CN115908999A

CN115908999A - Method for detecting corrosion of top hardware fitting of power distribution tower, medium and edge terminal equipment

Info

Publication number: CN115908999A
Application number: CN202211491468.6A
Authority: CN
Inventors: 陈超; 赵裕成; 张志勇; 艾坤; 刘海峰
Original assignee: Hefei Zhongke Leinao Intelligent Technology Co ltd
Current assignee: Hefei Zhongke Leinao Intelligent Technology Co ltd
Priority date: 2022-11-25
Filing date: 2022-11-25
Publication date: 2023-04-04
Anticipated expiration: 2042-11-25
Also published as: CN115908999B

Abstract

The invention discloses a method for detecting the corrosion of a top hardware fitting of a power distribution tower, a medium and edge terminal equipment, and relates to the technical field of image processing. The method comprises the following steps: extracting a hardware fitting area image from a top image of a power distribution tower; performing feature extraction on the hardware fitting region image through a hardware fitting corrosion detection model to obtain feature data to be processed of the hardware fitting region image; the neck network structure of the hardware rust detection model is constructed by adopting a first general convolution kernel and a first fusion convolution kernel, and the first fusion convolution kernel is constructed by adopting a first standard convolution kernel and a first depth separable convolution kernel; and fusing the characteristic data through the first standard convolution kernel and the first depth separable convolution kernel to obtain a fusion result, and generating a hardware rust detection result. According to the method, the robustness and the precision of the model are further improved by improving the convolution block of the neck structure in the traditional network model, so that the detection speed of hardware rust at the top of the power distribution tower is greatly increased.

Description

Method for detecting rust corrosion of top hardware of power distribution tower, medium and edge terminal equipment

Technical Field

The invention relates to the technical field of image processing, in particular to a method, medium and edge terminal equipment for detecting rust corrosion of a top hardware fitting of a power distribution tower.

Background

Distribution lines are important components of national power grids, and metal accessories widely used in the lines, such as iron, aluminum, or aluminum alloy accessories, are connected in series with conductors, conductors and wires, transmission line wires themselves, and insulators in distribution devices, wherein the metal accessories (made of iron, aluminum, or aluminum alloy) used for protecting the wires, insulators themselves, and the like are collectively called electric power fittings. Electric power fittings used for inter-wire connection, inter-insulator connection, insulator-tower connection, and insulator-wire connection at the top of a distribution line are called line fittings. As distribution lines in China are widely distributed and geographical environments are very complex, distribution cables and tower poles are exposed outdoors for a long time, and in many areas, the conditions of abundant rainwater and humid climate exist, oxygen can gradually oxidize metal parts under the condition of water, so that various hardware fittings are rusted and corroded to different degrees. The corrosion not only can shorten the service life of gold utensil, more probably causes the gold utensil not hard up to drop, if can't in time investigate and change, probably causes the short circuit even, destroys the power supply line, and then causes serious accidents such as forest fire. In the related technology, whether the distribution line works normally or not is judged by manually and periodically checking the distribution line, the method is low in efficiency and has certain danger, and hidden dangers cannot be timely checked when the hardware at the top of the distribution pole is rusted.

Disclosure of Invention

The present invention is directed to solving, at least to some extent, one of the technical problems in the related art.

Therefore, the first purpose of the invention is to provide a method for detecting the rust of the top hardware of the power distribution tower, which further improves the robustness and the precision of a model by improving a convolution block of a neck structure in a traditional network model, so that the detection speed of the rust of the top hardware of the power distribution tower is greatly improved.

A second object of the invention is to propose a computer-readable storage medium.

A third object of the invention is to propose an edge termination device.

In order to achieve the above object, an embodiment of a first aspect of the present invention provides a method for detecting corrosion of a top hardware fitting of a power distribution tower, where the method includes: extracting a hardware fitting area image from a top image of a power distribution tower; performing feature extraction on the hardware fitting region image through a hardware fitting corrosion detection model to obtain feature data to be processed of the hardware fitting region image; the neck network structure of the hardware rust detection model is constructed by adopting a first general convolution kernel and a first fusion convolution kernel, and the first fusion convolution kernel is constructed by adopting a first standard convolution kernel and a first depth separable convolution kernel; performing standard convolution processing on the feature data to be processed through the first standard convolution core to obtain first feature data; performing feature fusion processing on the first feature data through the first depth separable convolution kernel to obtain second feature data; and generating a hardware rust detection result based on the fusion result of the first characteristic data and the second characteristic data.

According to one embodiment of the invention, the hardware region image is extracted from the top image of the power distribution tower, and the method comprises the steps of inputting the top image of the power distribution tower into a tower top detection model; wherein the neck network structure of the tower top detection model comprises a second common convolution kernel and a second fused convolution kernel, and the second fused convolution kernel comprises a second standard convolution kernel and a second depth separable convolution kernel; performing standard convolution processing on the image at the top of the power distribution tower through the second standard convolution core to obtain third characteristic data; performing feature fusion processing on the third feature data through the second depth separable convolution kernel to obtain fourth feature data; based on the third characteristic data and the fourth characteristic data, the tower pole top detection model outputs the hardware region image.

According to an embodiment of the invention, the hardware corrosion detection model further comprises a backbone network structure, and the neck network structure is connected to the backbone network structure; the neck network structure comprises a plurality of network branches; the network branch is constructed by adopting the first general convolution kernel and the first fusion convolution kernel; the hardware fitting regional image is subjected to feature extraction through a hardware fitting corrosion detection model, and the to-be-processed feature data of the hardware fitting regional image is obtained, and the method comprises the following steps: performing feature extraction on the hardware fitting region image through the trunk network structure to obtain trunk feature data; inputting the main characteristic data into the neck network structure, and performing standard convolution processing on the main characteristic data through the first general convolution kernel of the network branch to obtain standard convolution characteristic data; and the standard convolution characteristic data is used as the characteristic data to be processed.

According to one embodiment of the invention, the network branch comprises a first general convolution kernel, a first branch convolution kernel and a second branch convolution kernel which are connected in sequence; the first branch convolution kernel and the second branch convolution kernel adopt a first fusion convolution kernel; correspondingly, after the standard convolution processing is performed on the main feature data through the first general convolution kernel of the network branch, so as to obtain standard convolution feature data, the method further includes: processing the feature data to be processed sequentially through a first standard convolution kernel and a first depth separable convolution kernel in the first branch convolution kernel to obtain second feature data; second feature data obtained based on the first branch convolution kernel corresponds to first intermediate fusion feature data output by a first standard convolution kernel in the first branch convolution kernel; and sequentially processing the first intermediate fusion characteristic data through a first standard convolution kernel and a first depth separable convolution kernel in a second branch convolution kernel of the network branch to obtain target fusion convolution characteristic data of the network branch.

