map label to guide the positioning of the prediction box in target detection. Compared with the circular Gaussian kernel, the elliptical Gaussian kernel can better make the prediction box accurately match the target's true value box for targets with inconsistent aspect ratios. The elliptical Gaussian kernel function Y xyc include:

[0008]

[0009] in, is the first intermediate variable, W represents the width of the target, is the second intermediate variable, H represents the height of the target, min_iou takes the minimum intersection-union ratio as 0.7, x is the horizontal coordinate of the center point of the target, y is the vertical coordinate of the center point of the target, is the horizontal coordinate of the predicted target center, is the vertical coordinate of the predicted target center.

[0010] In the second aspect, an embodiment of the present invention provides a transmission line power component and abnormal target detection device, including: a first main module, used to implement step 1, the feature extraction stage, replacing the ResNet101 backbone feature extraction network with MobileNetV3; a second main module, used to implement step 2, the decoding stage, inputting the features output by the bottleneck3, bottleneck5 and bottleneck7 structures in the MobileNetV3 feature extraction network into the feature enhancement structure, and fusing the output features of the feature enhancement structure with the upsampling layer features through residual attention fusion; a third main module, used to implement step 3, the training stage, encoding the true value frame of the target into the center point coordinates and size information of the object, and projecting the center point of the object onto the heat map through an elliptical Gaussian scattering kernel.

[0011] In a third aspect, an embodiment of the present invention provides an electronic device, including:

[0012] at least one processor; and

[0013] at least one memory communicatively coupled to the processor, wherein:

[0014] The memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the power transmission line power component and abnormal target detection method provided by any one of the various implementations of the first aspect.

[0015] In a fourth aspect, an embodiment of the present invention provides a non-transitory computer-readable storage medium, which stores computer instructions, and the computer instructions enable a computer to execute the transmission line power component and abnormal target detection method provided by any one of the various implementation methods of the first aspect.

[0016] The method and device for detecting power components and abnormal targets in transmission lines provided by the embodiments of the present invention improve the detection accuracy of the model by constructing a feature enhancement structure and introducing underlying detailed features. A residual attention feature fusion module is designed to replace the fusion method of direct feature addition in the decoding network, fit the same-level features from different branches, and use an elliptical Gaussian kernel instead of a circular Gaussian kernel for label encoding to improve the matching degree between the prediction box and the true value box. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are 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.

[0018] Figure 1 A flow chart of a method for detecting power components and abnormal targets in a transmission line according to an embodiment of the present invention;

[0019] Figure 2 A schematic diagram of the structure of a power transmission line power component and an abnormal target detection device provided by an embodiment of the present invention;

[0020] Figure 3 A schematic diagram of the physical structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, 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 part of the embodiments of the present invention, not all of 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. In addition, the technical features in the various embodiments or single embodiments provided by the present invention can be arbitrarily combined with each other to form a feasible technical solution. This combination is not restricted by the sequence of steps and / or structural composition mode, but must be based on the ability of ordinary technicians in this field to implement it. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0022] The embodiment of the present invention provides a method for detecting power components and abnormal targets in a transmission line. Figure 1 The method includes: step 1, feature extraction stage, replacing the ResNet101 backbone feature extraction network with MobileNetV3; step 2, decoding stage, inputting the features output by the bottleneck3, bottleneck5 and bottleneck7 structures in the MobileNetV3 feature extraction network into the feature enhancement structure, and fusing the output features of the feature enhancement structure with the upsampling layer features through residual attention fusion; step 3, training stage, encoding the target's true value frame into the center point coordinates and size information of the object, and projecting the center point of the object onto the heat map through the elliptical Gaussian scattering kernel.

[0023] Based on the content of the above method embodiment, as an optional embodiment, the transmission line power component and abnormal target detection method provided in the embodiment of the present invention, step 1 specifically includes: step 1.1, inputting a 512×512 three-channel power inspection image into the MobileNetV3 network, and the feature map size obtained after 3×3 convolution and seven bottleneck structures is 16×16, and the number of channels is 320; step 1.2, the bottleneck in step 1.1 is different from the residual block of ResNet. The bottleneck in step 1.1 adopts an inverted residual structure. The input feature matrix is ​​dimensionally upgraded by 1×1 convolution to increase the channel size, processed by 3×3 depth convolution, and reduced in dimension by 1×1 convolution. MobileNetV3 introduces the SE channel attention module based on the MobileNetV2 structure. The SE module constructs the mutual dependence relationship between channels.

