A power transmission channel hidden danger detection method and system

Abstract

This invention belongs to the field of power transmission channel hazard detection technology, specifically disclosing a method and system for power transmission channel hazard detection. The method includes: inputting acquired power transmission channel images into a trained hazard detection model to obtain hazard detection results; the hazard detection model includes a backbone network, a feature fusion network, and a detection head connected in sequence; a multi-scale feature extraction module is provided in the backbone network and / or the feature fusion network, the multi-scale feature extraction module includes n feature extraction branches, n≥2; the output of the i-th feature extraction branch is added to the input of the (i+1)-th feature extraction branch, and then fused through multi-scale frequency domain features to obtain the output of the (i+1)-th feature extraction branch; i=1,2,…,n; the outputs of all feature extraction branches are concatenated to obtain the output of the multi-scale feature extraction module. This invention utilizes different frequency features to better identify interference targets and achieve accurate detection of hazard targets.

Description

Technical Field

[0001] This invention relates to the field of power transmission channel hazard detection technology, and in particular to a method and system for power transmission channel hazard detection. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] As the coverage of the power system continues to expand, the differences and complexity of the surrounding environment are also increasing. In the case of transmission line corridors, they will inevitably be affected by many external hazards (such as construction machinery, smoke, wildfires and foreign objects on the conductors), which brings greater challenges to the stable operation of the power system.

[0004] In existing technologies, deep learning technology has been widely used for the detection of potential hazards. However, there are many highly similar and interfering targets in nature that are highly similar to external hazards, such as fog and smoke in the mountains, some lights at night and wildfires, and cement poles and cranes on the roadside. In the process of optimizing the balance between false alarms and missed detections to improve alarm quality, previous deep learning target detection models are prone to problems such as missed detections and false detections for certain difficult-to-distinguish samples due to factors such as weather, scene changes, the shape of the potential hazard, and differences in the imaging quality of monitoring equipment. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes a method and system for detecting potential hazards in power transmission channels. By constructing a multi-scale frequency domain feature fusion extraction network module, the system comprehensively extracts multi-scale and multi-frequency fusion features of the input features. This allows the system to learn the differences in visual features between the input and interference targets, thereby improving the accuracy of hazard detection.

[0006] In some implementations, the following technical solutions are adopted:

[0007] A method for detecting hidden dangers in power transmission channels includes:

[0008] The acquired images of the power transmission channel are input into the trained hazard detection model to obtain the hazard detection results of the power transmission channel;

[0009] The hazard detection model includes a backbone network, a feature fusion network, and a detection head connected in sequence. A multi-scale feature extraction module is provided in the backbone network and / or the feature fusion network. The multi-scale feature extraction module includes n feature extraction branches, where n≥2. The output of the i-th feature extraction branch is added to the input of the (i+1)-th feature extraction branch, and then fused with multi-scale frequency domain features to obtain the output of the (i+1)-th feature extraction branch; i = 1, 2, ..., n. The outputs of all feature extraction branches are concatenated to obtain the output of the multi-scale feature extraction module.

[0010] Traditional methods typically involve adding a sufficient number of positive and negative samples to the training dataset, allowing the model to explore the boundary between potential hazards and interfering targets during training. However, due to the limited source of features, many indistinguishable samples still exist near the boundary, leading to missed or false detections of potential hazards in power transmission channels. This invention introduces a multi-scale frequency domain feature fusion extraction module into the backbone network and / or feature fusion network. This module can extract rich multi-scale hazard features and extract and fuse multi-level frequency domain features, resulting in multi-scale, multi-resolution, and multi-frequency hazard feature maps with a greater number of features, and achieving global perception capabilities. By leveraging different frequency features, it can better identify interfering targets, thereby achieving accurate detection of potential hazards and solving the problem of missed or false detections caused by weather, scene changes, the morphology of potential hazards, and differences in the imaging quality of monitoring equipment.

[0011] As an optional solution, the backbone network is used to extract multi-scale frequency domain fusion features at different levels of the power transmission channel image, the feature fusion network is used to further fuse the multi-scale frequency domain fusion features at different levels, and the detection head is used to output the detection results of potential hazards in the power transmission channel.

[0012] As an optional solution, the specific processing steps of the multi-scale feature extraction module include:

[0013] The input features are expanded by convolution, dividing them into n independent features.