According to one embodiment of the invention, the plurality of network branches includes a first network branch; the first network branch comprises a third branch convolution kernel, a first general convolution kernel, a first branch convolution kernel and a second branch convolution kernel which are connected in sequence, and the third branch convolution kernel adopts a first fusion convolution kernel; correspondingly, before the standard convolution processing is performed on the main feature data through the first general convolution kernel of the network branch, so as to obtain standard convolution feature data, the method further includes: sequentially processing the trunk characteristic data through a first standard convolution kernel and a first depth separable convolution kernel in the third branch convolution kernel to obtain second characteristic data; second feature data obtained based on the third branch convolution kernel corresponds to first feature data output by a first standard convolution kernel in the third branch convolution kernel; correspondingly, the standard convolution processing is performed on the main feature data through the first general convolution kernel of the network branch to obtain standard convolution feature data, and the standard convolution feature data includes: and performing standard convolution processing on the second intermediate fusion characteristic data through a first general convolution core of the first network branch to obtain the standard convolution characteristic data.

According to one embodiment of the invention, the plurality of network branches includes a second network branch; the second network branch comprises a first general convolution kernel, a first branch convolution kernel, a second branch convolution kernel, a fourth branch convolution kernel and a fifth branch convolution kernel which are connected in sequence, wherein the fourth branch convolution kernel adopts a first general convolution kernel, and the fifth branch convolution kernel adopts a first fusion convolution kernel; correspondingly, standard convolution processing is performed on the main feature data at the first general convolution kernel passing through the network branch to obtain standard convolution feature data, and the method further includes: performing standard convolution processing on the main characteristic data through a first general convolution kernel of the second network branch to obtain standard convolution characteristic data; correspondingly, after the standard convolution processing is performed on the main feature data through the first general convolution kernel of the network branch to obtain standard convolution feature data, the method includes: processing the standard convolution characteristic data through a first branch convolution kernel and a second branch convolution kernel in the second network branch in sequence to obtain target fusion convolution characteristic data of the second network branch; performing standard convolution processing on the target fusion convolution characteristic data of the second network branch through the fourth branch convolution core to obtain first intermediate standard convolution characteristic data; sequentially processing the first intermediate standard convolution characteristic data through a first standard convolution kernel and a first depth separable convolution kernel in the fifth branch convolution kernel to obtain second characteristic data; and the second characteristic data obtained based on the fifth branch convolution kernel corresponds to the first characteristic data output by the first standard convolution kernel in the fifth branch convolution kernel.

According to one embodiment of the invention, the plurality of network branches includes a third network branch; the third network branch comprises a sixth branch convolution kernel, a first general convolution kernel, a first branch convolution kernel and a second branch convolution kernel which are connected in sequence, and the sixth branch convolution kernel adopts a first fusion convolution kernel; correspondingly, before the standard convolution processing is performed on the main feature data through the first general convolution kernel of the network branch to obtain standard convolution feature data, the method further includes: sequentially processing the trunk characteristic data through a first standard convolution kernel and a first depth separable convolution kernel in the sixth branch convolution kernel to obtain second characteristic data; the second feature data obtained based on the sixth branch convolution kernel corresponds to the fourth intermediate fusion feature data with respect to the first feature data output by the first standard convolution kernel in the sixth branch convolution kernel; correspondingly, the performing standard convolution processing on the main feature data through the first general convolution kernel of the network branch to obtain standard convolution feature data includes: and performing standard convolution processing on the fourth intermediate fusion characteristic data through the first general convolution core of the third network branch to obtain the standard convolution characteristic data.

According to an embodiment of the present invention, the plurality of network branches include a first intermediate convolution kernel disposed between a second network branch and a third network branch, and a second intermediate convolution kernel disposed between a first network branch and a second network branch, and the first intermediate convolution kernel and the second intermediate convolution kernel respectively adopt a first fusion convolution kernel; the second intermediate convolution kernel is configured to process an output of the first network branch to obtain fifth intermediate fusion feature data, and the fifth intermediate fusion feature data is fused with the output of the second network branch to obtain sixth intermediate fusion feature data input to the first intermediate convolution kernel; correspondingly, before the standard convolution processing is performed on the fourth intermediate fusion feature data through the first general convolution kernel of the third network branch to obtain the standard convolution feature data, the method further includes: processing the sixth intermediate fusion feature data sequentially through a first general convolution kernel and a first depth separable convolution kernel in the first intermediate convolution kernel to obtain second feature data; the second characteristic data obtained based on the first intermediate convolution kernel corresponds to the seventh intermediate fusion characteristic data with the first characteristic data output by the first standard convolution kernel in the first intermediate convolution kernel; performing fusion processing on the seventh intermediate fusion feature data and the fourth intermediate fusion feature data to obtain eighth intermediate fusion feature data; performing standard convolution processing on the fourth intermediate fusion feature data and the seventh intermediate fusion feature data through the first general convolution kernel of the third network branch to obtain the standard convolution feature data, including: and performing standard convolution processing on the eighth intermediate fusion characteristic data through the first general convolution kernel of the third network branch to obtain the standard convolution characteristic data.

According to an embodiment of the present invention, based on a fusion result of the first feature data and the second feature data, includes: and multiplying the first characteristic data by the second characteristic data, and performing information fusion on the multiplication result of the first characteristic data and the second characteristic data through a shuffle function to obtain the fusion result.

According to one embodiment of the invention, the tower top detection model and the hardware rust detection model adopt EIoU Loss as a candidate box regression Loss function.

According to one embodiment of the invention, the tower pole top detection model and the hardware rust detection model adopt an exponential moving average method to carry out parameter weight correction.

In order to achieve the above embodiments, a second aspect of the present invention provides a computer-readable storage medium, on which a computer program is stored, where the computer program, when executed by a processor, implements the method for detecting corrosion of top hardware of a power distribution tower according to the above embodiments of the present invention.

In order to achieve the above embodiments, a third aspect of the present invention provides an edge terminal device, which is characterized by including a memory, a processor, and a power distribution tower top hardware corrosion detection program that is stored in the memory and can be run on the processor, where when the processor executes the power distribution tower top hardware corrosion detection program, the method for detecting power distribution tower top hardware corrosion according to the above embodiments of the present invention is implemented.

Additional aspects and advantages of the invention will be set forth in part in the description which follows and, in part, will be obvious from the description, or may be learned by practice of the invention.

Drawings

Fig. 1 is a schematic flow chart of a method for detecting corrosion of power distribution tower top hardware according to an embodiment of the invention;

FIG. 2 is a schematic diagram of a first fused convolution kernel according to one embodiment of the present invention;

fig. 3 is a schematic flow chart of a method for detecting rust corrosion of a power distribution tower top hardware according to a first embodiment of the invention;

fig. 4 is a schematic flow chart of a method for detecting rust corrosion of hardware on the top of a power distribution tower according to a second embodiment of the invention;

FIG. 5 is a schematic diagram of a network branch according to one embodiment of the present invention;

FIG. 6 is a schematic diagram of the structure of a first network branch according to one embodiment of the invention;

FIG. 7 is a schematic diagram of a second network branch, according to one embodiment of the present invention;

FIG. 8 is a schematic diagram of a second network branch, according to one embodiment of the present invention;

fig. 9 is a schematic structural diagram of a neck network of a hardware rust detection model according to an embodiment of the present invention;

fig. 10 is a schematic diagram of the structure of an edge termination device according to one embodiment of the present invention.