[0024] Based on the content of the above method embodiment, as an optional embodiment, the power transmission line power component and abnormal target detection method provided in the embodiment of the present invention, step 2 specifically includes: step 2.1, inputting the bottleneck3, bottleneck5, and bottleneck7 layer features of the feature extraction network into the feature enhancement structure, and outputting features with the attention channel level and features with multi-scale receptive fields respectively, thereby improving the network model's detection ability for small targets and multi-scale targets; step 2.2, dividing the features output by the bottleneck3 layer in the MobileNetV3 structure into two paths, one of which transforms the channel and size of the features and adds them to the bottleneck5 output features to obtain intermediate-level features, and the other path is a direct connection to the features output by bottleneck3 for efficient The channel attention mechanism ECA is used to obtain attention to channel features; a skip branch is added between the bottleneck5 layer and the efficient channel attention ECA; the features output by bottleneck7 are added to the intermediate-level features, and the features are extracted by the void space pyramid structure to obtain multi-scale features; step 2.3, in order to solve the problem that CenterNet cannot fully fit two similar features through summation operation in the decoding network feature recovery stage, a residual attention feature fusion module RAFF is designed; the three features in step 2.1 are output after the feature enhancement structure designed in step 2.2, and the output features of the skip branch, deconvolution layer 1 and deconvolution layer 2 are recorded as B{B1, B2, B3}; the RAFF module is used to fit the two feature sets A and B.

[0025] Based on the content of the above method embodiment, as an optional embodiment, the power transmission line power component and abnormal target detection method provided in the embodiment of the present invention, step 3 specifically includes: introducing an elliptical Gaussian kernel to generate a heat map label to guide the positioning of the prediction box in target detection. Compared with the circular Gaussian kernel, the elliptical Gaussian kernel can better make the prediction box accurately match the target's true value box for targets with inconsistent aspect ratios. The elliptical Gaussian kernel function Y xyc include:

[0026]

[0027] in, is the first intermediate variable, W represents the width of the target, is the second intermediate variable, H represents the height of the target, min_iou takes the minimum intersection-union ratio as 0.7, x is the horizontal coordinate of the center point of the target, y is the vertical coordinate of the center point of the target, is the horizontal coordinate of the predicted target center, is the vertical coordinate of the predicted target center.

[0028] The method for detecting power components and abnormal targets in transmission lines provided by an embodiment of the present invention improves the detection accuracy of the model by constructing a feature enhancement structure and introducing underlying detailed features. A residual attention feature fusion module is designed to replace the fusion method of direct feature addition in the decoding network, fits the same-level features from different branches, and uses an elliptical Gaussian kernel instead of a circular Gaussian kernel for label encoding to improve the matching degree between the prediction box and the true value box.

[0029] In another embodiment, this embodiment selects the anchor-free detection algorithm CenterNet in the process of realizing the detection of power components and abnormal targets of transmission lines.

[0030] The CenterNet algorithm does not require pre-setting anchor boxes, thus simplifying the network structure. Anchor-free detection algorithms are primarily composed of two categories: center-based detection algorithms and keypoint-based detection algorithms. Center-based anchor-free detection algorithms use the central area or center point of an object to determine positive samples, and then regress the distance from the center point to the four sides to produce a bounding box. Typical center-based algorithms include FCOS and SAPD. Another keypoint-based anchor-free detection algorithm identifies the key points of an object and then regresses the bounding box of the target. For example, CornetNet transforms the target detection problem into detecting a pair of key points of an object without using anchor points, and then uses the top left and bottom right corners of the target to predict the object's bounding box. Unlike the aforementioned models that detect multiple key points, CenterNet uses the target as a key point, determines the center coordinates of the object through Gaussian operations, and then regresses the target's size and position.

[0031] This embodiment directly processes aerial transmission line images or videos captured by drones, identifying the locations and categories of six types of power components and two types of abnormal targets from transmission line inspection images. This approach consists of four parts: 1. A lightweight feature extraction network; 2. Building a feature enhancement architecture; 3. A residual attention feature fusion architecture; and 4. Elliptical Gaussian scattering kernel optimized label encoding.