[0014] n independent features are input into n feature extraction branches respectively;

[0015] The first feature extraction branch performs an identity mapping operation on the independent input features to obtain the output of the first feature extraction branch;

[0016] The output of the i-th feature extraction branch is added to the independent features input to the (i+1)-th feature extraction branch, and then the output of the (i+1)-th feature extraction branch is obtained by fusing multi-scale frequency domain features.

[0017] The outputs of all feature extraction branches are concatenated to form the output of the multi-scale feature extraction module.

[0018] As an optional approach, the fusion process of the multi-scale frequency domain features is as follows:

[0019] The basic features are obtained by performing a convolution operation on the input features;

[0020] The input features are subjected to at least two levels of wavelet transform to obtain the frequency domain features;

[0021] The basic features and frequency domain features are fused to obtain the final output.

[0022] As an optional approach, the fusion of the multi-scale frequency domain features is carried out as follows:

[0023] For input feature X, the first low-frequency feature of input feature X is obtained by the first wavelet transform. and the first high-frequency feature By analyzing the first low-frequency feature Perform a second wavelet transform to obtain the second low-frequency feature. Second high frequency characteristics Regarding the second low-frequency feature Second high frequency characteristics Perform the first convolution operation and inverse wavelet transform to obtain the feature Z extracted by the first inverse wavelet transform. 1 ;

[0024] Feature Z extracted from the first wavelet inverse transform 1 With the first low-frequency feature and the first high-frequency feature After addition, a second convolution operation and inverse wavelet transform are performed to obtain the feature Z extracted by the second inverse wavelet transform. 2 ;

[0025] Perform a third convolution operation on the input feature X to obtain the third convolution domain feature Y. 3 ;

[0026] The feature Z extracted by the inverse wavelet transform of the second wavelet 2 and the third convolutional domain feature Y 3 After addition, an inverse wavelet transform is performed to obtain the final output feature Z. 3 .

[0027] As an optional solution, the backbone network is specifically: on the basis of a conventional backbone network, the residual network module is replaced with a multi-scale frequency domain feature fusion and extraction module;

[0028] or,

[0029] The feature fusion network is specifically designed to replace the residual network module with a multi-scale frequency domain feature fusion extraction module, based on a conventional feature fusion network.

[0030] In other embodiments, the following technical solutions are adopted:

[0031] A power transmission channel hazard detection system, comprising:

[0032] The hazard detection module is used to input the acquired images of the power transmission channel into the trained hazard detection model to obtain the hazard detection results of the power transmission channel;

[0033] The hazard detection model includes a backbone network, a feature fusion network, and a detection head connected in sequence. A multi-scale feature extraction module is provided in the backbone network and / or the feature fusion network. The multi-scale feature extraction module includes n feature extraction branches, where n≥2. The output of the i-th feature extraction branch is added to the input of the (i+1)-th feature extraction branch, and then fused with multi-scale frequency domain features to obtain the output of the (i+1)-th feature extraction branch; i = 1, 2, ..., n. The outputs of all feature extraction branches are concatenated to obtain the output of the multi-scale feature extraction module.

[0034] In other embodiments, the following technical solutions are adopted:

[0035] A terminal device includes a processor and a memory, wherein the processor is used to implement instructions; and the memory is used to store multiple instructions, which are adapted to be loaded and executed by the processor to perform the above-described method for detecting potential hazards in power transmission channels.

[0036] In other embodiments, the following technical solutions are adopted:

[0037] A computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the above-described method for detecting potential hazards in power transmission channels.

[0038] In other embodiments, the following technical solutions are adopted:

[0039] A computer program product includes a computer program / instructions that, when executed by a processor, implement the aforementioned method for detecting potential hazards in power transmission channels.

[0040] Compared with the prior art, the beneficial effects of the present invention are:

[0041] (1) The present invention constructs a multi-scale feature extraction module, which extracts multi-scale hidden danger features from the input features. This can optimize the problem of different scales in the target detection task. Through multi-scale fusion, it can better handle scale changes and improve the model's adaptability at different scales. It can enhance the features of the hidden danger target such as edge, texture and shape, and improve the ability to locate the hidden danger target. In addition, features at different scales will provide different contextual information, which can further enhance the model's recognition ability and improve the performance of detection and recognition.