Detailed Description

Reference will now be made in detail to embodiments of the present invention, examples of which are illustrated in the accompanying drawings, wherein like or similar reference numerals refer to the same or similar elements or elements having the same or similar function throughout. The embodiments described below with reference to the drawings are illustrative and intended to be illustrative of the invention and are not to be construed as limiting the invention.

The method, medium and edge terminal equipment for detecting rust on the top hardware of the power distribution tower according to the embodiment of the invention are described below with reference to the accompanying drawings 1 to 10.

Fig. 1 is a schematic flow chart of a method for detecting corrosion of power distribution tower top hardware according to an embodiment of the invention.

As shown in fig. 1, the method for detecting the corrosion of the hardware on the top of the power distribution tower may include the following steps:

and S110, extracting a hardware region image from the top image of the power distribution tower.

It can be understood that the distribution line is an important component of a national power grid, the top of the distribution line is used for connecting wires, insulators and pole towers, and power fittings connected between the insulators and the wires, and equipment and conductors, conductors and wires in the distribution device, and the power fittings connected with the conductors, the conductors and the wires of the transmission line are connected into strings, when the power fittings used for protecting the wires and the insulators are corroded, the service life of the fittings can be shortened, and hidden danger troubleshooting needs to be carried out in time. In this embodiment, by acquiring an image of the top of the distribution tower, the position of the hardware region is extracted from the top of the distribution tower.

As an example, after the top image of the power distribution tower is obtained, the part of the hardware region in the top image of the power distribution tower can be cut out manually, the part of the hardware region in the top image of the power distribution tower can be detected through a network detection model, and the detected part of the hardware region image can be extracted from the top image of the power distribution tower.

For example, the top of the distribution pole may be photographed by an unmanned aerial vehicle to obtain the image at the top of the distribution pole, or the image at the top of the distribution pole may be obtained by other aerial photographing devices, which is not limited in this invention.

S120, extracting the features of the hardware regional image through the hardware corrosion detection model to obtain the to-be-processed feature data of the hardware regional image; the neck network structure of the hardware rust detection model is constructed by adopting a first general convolution kernel and a first fusion convolution kernel, and the first fusion convolution kernel is constructed by adopting a first standard convolution kernel and a first depth separable convolution kernel.

Specifically, the hardware fitting region image is input into a hardware fitting corrosion detection model to be detected, feature data to be processed of the hardware fitting region is obtained, after the feature data to be processed are obtained, the feature data to be processed are processed through a neck network of the hardware fitting corrosion detection model, wherein the neck network structure of the hardware fitting corrosion detection model is constructed by adopting a first general convolution kernel and a first fusion convolution kernel, namely the convolution kernel of the neck network comprises a first general convolution kernel and a first fusion convolution kernel, and the first fusion convolution kernel is constructed by adopting a first standard convolution kernel and a first depth separable convolution kernel.

And S130, performing standard convolution processing on the characteristic data to be processed through the first standard convolution kernel to obtain first characteristic data.

And S140, performing feature fusion processing on the first feature data through the first depth separable convolution kernel to obtain second feature data.

It can be understood that the first standard convolution kernel is used for performing standard convolution on the feature data, the first fusion convolution kernel is used for performing fusion processing on the feature data, the first fusion convolution kernel firstly processes the feature data through the first standard convolution kernel therein to obtain first feature data, and further processes the first feature data through the first depth separable convolution kernel to obtain second feature data.

And S150, generating a hardware corrosion detection result based on the fusion result of the first characteristic data and the second characteristic data.

And further, fusing the first characteristic data and the second characteristic data to obtain a fusion result, wherein the fusion result is used as the characteristic data used for generating the hardware corrosion detection model in the hardware corrosion detection model.

As a possible implementation manner, a schematic structural diagram of the first fusion convolution kernel is shown in fig. 2, where the first fusion convolution kernel GSConv is constructed by using a first standard convolution kernel Conv and a first depth-separable convolution kernel DSC, where the first standard convolution kernel Conv is used to perform standard convolution processing on the feature data, the first depth-separable convolution kernel DSC is used to perform channel-by-channel convolution processing on the feature image, that is, the first standard convolution kernel Conv is used to generate first feature data, the first feature data is further processed by the first depth-separable convolution kernel DSC to generate second feature data, and then the first feature data and the second feature data are multiplied to finally obtain a fusion result.

As an example, before obtaining a hardware corrosion detection model, labeling an image by using a label tool with respect to a hardware corrosion target detection production dataset on the top of a power distribution pole, where a region labeled in the image is a position region of hardware corrosion in the pole top, and normalizing the image to generate a label, where the format of the label is a central point and a width and a height of a category and a rectangle. The data normalization mode adopts a mode of dividing the coordinates of the labeling frame by the width and height values of the image, and the data enhancement mode can adopt modes of horizontal turning, left-right turning, tone variation, rotation enhancement and the like. And training the Yolov5 model with the improved neck convolution structure by adopting the power distribution rod top hardware corrosion target detection data set to obtain a hardware corrosion detection model.

Optionally, a preset number of rust images of the hardware at the top of the distribution pole can be shot by the unmanned aerial vehicle to manufacture a target detection data set at the top of the distribution pole.

Fig. 3 is a schematic flow chart of a method for detecting rust corrosion of a power distribution tower top fitting according to a first embodiment of the invention.

As shown in fig. 3, extracting the hardware region image from the top image of the power distribution tower may include the following steps:

s310, inputting the top image of the power distribution tower into a tower top detection model; the neck network structure of the tower top detection model comprises a second general convolution kernel and a second fusion convolution kernel, and the second fusion convolution kernel comprises a second standard convolution kernel and a second depth separable convolution kernel.

Specifically, after acquiring a power distribution pole top image, firstly, inputting the power distribution pole top image into a tower top detection model, wherein the model is used for detecting the position of hardware fittings in the power distribution pole top image, so as to acquire an area image where the hardware fittings are located in the power distribution pole top image; and inputting the hardware fitting region image to a hardware fitting corrosion detection model, wherein the model is used for detecting the position of hardware fitting corrosion in the hardware fitting region image.

Illustratively, the tower top detection model and the hardware rust detection model have the same structure, that is, the second general convolution kernel has the same structure as the first general convolution kernel, and the second fusion convolution kernel has the same structure as the first fusion convolution kernel.

As an example, before obtaining a tower top detection model, firstly, a data set is created for a power distribution pole top target detection, an image is labeled by using a label tool, wherein a region labeled in the image is a position region of the power distribution pole top, and then a label is generated after normalizing the image, wherein the label format is a category and a central point and a width and a height of a rectangle. The data normalization mode adopts a mode of dividing the coordinates of the labeling frame by the width and height values of the image, and the data enhancement mode can adopt modes of horizontal turning, left-right turning, tone variation, rotation enhancement and the like. And training the model with the improved neck convolution structure by adopting the distribution pole top target detection data set to obtain the tower pole top detection model.