[0032] Specifically, the present invention includes: (1) replacing the feature extraction network of the original CenterNet with MobileNetV3 to achieve the purpose of lightweight and shorten the detection inference time of the network model; (2) since the CenterNet network only uses the output of the bottleneck7 layer for feature upsampling after feature extraction, and the output features of the bottleneck7 layer only have more high-dimensional semantic information, the low-dimensional detail information is lost in the downsampling process, resulting in low detection accuracy of the model, the present invention designs a feature enhancement FE structure to improve the detection performance; (3) the output features of the feature enhancement structure and the output features of each upsampling layer are fused through the residual attention feature fusion structure to achieve the fitting of the same-level features; (4) when training the network, the elliptical Gaussian scattering kernel is used to optimize the label encoding to generate the heat map label. For targets with large differences in aspect ratio, after improvement, the major axis (or minor axis) of the ellipse can be used to control the key point to move in the long side (or short side) direction of the bounding box, so that the predicted box accurately matches the true value box of the target.

[0033] This embodiment is implemented through the following technical solution: a method for detecting power components and abnormalities in a transmission line based on an improved CenterNet, comprising the following steps:

[0034] S1. Lightweight feature extraction network. The original CenterNet target detection algorithm uses ResNet101 for feature extraction, which results in too many network model parameters and long model detection and inference time. The MobileNetV3 network, which is friendly to embedded devices, is used to replace the original CenterNet feature extraction network to achieve the purpose of lightweighting and improve the detection and inference speed of the network model.

[0035] S2. Feature enhancement structure. In view of the fact that the original CenterNet target detection algorithm is prone to loss of target information during the upsampling process, a multi-channel feature enhancement structure is constructed. The differences between the feature enhancement structure and the three-branch FPN structure are as follows: (1) The features output by the bottleneck3 layer in the MobileNetV3 structure are divided into two paths. One path transforms the channel and size of the features and adds them to the bottleneck5 output features to obtain intermediate-level features. The other path is the feature output by bottleneck3 directly connected to the efficient channel attention mechanism ECA to obtain attention to the channel features; (2) A crossing branch is added between the bottleneck5 layer and the efficient channel attention ECA; (3) The features output by bottleneck7 are added to the intermediate-level features, and then this feature is extracted through the hollow spatial pyramid structure to obtain multi-scale features.

[0036] S3. Residual attention feature fusion structure. The main purpose of designing this structure is to fit two levels of similar features. The output of the feature enhancement structure and the upsampled features are fitted using the residual attention fusion module. Experiments have shown that this structure can better adaptively emphasize the information that needs to be supplemented.

[0037] S4. Optimize label encoding. The CenterNet network label encoding method uses a circular Gaussian scattering kernel. Circular Gaussian scattering kernel label encoding is friendly to targets with consistent length and width, but the length and width of objects are usually inconsistent. When detecting objects with a large difference between length and width, the predicted center point should move the same distance in the horizontal and vertical directions. During bounding box regression, the predicted box obtained does not completely surround the true value box. Therefore, using an elliptical Gaussian scattering kernel label encoding method is beneficial to the bounding box regression of objects with inconsistent length and width, improving the positioning quality of the predicted box.

[0038] In specific implementation, a method for detecting power components and abnormal targets on transmission lines based on an improved CenterNet is proposed to achieve rapid and accurate detection of various power components and abnormal targets on high-voltage transmission lines. The present invention uses 3,000 images collected by drone aerial photography of transmission lines for annotation to prepare for subsequent training. In view of the low detection accuracy and large number of model parameters of the original CenterNet algorithm, the present invention uses MobileNetV3 as the feature extraction network in CenterNet and adjusts the number of output channels of the upsampling layer; at the same time, the detection ability of the model is improved by constructing a feature enhancement structure; in addition, a residual attention feature fusion module is designed to replace the method of directly adding and fusing features in the upsampling process to supplement the missing information of the two different branch features; finally, the method uses an elliptical Gaussian kernel to optimize the label encoding to improve the quality of target prediction box regression. The steps include: lightweight feature extraction network stage, multi-channel feature enhancement structure, same-scale residual attention feature fusion structure, and optimized label encoding.

[0039] 1) In the lightweight feature extraction network stage, the original CenterNet target detection algorithm uses ResNet101 for feature extraction, which results in too many network model parameters and long model detection and inference time. Therefore, this paper uses the embedded-friendly MobileNetV3 feature extraction to replace ResNet101.

[0040] 2) Multi-channel feature enhancement structure, which connects the three layers of features of the MobileNetV3 feature extraction network respectively, and then connects the efficient attention mechanism, efficient attention mechanism, and void space pyramid in series respectively.

[0041] 3) Residual attention feature fusion structure: The output of the feature enhancement structure in step 2) and the output features of the upsampling layer are subjected to the feature fusion structure designed by the present invention to adaptively emphasize the feature information that needs to be supplemented.