It can be understood that, when carrying out the gold utensil corrosion detection in the distribution pole top of the tower image, the equipment through unmanned aerial vehicle or other acquisition image shoots and joins in marriage pole top of the tower image, the position of gold utensil corrosion is not obvious in this image, in this embodiment, at first predict the mark through the position at tower pole top in the distribution pole top of the tower image of tower pole top detection model pair acquireing, then will predict the position at the tower pole top of marking as gold utensil corrosion detection model input, whether the corrosion appears in the metal utensil in the detection tower pole top.

And S320, performing standard convolution processing on the image at the top of the power distribution tower through a second standard convolution core to obtain third characteristic data.

And S330, performing feature fusion processing on the third feature data through the second depth separable convolution kernel to obtain fourth feature data.

And S240, outputting a hardware area image by the tower pole top detection model based on the third characteristic data and the fourth characteristic data.

Optionally, the hardware region image is output through the tower rod top detection model and serves as an input for feature extraction of the hardware corrosion detection model.

In some embodiments, the hardware corrosion detection model further includes a backbone network structure, and the neck network structure is connected to the backbone network structure; the neck network structure comprises a plurality of network branches; the network branch is constructed by adopting a first general convolution kernel and a first fusion convolution kernel; the hardware fitting rust detection model is used for extracting the features of the hardware fitting region image to obtain the feature data to be processed of the hardware fitting region image, and the method can comprise the following steps of:

and S310, performing feature extraction on the hardware fitting region image through the backbone network structure to obtain backbone feature data.

S320, inputting the main characteristic data into a neck network structure, and performing standard convolution processing on the main characteristic data through a first general convolution core of a network branch to obtain standard convolution characteristic data; and taking the standard convolution characteristic data as the characteristic data to be processed.

Specifically, feature extraction is carried out on the hardware fitting region image through a main network of the hardware fitting corrosion detection model to obtain main feature data, the main feature data are input into a neck network of the hardware fitting corrosion detection model, standard convolution processing is carried out on the main feature network through a first general convolution core, and the main feature data are used as feature data to be processed of the neck network of the hardware fitting corrosion detection model.

Fig. 4 is a schematic flow chart of a method for detecting corrosion of hardware on the top of a distribution tower according to a second embodiment of the present invention.

It should be noted that the network branch includes a first general convolution kernel, a first branch convolution kernel, and a second branch convolution kernel, which are connected in sequence; the first branch convolution kernel and the second branch convolution kernel adopt a first fusion convolution kernel.

Correspondingly, after standard convolution processing is carried out on the main characteristic data through the first general convolution kernel of the network branch to obtain standard convolution characteristic data, the method further comprises the following steps:

s410, processing the characteristic data to be processed sequentially through a first standard convolution kernel and a first depth separable convolution kernel in the first branch convolution kernel to obtain second characteristic data; and the second feature data obtained based on the first branch convolution kernel corresponds to first intermediate fusion feature data output by a first standard convolution kernel in the first branch convolution kernel.

It can be understood that the first branch convolution kernel adopts a structure of a first fusion convolution kernel, that is, includes a first standard convolution kernel and a first depth separable convolution, and the feature data to be processed is processed through the first standard convolution kernel and the first depth separable convolution to obtain first intermediate fusion feature data.

And S420, sequentially processing the first intermediate fusion characteristic data through a first standard convolution kernel and a first depth separable convolution kernel in a second branch convolution kernel of the network branch to obtain target fusion convolution characteristic data of the network branch.

The second branch convolution kernel is connected with the first branch convolution kernel by adopting a structure of the first fusion convolution kernel, so that the first intermediate fusion characteristic data is processed through the first standard convolution kernel and the first depth separable convolution in the second branch convolution kernel to obtain target fusion convolution characteristic data.

Exemplarily, fig. 5 is a schematic structural diagram of a network branch according to an embodiment of the present invention, as shown in fig. 5, the structure of the network branch may include a first general convolution kernel Conv1, a first branch convolution kernel GSConv1, and a second branch convolution kernel GSConv2, where a specific structure of the first branch convolution kernel GSConv1 and the second branch convolution kernel GSConv2 is as shown in fig. 2, and a feature image obtained by sequentially passing through the first general convolution kernel Conv1, the first branch convolution kernel GSConv1, and the second branch convolution kernel GSConv2 corresponds to target fusion convolution feature data with feature data to be processed obtained by the first general convolution kernel Conv 1.

In some embodiments, the plurality of network branches includes a first network branch; the first network branch comprises a third branch convolution kernel, a first general convolution kernel, a first branch convolution kernel and a second branch convolution kernel which are connected in sequence, and the third branch convolution kernel adopts a first fusion convolution kernel; correspondingly, before standard convolution processing is carried out on the main characteristic data through the first general convolution kernel of the network branch to obtain standard convolution characteristic data, the method further comprises the following steps: sequentially processing the main feature data through a first standard convolution kernel and a first depth separable convolution kernel in the third branch convolution kernel to obtain second feature data; and the second feature data obtained based on the third branch convolution kernel corresponds to the second intermediate fusion feature data output by the first standard convolution kernel in the third branch convolution kernel.

Correspondingly, standard convolution processing is carried out on the main characteristic data through a first general convolution kernel of the network branch to obtain standard convolution characteristic data, and the standard convolution characteristic data comprises the following steps: and standard convolution processing is carried out on the second intermediate fusion characteristic data through the first general convolution kernel of the first network branch to obtain standard convolution characteristic data.

Exemplarily, fig. 6 is a schematic structural diagram of a first network branch according to an embodiment of the present invention, and as shown in fig. 6, the first network branch includes a third branch convolution kernel GSConv3, a first general convolution kernel Conv1, a first branch convolution kernel GSConv1, and a second branch convolution kernel GSConv2, the trunk feature data output by the hardware rust detection model trunk network is processed by the third branch convolution kernel GSConv3 of the first network branch to obtain second intermediate fusion feature data, and the second intermediate fusion feature data is processed by the first general convolution kernel Conv1 of the first network branch to obtain standard convolution feature data, that is, feature data to be processed.

In some embodiments, the plurality of network branches includes a second network branch; the second network branch comprises a first general convolution kernel, a first branch convolution kernel, a second branch convolution kernel, a fourth branch convolution kernel and a fifth branch convolution kernel which are sequentially connected, wherein the fourth branch convolution kernel adopts the first general convolution kernel, and the fifth branch convolution kernel adopts the first fusion convolution kernel;

correspondingly, standard convolution processing is carried out on the main characteristic data through the first general convolution core of the network branch to obtain standard convolution characteristic data, and the method further comprises the following steps: and carrying out standard convolution processing on the main characteristic data through the first general convolution core of the second network branch to obtain standard convolution characteristic data.

Correspondingly, after performing standard convolution processing on the main feature data through the first general convolution kernel of the network branch to obtain standard convolution feature data, the method may further include: processing the standard convolution characteristic data through a first branch convolution kernel and a second branch convolution kernel in the second network branch in sequence to obtain target fusion convolution characteristic data of the second network branch; standard convolution processing is carried out on the target fusion convolution characteristic data of the second network branch through a fourth branch convolution core to obtain first intermediate standard convolution characteristic data; sequentially processing the first intermediate standard convolution characteristic data through a first standard convolution kernel and a first depth separable convolution kernel in the fifth branch convolution kernel to obtain second characteristic data; and the second characteristic data obtained based on the fifth branch convolution kernel corresponds to the third intermediate fusion characteristic data with the first characteristic data output by the first standard convolution kernel in the fifth branch convolution kernel.

Exemplarily, fig. 7 is a schematic structural diagram of a second network branch according to an embodiment of the present invention, as shown in fig. 7, the second network branch includes a first general convolution kernel Conv1, a first branch convolution kernel GSConv1, a second branch convolution kernel GSConv2, a fourth branch convolution kernel Conv4, and a fifth branch convolution kernel GSConv5, and the standard convolution feature data, that is, the feature data to be processed, is obtained by performing standard convolution processing on the main feature data through the first general convolution kernel Conv 1. Further, after the standard convolution feature data of the second network branch are processed through the first branch convolution kernel GSConv1 and the second branch convolution kernel GSConv2, target fusion convolution feature data of the second branch network are obtained, and third intermediate fusion feature data are output through a fourth branch convolution kernel GSConv4 and a fifth branch convolution kernel GSConv5 of the second network branch.

In some embodiments, the plurality of network branches further includes a third network branch; the third network branch comprises a sixth branch convolution kernel, a first general convolution kernel, a first branch convolution kernel and a second branch convolution kernel which are connected in sequence, wherein the sixth branch convolution kernel adopts a first fusion convolution kernel;

correspondingly, before standard convolution processing is carried out on the main characteristic data through the first general convolution kernel of the network branch to obtain standard convolution characteristic data, the method further comprises the following steps: sequentially processing the main feature data through a first standard convolution kernel and a first depth separable convolution kernel in the sixth branch convolution kernel to obtain second feature data; the second characteristic data obtained based on the sixth branch convolution kernel corresponds to the fourth intermediate fusion characteristic data with the first characteristic data output by the first standard convolution kernel in the sixth branch convolution kernel;

correspondingly, standard convolution processing is carried out on the main characteristic data through a first general convolution kernel of the network branch to obtain standard convolution characteristic data, and the standard convolution characteristic data comprises the following steps: and performing standard convolution processing on the fourth intermediate fusion characteristic data through the first general convolution core of the third network branch to obtain standard convolution characteristic data.

Exemplarily, fig. 8 is a schematic structural diagram of a third network branch according to an embodiment of the present invention, and as shown in fig. 8, the third network branch includes a sixth branch convolution kernel GSConv6, a first general convolution kernel Conv1, a first branch convolution kernel GSConv1, and a second branch convolution kernel GSConv2, which are connected in sequence. And processing the trunk characteristic data output by the trunk network in the hardware rust detection model through a sixth branch convolution kernel GSConv6 of the third network branch to obtain fourth intermediate fusion characteristic data, and processing the fourth intermediate fusion characteristic data through a first general convolution kernel Conv1 of the third network branch to obtain standard convolution characteristic data, namely the characteristic data to be processed.

In some embodiments, the plurality of network branches includes a first intermediate convolution kernel disposed between the second network branch and the third network branch, and a second intermediate convolution kernel disposed between the first network branch and the second network branch, and the first intermediate convolution kernel and the second intermediate convolution kernel respectively adopt a first fused convolution kernel; the second intermediate convolution kernel is used for processing the output of the first network branch to obtain fifth intermediate fusion feature data, and the fifth intermediate fusion feature data and the output of the second network branch are subjected to fusion processing to obtain sixth intermediate fusion feature data input to the first intermediate convolution kernel.

Correspondingly, before standard convolution processing is carried out on the fourth intermediate fusion characteristic data through the first general convolution kernel of the third network branch to obtain standard convolution characteristic data, the method further comprises the following steps: processing the sixth intermediate fusion characteristic data in sequence through a first general convolution kernel and a first depth separable convolution kernel in the first intermediate convolution kernel to obtain second characteristic data; the second characteristic data obtained based on the first intermediate convolution kernel corresponds to the seventh intermediate fusion characteristic data with the first characteristic data output by the first standard convolution kernel in the first intermediate convolution kernel; performing fusion processing on the seventh intermediate fusion feature data and the fourth intermediate fusion feature data to obtain eighth intermediate fusion feature data; performing standard convolution processing on the fourth intermediate fusion feature data and the seventh intermediate fusion feature data through a first general convolution core of a third network branch to obtain standard convolution feature data, wherein the standard convolution feature data comprises the following steps: and carrying out standard convolution processing on the eighth intermediate fusion characteristic data through the first general convolution kernel of the third network branch to obtain standard convolution characteristic data.

It can be understood that the first network branch and the second network branch are connected through the first intermediate convolution kernel, the second network branch is connected through the second intermediate convolution kernel, bottom-up and top-down feature fusion is performed, three output results of the neck detection network of the hardware corrosion detection model are obtained, and the three output results are respectively input into the hardware corrosion detection model prediction layer to perform prediction and regression on the target.

As an example, fig. 9 is a schematic structural diagram of a neck network of a hardware corrosion detection model according to an embodiment of the present invention.

As shown in fig. 9, in this embodiment, after being preprocessed, a power distribution tower top detection image or a hardware area image is first input to an input layer of a hardware corrosion detection model, then is subjected to multi-layer feature extraction through a backbone network, and then is passed through a first branch network, a second branch network, and a third branch network, where the first branch network is connected with the second branch network through a second intermediate convolution kernel GSConv8, and the second network branch is connected through a first intermediate convolution kernel GSConv7 and a third network branch, so that features of the image are sampled from bottom to top and from top to bottom through the first branch network, the second branch network, and the third branch network, and feature data are combined and mixed, and then transmitted to a prediction layer of the model, that is, a prediction layer of the tower top detection model or a prediction layer of the hardware corrosion detection model performs processing on a first target detection head1, a second target detection head2, and a third target detection head3 obtained by the neck network, that three feature maps with different sizes are obtained, where a head of the model includes three detection results, and a head portion of the three detection layers respectively obtains a high-to-target regression graph and a non-regression graph, and a target width ratio regression graph is obtained by using a grid.

It should be noted that the sizes of each standard convolution layer and each fusion convolution layer can be set according to the required number of channels, and the required number of output channels can be set according to actual needs, which is not limited in this invention.

As a possible implementation manner, based on the fusion result of the first feature data and the second feature data, the method may further include: and multiplying the first characteristic data and the second characteristic data, and performing information fusion on the multiplication result of the first characteristic data and the second characteristic data through a shuffle function to obtain a fusion result.