[0042] 4) Optimizing label encoding: Using elliptical Gaussian scattering kernel label encoding to generate heatmaps is beneficial for bounding box regression of targets with inconsistent aspect ratios.

[0043] The original CenterNet algorithm simply uses the last layer of features extracted by the backbone network to predict the target. In order to improve the detection accuracy, a FE feature enhancement structure is designed, and the output features of the bottleneck3, bottleneck5 and bottleneck7 layers of MobileNetV3 are selected for feature enhancement. The differences between the FE structure and the three-branch FPN structure are as follows: (1) The features output by the bottleneck3 layer in the MobileNetV3 structure are divided into two paths. One path transforms the channel and size of the features and adds them to the bottleneck5 output features to obtain intermediate-level features. The other path is the feature output of bottleneck3 directly connected to the efficient channel attention mechanism ECA to obtain attention to channel features; (2) A crossing branch is added between the bottleneck5 layer and the efficient channel attention ECA; (3) The features output by bottleneck7 are added to the intermediate-level features, and then this feature is extracted through the hollow spatial pyramid structure to obtain multi-scale features. In general, this feature enhancement structure can obtain attention to channel features and realize the extraction of multi-scale features.

[0044] In order to solve the problem that CenterNet cannot fully fit two similar features through summation operation in the decoding network feature recovery stage, the present invention designs a residual attention feature fusion module RAFF; the present invention outputs feature A{A1, A2, A3} after the three features in step 2.1 are subjected to the feature enhancement structure designed in step 2.2, and the output features of the skip branch, deconvolution layer 1 and deconvolution layer 2 are recorded as B{B1, B2, B3}; then, the RAFF module designed by the invention is used to fit the two feature sets A and B.

[0045] When performing label encoding, an elliptical Gaussian kernel is introduced to generate a heat map label to guide the positioning of the prediction box in target detection. Compared with the circular Gaussian kernel, the elliptical Gaussian kernel can better make the prediction box accurately match the target's true value box for targets with inconsistent aspect ratios. The elliptical Gaussian kernel function is shown in formula (1).

[0046] The implementation basis of each embodiment of the present invention is achieved through programmed processing by a device with processor functions. Therefore, in engineering practice, the technical solutions and functions of each embodiment of the present invention can be encapsulated into various modules. Based on this reality, on the basis of the above embodiments, an embodiment of the present invention provides a power transmission line power component and abnormal target detection device, which is used to execute the power transmission line power component and abnormal target detection method of the above method embodiment. Figure 2The device includes: a first main module, used to implement step 1, the feature extraction stage, replacing the ResNet101 backbone feature extraction network with MobileNetV3; a second main module, used to implement step 2, the decoding stage, inputting the features output by the bottleneck3, bottleneck5 and bottleneck7 structures in the MobileNetV3 feature extraction network into the feature enhancement structure, and fusing the output features of the feature enhancement structure with the upsampling layer features through residual attention fusion; a third main module, used to implement step 3, the training stage, encoding the target's true value frame into the center point coordinates and size information of the object, and projecting the center point of the object onto the heat map through an elliptical Gaussian scattering kernel.

[0047] The power transmission line power component and abnormal target detection device provided by the embodiment of the present invention adopts Figure 2 Several modules in the proposed method improve the detection accuracy of the model by constructing a feature enhancement structure and introducing underlying detail features. A residual attention feature fusion module is designed to replace the direct feature addition fusion method in the decoding network, fit the same-level features from different branches, and use an elliptical Gaussian kernel instead of a circular Gaussian kernel for label encoding to improve the matching degree between the predicted box and the true value box.

[0048] It should be noted that the device in the device embodiment provided by the present invention can be used to implement the method in the above-mentioned method embodiment as well as the method in other method embodiments provided by the present invention. The only difference is that the corresponding functional modules are set. The principle is basically the same as the principle of the above-mentioned device embodiment provided by the present invention. As long as those skilled in the art refer to the specific technical solutions in other method embodiments on the basis of the above-mentioned device embodiment, obtain the corresponding technical means and the technical solutions composed of these technical means by combining technical features, and ensure the practicality of the technical solutions, they can improve the device in the above-mentioned device embodiment to obtain the corresponding device class embodiment, thereby obtaining the corresponding device class embodiment for implementing the methods in other method class embodiments. For example:

[0049] Based on the content of the above-mentioned device embodiment, as an optional embodiment, the transmission line power component and abnormal target detection device provided in the embodiment of the present invention further includes: a first submodule, which is used to implement step 1, specifically including: step 1.1, inputting a 512×512 three-channel power inspection image into the MobileNetV3 network, and the feature map size obtained after 3×3 convolution and seven bottleneck structures is 16×16, and the number of channels is 320; step 1.2, the bottleneck in step 1.1 is different from the residual block of ResNet. The bottleneck in step 1.1 adopts an inverted residual structure. The input feature matrix is ​​dimensionally upgraded through 1×1 convolution to increase the channel size, processed through 3×3 depth convolution, and reduced in dimension through 1×1 convolution. MobileNetV3 introduces the SE channel attention module based on the MobileNetV2 structure. The SE module constructs the mutual dependence relationship between channels.

[0050] Based on the content of the above-mentioned device embodiment, as an optional embodiment, the power transmission line power component and abnormal target detection device provided in the embodiment of the present invention further includes: a second submodule, which is used to implement step 2, specifically including: step 2.1, inputting the bottleneck3, bottleneck5, and bottleneck7 layer features of the feature extraction network into the feature enhancement structure, and outputting features with the attention channel level and features with multi-scale receptive fields respectively, so as to improve the network model's detection ability for small targets and multi-scale targets; step 2.2, dividing the features output by the bottleneck3 layer in the MobileNetV3 structure into two paths, one of which converts the channel and size of the features and adds them to the bottleneck5 output features to obtain intermediate-level features, and the other path is the bottleneck3 output The features are directly connected to the efficient channel attention mechanism ECA to obtain attention to the channel features; a skip branch is added between the bottleneck5 layer and the efficient channel attention ECA; the features output by bottleneck7 are added to the intermediate-level features, and the features are extracted by the void space pyramid structure to obtain multi-scale features; step 2.3, in order to solve the problem that CenterNet cannot fully fit two similar features through summation operation in the decoding network feature recovery stage, a residual attention feature fusion module RAFF is designed; the three features in step 2.1 are output after the feature enhancement structure designed in step 2.2, and the output features of the skip branch, deconvolution layer 1 and deconvolution layer 2 are recorded as B{B1, B2, B3}; the RAFF module is used to fit the two feature sets A and feature set B.

[0051] Based on the content of the above device embodiment, as an optional embodiment, the power transmission line power component and abnormal target detection device provided in the embodiment of the present invention further includes: a third submodule for implementing step 3 specifically including: introducing an elliptical Gaussian kernel to generate a heat map label to guide the positioning of the prediction box in target detection. Compared with the circular Gaussian kernel, the elliptical Gaussian kernel can better make the prediction box accurately match the target's true value box for targets with inconsistent aspect ratios. The elliptical Gaussian kernel function Y xyc include:

[0052]

[0053] in, is the first intermediate variable, W represents the width of the target, is the second intermediate variable, H represents the height of the target, min_iou takes the minimum intersection-union ratio as 0.7, x is the horizontal coordinate of the center point of the target, y is the vertical coordinate of the center point of the target, is the horizontal coordinate of the predicted target center, is the vertical coordinate of the predicted target center.

[0054] The method of the embodiment of the present invention is implemented by electronic devices, so it is necessary to introduce the relevant electronic devices. Based on this purpose, the embodiment of the present invention provides an electronic device, such as Figure 3 As shown, the electronic device includes: at least one processor, a communications interface, at least one memory, and a communications bus, wherein the at least one processor, the communications interface, and the at least one memory communicate with each other via the communications bus. The at least one processor can call logic instructions in the at least one memory to execute all or part of the steps of the methods provided in the aforementioned method embodiments.

[0055] In addition, the logic instructions in the at least one memory mentioned above can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each method embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0056] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0057] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course can also be implemented by hardware. Based on this understanding, the essence of the above technical solution or the part that contributes to the existing technology can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiment.

[0058] The flowcharts and block diagrams in the accompanying drawings show the possible architectures, functions and operations of the systems, methods and computer program products according to multiple embodiments of the present invention. Based on this understanding, each box in the flowchart or block diagram can represent a module, program segment or part of the code, and the module, program segment or part of the code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in an order different from that marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, or sometimes in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of boxes in the block diagram and / or flowchart, can be implemented with a dedicated hardware-based system that performs the specified function or action, or can be implemented with a combination of dedicated hardware and computer instructions.

[0059] It should be noted that the terms "comprise," "include," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or apparatus comprising a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, the elements defined by the phrase "comprise..." do not preclude the presence of additional identical elements in the process, method, article, or apparatus comprising the elements.