It can be understood that the deep separable convolution belongs to sparse convolution and consists of two parts of channel-by-channel convolution and point-by-point convolution, the characteristic values are separated according to the channels, and the required output is obtained through two steps of processing. For the channel-by-channel convolution, one convolution kernel of the convolution kernel is responsible for one channel, one channel is convoluted by only one convolution kernel, and the number of the convolution kernels is the same as that of the channels of the previous layer. The number of the feature maps after the channel-by-channel convolution is the same as that of the channels of the input layer, the feature maps cannot be expanded, and the feature maps are combined through point-by-point convolution to generate a new feature map.

It should be noted that, as shown in fig. 2, the structure of the first fusion convolution kernel GSConv1 may be, in this embodiment, obtain first feature data of half the required number of channels by the first standard convolution Conv, and then obtain second feature data of half the required number of channels by the first feature data by the depth separable convolution DSC. And multiplying the first characteristic data and the second characteristic data to obtain a characteristic diagram of the required output channel number.

Further, information of the first characteristic data and the second characteristic data is fused through shuffle fusion, and the generalization capability of the model is improved.

For example, for a 5 × 5 pixel image, three channels (shape is 5 × 5 × 3), a convolution layer passing through a 3 × 3 convolution kernel (assuming that the number of output channels is 4, the shape of the convolution kernel is 3 × 3 × 3 × 4, and finally 4 feature maps are output.4 filters are included in the convolution layer), each Filter includes 3 convolution kernels, and each convolution kernel has a size of 3 × 3, so that the number of parameters of the convolution layer is N _ std =4 × 3 × 3 × 3=108, and the amount of calculation is C _ std =3 × 3= (5-2) × (3 × 4) =, i.e., the number of convolution kernels W (wide) × convolution kernel H (high) ((picture W-convolution kernel W + 1) × (H-convolution kernel H + 1) × number of input channels × number of output channels) × for convolution in a depth convolution, one channel is responsible for a channel convolution kernel, one channel by one channel, and one channel is calculated by one channel after a convolution operation, the number of pixels of 5 × 5 × 3 × 3H-convolution kernel (DW) is 3H = 3H) × (DW 3H) =, so that the number of convolution kernels is calculated by one layer after a convolution kernel is calculated by 3H _ th _ convolution kernel, and a convolution kernel (DW convolution kernel), and a convolution kernel is calculated by 3H) × (DW 3 × 3 × 3H) × (W) × (DW convolution kernel, so that the number of a convolution kernel) After the integration, the feature maps of the previous step are weighted and combined in the depth direction through point-by-point convolution to generate a new feature map. The operation of point-by-point convolution is very similar to the conventional convolution operation, the size of the convolution kernel is 1 multiplied by 3, and 3 is the number of channels of the previous layer. Since the method of 1 × 1 convolution is adopted, the number of parameters involved in convolution at this step is N _ pointwise =1 × 1 × 3 × 4=12, the calculated amount is C _ pointwise =1 × 1 × 3 × 3 × 4=108, C _pointwise =1 × 1 × feature layer W × feature layer H × the number of input channels × the number of output channels (in the above example, the number of channels is 4). Thus, for the entire depth separable convolution, its total number of parameters is N _ partial = N _ depthwise27+ N _ pointwise12=39, and its total calculated amount is C _ partial = C _ depthwise243+ C _ pointwise 108=351. From the above analysis, it can be seen that the deep separable convolution reduces a large number of parameters and calculations compared to the conventional standard convolution algorithm. However, the traditional dense convolution algorithm can maximally preserve the hidden connections between each channel, while the sparse convolution algorithm completely cuts off these connections, and although the speed is increased, the feature extraction and fusion capabilities are greatly reduced. In this embodiment, the standard convolution kernel and the depth separable convolution are used in combination, that is, the convolution kernel is fused, first, the standard convolution kernel is used to obtain a feature map of half of the number of output channels required, then, the feature maps are used to obtain a feature map of half of the number of output channels required by the first depth separable convolution kernel, the two feature maps are multiplied to obtain a feature map of the number of output channels required, and then, shuffle is used to fuse information obtained by the standard convolution kernel and information obtained by the depth separable convolution kernel together, so that the amount of operation of the parameter kernel is greatly reduced, and features of the image are greatly retained.

In some embodiments, the tower top detection model and the hardware rust detection model employ EIoU Loss as a candidate box regression Loss function.

Specifically, the Loss functions of the tower mast top detection model and the hardware rust detection model are divided into three parts, including candidate frame regression Loss, foreground background classification Loss and multi-class classification Loss, and in this embodiment, the candidate frame regression Loss function CIoU is replaced by an EIoU Loss function. The expression of the EIoU Loss function is shown as follows:

wherein A and B represent the area of the real bounding box and the area of the predicted bounding box, respectively, p represents the Euclidean distance between the centers of the real bounding box and the predicted bounding box, and B ^gt Representing the center points, w and w, of the true bounding box and the predicted bounding box, respectively ^gt Respectively representing the width of the real bounding box and the width of the predicted bounding box, d representing the Euclidean distance of the diagonal vertices of the closed box, h and h ^gt Height, C, representing true bounding box and high prediction bounding box, respectively _w And C _h Representing the area of the real bounding box and the width and height, respectively, of the minimum bounding rectangle of the predicted bounding box.

It will be appreciated that the CIoU loss function is the most widely used loss function in current anchor-based detectors, but the CIoU loss function still suffers from drawbacks. Two variables (width and height) of the method can be updated only in the same direction, and are increased or reduced at the same time, so that the scene adaptability is weak. In this embodiment, the CIoU loss function is replaced with the EIoU loss function, and the EIoU loss function directly uses the width and height of the prediction bounding box as penalty terms, instead of the ratio of the width to the height, to further improve the precision of the trained model and improve the accuracy of the hardware rust part detection.

As an example, the tower pole top detection model and the hardware rust detection model adopt an exponential moving average method for parameter weight correction.

The expression of the exponential moving average method is as follows:

v _t ＝β·v _t-1 +(1-β)·θ _t

wherein, beta is epsilon [0,1 ]]When β =0, it means that the moving average is not used, and the larger the value of β, the more the value obtained by the moving average is correlated with the history value of θ, and θ _t Representing the value of the variable at time t.

It can be understood that the conventional parameter weight modification process is equivalent to updating the gradient of the whole training process cumulatively all the time, and the parameter weight using the exponential moving average method is equivalent to using the weighted average of the gradient of the training process. Because the training is unstable at the beginning, the obtained gradient is more reasonable for smaller weight, and therefore, the exponential moving average method is effective.

According to the method for detecting the hardware corrosion at the top of the power distribution tower, disclosed by the embodiment of the invention, part of standard convolution of the neck of the model is replaced by the fused convolution block, and the loss function EIoU which is more in line with the hardware corrosion scene and the exponential sliding average method for improving the robustness of the model are applied to the tower top detection model and the hardware corrosion detection model, so that the memory consumption of the model is reduced, the detection precision is improved, the real-time monitoring on the hardware corrosion at the top of the power distribution tower can be automatically and accurately carried out, the corrosion hidden danger is timely treated when the hardware corrosion at the top of the power distribution tower occurs, and the potential safety hazard is further reduced.

In order to implement the foregoing embodiment, the present invention further provides a computer-readable storage medium, on which a computer program is stored, where the computer program is configured to, when executed by a processor, implement the method for detecting rust corrosion of top hardware fittings of a power distribution tower according to the foregoing embodiment of the present invention.

Fig. 10 is a schematic structural diagram of an edge termination device according to an embodiment of the present invention.

In order to implement the foregoing embodiment, the present invention further provides an edge terminal device 1000, which includes a memory 1001, a processor 1002, and a power distribution tower top hardware rust detection program 1003 that is stored in the memory and is capable of running on the processor, where when the processor 1002 executes the power distribution tower top hardware rust detection program, the method for detecting power distribution tower top hardware rust according to the foregoing embodiment of the present invention is implemented.

It should be noted that the logic and/or steps shown in the flowcharts or otherwise described herein, such as an ordered listing of executable instructions that can be considered to implement logical functions, can be embodied in any computer-readable medium for use by or in connection with an instruction execution system, apparatus, or device, such as a computer-based system, processor-containing system, or other system that can fetch the instructions from the instruction execution system, apparatus, or device and execute the instructions. For the purposes of this description, a "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of the computer-readable medium would include the following: an electrical connection (electronic device) having one or more wires, a portable computer diskette (magnetic device), a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Further, the computer-readable medium could even be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, via for instance optical scanning of the paper or other medium, then compiled, interpreted or otherwise processed in a suitable manner if necessary, and then stored in a computer memory.

It should be understood that portions of the present invention may be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, the various steps or methods may be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, any one or combination of the following techniques, which are known in the art, may be used: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application specific integrated circuit having an appropriate combinational logic gate circuit, a Programmable Gate Array (PGA), a Field Programmable Gate Array (FPGA), or the like.

In the description of the specification, reference to the description of "one embodiment," "some embodiments," "an example," "a specific example," or "some examples" or the like means that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the invention. In this specification, the schematic representations of the terms used above do not necessarily refer to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

In the description of the present invention, it is to be understood that the terms "central," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," "circumferential," and the like are used in the orientations and positional relationships indicated in the drawings for convenience in describing the invention and to simplify the description, and are not intended to indicate or imply that the referenced devices or elements must have a particular orientation, be constructed and operated in a particular orientation, and are therefore not to be considered limiting of the invention.

Furthermore, the terms "first", "second", and the like, used in the embodiments of the present invention, are used for descriptive purposes only and are not to be construed as indicating or implying relative importance or implying any indication of the number of technical features indicated in the embodiments. Thus, a feature of an embodiment of the present invention that is defined by the terms "first," "second," etc. may explicitly or implicitly indicate that at least one of the feature is included in the embodiment. In the description of the present invention, the word "plurality" means at least two or two and more, such as two, three, four, etc., unless specifically limited otherwise in the examples.

In the present invention, unless otherwise explicitly stated or limited by the relevant description or limitation, the terms "mounted," "connected," and "fixed" in the embodiments are to be understood in a broad sense, for example, the connection may be a fixed connection, a detachable connection, or an integrated connection, and it may be understood that the connection may also be a mechanical connection, an electrical connection, etc.; of course, they may be directly connected or indirectly connected through intervening media, or they may be interconnected within one another or in an interactive relationship. Those of ordinary skill in the art will understand the specific meaning of the above terms in the present invention according to their specific implementation.

In the present invention, unless otherwise expressly stated or limited, the first feature "on" or "under" the second feature may be directly contacting the first and second features or indirectly contacting the first and second features through an intermediate. Also, a first feature "on," "over," and "above" a second feature may be directly or diagonally above the second feature, or may simply indicate that the first feature is at a higher level than the second feature. A first feature being "under," "below," and "beneath" a second feature may be directly under or obliquely under the first feature, or may simply mean that the first feature is at a lesser elevation than the second feature.

Although embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are exemplary and not to be construed as limiting the present invention, and that changes, modifications, substitutions and alterations can be made to the above embodiments by those of ordinary skill in the art within the scope of the present invention.

Claims

1. A method for detecting corrosion of hardware on the top of a power distribution tower is characterized by comprising the following steps:

extracting a hardware fitting area image from a top image of a power distribution tower;

performing feature extraction on the hardware fitting region image through a hardware fitting corrosion detection model to obtain feature data to be processed of the hardware fitting region image; the neck network structure of the hardware rust detection model is constructed by adopting a first general convolution kernel and a first fusion convolution kernel, and the first fusion convolution kernel is constructed by adopting a first standard convolution kernel and a first depth separable convolution kernel;

performing standard convolution processing on the feature data to be processed through the first standard convolution core to obtain first feature data;

performing feature fusion processing on the first feature data through the first depth separable convolution kernel to obtain second feature data;

and generating a hardware corrosion detection result based on the fusion result of the first characteristic data and the second characteristic data.

2. The method for detecting the rust corrosion of the hardware on the top of the power distribution tower according to claim 1, wherein the step of extracting the hardware area image from the top image of the power distribution tower comprises the following steps:

inputting the top image of the power distribution tower into a tower top detection model; the neck network structure of the tower pole top detection model comprises a second general convolution kernel and a second fusion convolution kernel, and the second fusion convolution kernel comprises a second standard convolution kernel and a second depth separable convolution kernel;

performing standard convolution processing on the image at the top of the power distribution tower through the second standard convolution core to obtain third characteristic data;

performing feature fusion processing on the third feature data through the second depth separable convolution kernel to obtain fourth feature data;

based on the third characteristic data and the fourth characteristic data, the tower pole top detection model outputs the hardware area image.

3. The method for detecting the rust of the hardware on the top of the power distribution tower according to claim 1, wherein the hardware rust detection model further comprises a backbone network structure, and the neck network structure is connected to the backbone network structure; the neck network structure comprises a plurality of network branches; the network branch is constructed by adopting the first general convolution kernel and the first fusion convolution kernel; the hardware fitting regional image is subjected to feature extraction through a hardware fitting corrosion detection model, and the to-be-processed feature data of the hardware fitting regional image is obtained, and the method comprises the following steps:

performing feature extraction on the hardware fitting region image through the trunk network structure to obtain trunk feature data;

inputting the main characteristic data into the neck network structure, and performing standard convolution processing on the main characteristic data through the first general convolution kernel of the network branch to obtain standard convolution characteristic data; and the standard convolution characteristic data is used as the characteristic data to be processed.