[0060] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for detecting power components and abnormal targets in a transmission line, characterized in that: include: In step 1, during the feature extraction phase, the ResNet101 backbone feature extraction network is replaced with MobileNetV3. In step 2, during the decoding phase, the features output by the bottleneck3, bottleneck5, and bottleneck7 structures in the MobileNetV3 feature extraction network are input into the feature enhancement structure, and the output features of the feature enhancement structure are fused with the features of the upsampling layer through residual attention fusion. In step 3, during the training phase, the target's true value frame is encoded as the center point coordinates and size information of the object, and the center point of the object is projected onto the heat map through an elliptical Gaussian scattering kernel. Among them, step 2 includes: step 2.1, inputting the bottleneck3, bottleneck5, and bottleneck7 layer features of the feature extraction network into the feature enhancement structure, and outputting features with attention channel level and features with multi-scale receptive field respectively, so as to improve the network model's detection ability for small targets and multi-scale targets; step 2.2, dividing the features output by the bottleneck3 layer in the MobileNetV3 structure into two paths, one of which transforms the channel and size of the features and adds them to the bottleneck5 output features to obtain the intermediate-level features, and the other path is the bottleneck3 output The features of the bottleneck7 layer are directly connected to the efficient channel attention mechanism ECA to obtain attention to the channel features; a skip branch is added between the bottleneck5 layer and the efficient channel attention ECA; the features output by bottleneck7 are added to the intermediate-level features, and the features are extracted by the hollow spatial pyramid structure to obtain multi-scale features; step 2.3, in order to solve the problem that CenterNet cannot fully fit two similar features through the summation operation in the decoding network feature recovery stage, a residual attention feature fusion module RAFF is designed; the three features in step 2.1 are output after the feature enhancement structure designed in step 2.2, and the output features of the skip branch, deconvolution layer 1 and deconvolution layer 2 are recorded as B{B1, B2, B3}; the RAFF module is used to fit the two feature sets A and B; Step 3 includes: introducing an elliptical Gaussian kernel to generate a heat map label to guide the positioning of the prediction box in target detection. Compared with the circular Gaussian kernel, the elliptical Gaussian kernel can better match the prediction box accurately with the target's true value box for targets with inconsistent aspect ratios. The elliptical Gaussian kernel function include: in, is the first intermediate variable, W represents the width of the target, is the second intermediate variable, H represents the height of the target, Take the minimum intersection-union ratio as 0.7, x is the horizontal coordinate of the center point of the target, y is the vertical coordinate of the center point of the target, is the horizontal coordinate of the predicted target center, is the vertical coordinate of the predicted target center.

2. The method for detecting power components and abnormal targets in a transmission line according to claim 1, wherein: Step 1 specifically includes: Step 1.1, input a 512×512 three-channel power inspection image into the MobileNetV3 network. After 3×3 convolution and seven bottleneck structures, the feature map size is 16×16 and the number of channels is 320; Step 1.2, the bottleneck in step 1.1 is different from the residual block of ResNet. The bottleneck in step 1.1 adopts an inverted residual structure. The input feature matrix is ​​dimensionally upgraded through 1×1 convolution to increase the channel size, processed through 3×3 depth convolution, and reduced through 1×1 convolution. MobileNetV3 introduces the SE channel attention module based on the MobileNetV2 structure. The SE module builds the interdependence between channels.

3. A transmission line power component and abnormal target detection device, the transmission line power component and abnormal target detection device is used to implement the transmission line power component and abnormal target detection method according to claim 1, characterized in that: include: The first main module is used to implement step 1, the feature extraction stage, replacing the ResNet101 backbone feature extraction network with MobileNetV3; The second main module is used to implement step 2, the decoding stage, inputting the features output by the bottleneck3, bottleneck5 and bottleneck7 structures in the MobileNetV3 feature extraction network into the feature enhancement structure, and fusing the output features of the feature enhancement structure with the upsampling layer features through residual attention fusion; the third main module is used to implement step 3, the training stage, encoding the target's true value frame into the center point coordinates and size information of the object, and projecting the center point of the object onto the heat map through the elliptical Gaussian scattering kernel.

4. An electronic device, characterized in that: include: At least one processor, at least one memory and a communication interface; wherein, The processor, memory and communication interface communicate with each other; The memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the method according to any one of claims 1 to 2.

5. A non-transitory computer-readable storage medium, characterized in that The non-transitory computer-readable storage medium stores computer instructions, which cause the computer to execute the method of any one of claims 1 to 2.

Images