4. The method for detecting the rust corrosion of the hardware on the top of the power distribution tower according to claim 3, wherein the network branches comprise a first general convolution kernel, a first branch convolution kernel and a second branch convolution kernel which are sequentially connected; the first branch convolution kernel and the second branch convolution kernel adopt a first fusion convolution kernel; correspondingly, after the standard convolution processing is performed on the main feature data through the first general convolution kernel of the network branch to obtain standard convolution feature data, the method further includes:

sequentially processing the feature data to be processed through a first standard convolution kernel and a first depth separable convolution kernel in the first branch convolution kernel to obtain second feature data; second feature data obtained based on the first branch convolution kernel corresponds to first intermediate fusion feature data output by a first standard convolution kernel in the first branch convolution kernel;

and sequentially processing the first intermediate fusion characteristic data through a first standard convolution kernel and a first depth separable convolution kernel in a second branch convolution kernel of the network branch to obtain target fusion convolution characteristic data of the network branch.

5. The method of claim 4, wherein the plurality of network branches includes a first network branch; the first network branch comprises a third branch convolution kernel, a first general convolution kernel, a first branch convolution kernel and a second branch convolution kernel which are connected in sequence, and the third branch convolution kernel adopts a first fusion convolution kernel; correspondingly, before the standard convolution processing is performed on the main feature data through the first general convolution kernel of the network branch to obtain standard convolution feature data, the method further includes:

sequentially processing the trunk characteristic data through a first standard convolution kernel and a first depth separable convolution kernel in the third branch convolution kernel to obtain second characteristic data; second feature data obtained based on the third branch convolution kernel corresponds to first feature data output by a first standard convolution kernel in the third branch convolution kernel;

correspondingly, the standard convolution processing is performed on the main feature data through the first general convolution kernel of the network branch to obtain standard convolution feature data, and the standard convolution feature data includes:

and performing standard convolution processing on the second intermediate fusion characteristic data through a first general convolution core of the first network branch to obtain the standard convolution characteristic data.

6. The method of claim 5, wherein the plurality of network branches includes a second network branch; the second network branch comprises a first general convolution kernel, a first branch convolution kernel, a second branch convolution kernel, a fourth branch convolution kernel and a fifth branch convolution kernel which are connected in sequence, wherein the fourth branch convolution kernel adopts a first general convolution kernel, and the fifth branch convolution kernel adopts a first fusion convolution kernel; correspondingly, the standard convolution processing is performed on the main feature data through the first general convolution kernel of the network branch to obtain standard convolution feature data, and the method further includes:

performing standard convolution processing on the main characteristic data through a first general convolution core of the second network branch to obtain standard convolution characteristic data;

correspondingly, after the standard convolution processing is performed on the main feature data through the first general convolution kernel of the network branch to obtain standard convolution feature data, the method includes:

processing the standard convolution characteristic data through a first branch convolution kernel and a second branch convolution kernel in the second network branch in sequence to obtain target fusion convolution characteristic data of the second network branch;

performing standard convolution processing on the target fusion convolution characteristic data of the second network branch through the fourth branch convolution core to obtain first intermediate standard convolution characteristic data;

sequentially processing the first intermediate standard convolution characteristic data through a first standard convolution kernel and a first depth separable convolution kernel in the fifth branch convolution kernel to obtain second characteristic data; and the second characteristic data obtained based on the fifth branch convolution kernel corresponds to the third intermediate fusion characteristic data output by the first standard convolution kernel in the fifth branch convolution kernel.

7. The method of claim 6, wherein the plurality of network branches includes a third network branch; the third network branch comprises a sixth branch convolution kernel, a first general convolution kernel, a first branch convolution kernel and a second branch convolution kernel which are connected in sequence, and the sixth branch convolution kernel adopts a first fusion convolution kernel; correspondingly, before the standard convolution processing is performed on the main feature data through the first general convolution kernel of the network branch to obtain standard convolution feature data, the method further includes:

sequentially processing the trunk characteristic data through a first standard convolution kernel and a first depth separable convolution kernel in the sixth branch convolution kernel to obtain second characteristic data; the second feature data obtained based on the sixth branch convolution kernel corresponds to fourth intermediate fusion feature data output by the first standard convolution kernel in the sixth branch convolution kernel;

and performing standard convolution processing on the fourth intermediate fusion characteristic data through the first general convolution kernel of the third network branch to obtain the standard convolution characteristic data.

8. The method for detecting the rust corrosion of the power distribution tower top hardware fitting of claim 7, wherein the plurality of network branches comprise a first intermediate convolution kernel arranged between a second network branch and a third network branch, and a second intermediate convolution kernel arranged between the first network branch and the second network branch, and the first intermediate convolution kernel and the second intermediate convolution kernel respectively adopt a first fusion convolution kernel; the second intermediate convolution kernel is configured to process an output of the first network branch to obtain fifth intermediate fusion feature data, and the fifth intermediate fusion feature data is fused with the output of the second network branch to obtain sixth intermediate fusion feature data input to the first intermediate convolution kernel; correspondingly, before the standard convolution processing is performed on the fourth intermediate fusion feature data through the first general convolution kernel of the third network branch to obtain the standard convolution feature data, the method further includes:

processing the sixth intermediate fusion feature data sequentially through a first general convolution kernel and a first depth separable convolution kernel in the first intermediate convolution kernel to obtain second feature data; the second feature data obtained based on the first intermediate convolution kernel corresponds to seventh intermediate fusion feature data with the first feature data output by the first standard convolution kernel in the first intermediate convolution kernel;

performing fusion processing on the seventh intermediate fusion feature data and the fourth intermediate fusion feature data to obtain eighth intermediate fusion feature data;

performing standard convolution processing on the fourth intermediate fusion feature data and the seventh intermediate fusion feature data through the first general convolution kernel of the third network branch to obtain the standard convolution feature data, including:

and performing standard convolution processing on the eighth intermediate fusion characteristic data through the first general convolution kernel of the third network branch to obtain the standard convolution characteristic data.

9. The method for detecting the rusting of the hardware on the top of the power distribution tower according to claim 1, wherein based on the fusion result of the first characteristic data and the second characteristic data, the method comprises the following steps:

and multiplying the first characteristic data by the second characteristic data, and performing information fusion on the multiplication result of the first characteristic data and the second characteristic data through a shuffle function to obtain the fusion result.

10. The method for detecting corrosion of hardware on the top of a power distribution tower according to any one of claims 2 to 9, wherein the tower top detection model and the hardware corrosion detection model adopt EIoU Loss as a candidate frame regression Loss function.

11. The method for detecting the rust of the hardware on the top of the power distribution tower according to any one of claims 2 to 9, wherein the tower top detection model and the hardware rust detection model are subjected to parameter weight correction by an exponential sliding average method.

12. A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method for detecting rust corrosion on the top of power distribution towers according to any one of claims 1-11.

13. An edge terminal device, characterized by comprising a memory, a processor and a power distribution tower top hardware corrosion detection program stored on the memory and capable of running on the processor, wherein when the processor executes the power distribution tower top hardware corrosion detection program, the method for detecting the power distribution tower top hardware corrosion according to any one of claims 1-11 is realized.