[0042] Meanwhile, in the process of multi-scale hazard feature extraction, the fusion of multi-scale frequency domain features is introduced. Convolution operation and multi-level wavelet transform are performed on the input features respectively. The basic features extracted by the convolution operation and the multi-level frequency domain features extracted by the multi-level wavelet transform are fused to obtain hazard feature maps with more feature quantity, multi-scale, multi-resolution and different frequencies. With the help of different frequency features, the identification of interference targets can be better realized, the accurate detection of hazard targets can be achieved, and the problems of missed detection or false detection caused by weather, scene changes and the shape of hazard targets can be solved.

[0043] (2) In the process of multi-level wavelet transform, the present invention further extracts low-frequency features by combining two-level wavelet transform with convolution operation, which can enhance the low-frequency texture information in the input image, help to accurately extract low-frequency features with slowly changing gray levels, enhance image quality, and solve the problem of missed detection or false detection caused by poor imaging quality of monitoring equipment.

[0044] (3) By adding 1x1 convolutional units before and after frequency domain feature fusion extraction, the present invention can comprehensively balance the amount of computation and feature redundancy by increasing the number of input and output channels of the convolutional kernels, so as to obtain richer and more differentiated features in the channel dimension without increasing the amount of computation.

[0045] Other features 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. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the hazard detection model structure in an embodiment of the present invention;

[0047] Figure 2 This is a schematic diagram of the multi-scale frequency domain feature fusion and extraction module in an embodiment of the present invention;

[0048] Figure 3 This is a schematic diagram of the multi-scale frequency domain feature fusion extraction process in an embodiment of the present invention. Detailed Implementation

[0049] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0050] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0051] Example 1

[0052] In one or more embodiments, a method for detecting hidden dangers in power transmission channels is disclosed, which specifically includes the following process: acquiring images of power transmission channels; inputting the acquired images into a trained hidden danger detection model; and outputting the hidden danger detection results of the power transmission channels.

[0053] In this embodiment, images of the power transmission line are acquired using visual monitoring equipment installed on the transmission line. The acquired images are then transmitted to a terminal device configured with the hazard detection model of this embodiment to detect the types of external damage hazards to the power transmission line. These external damage hazards include, but are not limited to, construction machinery such as cranes and cement pump trucks, smoke, wildfires, and foreign objects in the conductors.

[0054] In this embodiment, combined with Figure 1 The hazard detection model mainly consists of a backbone network, a feature fusion network, and a detection head connected in sequence. The backbone network is used to extract multi-scale frequency domain fusion features at different levels of the transmission channel image. The feature fusion network is used to further fuse the multi-scale frequency domain fusion features at different levels. The detection head is used to output the hazard detection results of the transmission channel.

[0055] In this embodiment, the backbone network and / or feature fusion network include a multi-scale feature extraction module, combined with... Figure 2 The multi-scale feature extraction module includes a 1×1 convolutional unit. Channel expansion using this 1×1 convolutional unit divides the input potential hazard features into n independent features. These n independent features are then input to n feature extraction branches. The first feature extraction branch performs an identity mapping operation on the input independent features, yielding its output. The output of the ith feature extraction branch is added to the independent features input to the (i+1)th feature extraction branch, and then passed through the multi-scale frequency domain feature fusion unit (WTCF) to obtain the output of the (i+1)th feature extraction branch. The values ​​i = 1, 2, ..., n. The outputs of all feature extraction branches are concatenated to form the output of the multi-scale feature extraction module.

[0056] It should be noted that the multi-scale frequency domain feature fusion unit WTCF (Wavelet TransformConvolution Fusion) in this embodiment is used to fuse the frequency domain features extracted by wavelet transform with the basic features extracted by convolution operation.

[0057] As a specific implementation method, combined with Figure 2 The specific processing procedure of the multi-scale feature extraction module for the input features in this embodiment is as follows:

[0058] Input features X∈R H×W×C First, a 1×1 convolution operation is used to expand the channels, so that the output feature Y∈R is... H×W×C Divide the features Y into n independent groups on average i ;

[0059] The first set of independent features Y1 is used as the first feature Y1 in multi-scale feature extraction through the Identity operation. extract ;

[0060] Feature Y1 extract After being added to the second set of independent features Y2, the multi-scale frequency domain features are fused and extracted through the Multi-Scale Frequency Domain Feature Fusion Unit (WTCF), outputting the second feature extracted from the multi-scale features.

[0061]

[0062] Second feature After being added to the third independent feature Y3, the multi-scale frequency domain features are fused and extracted through the Multi-Scale Frequency Domain Feature Fusion Unit (WTCF), outputting the third feature extracted from the multi-scale features.

[0063] Following this pattern, we can finally obtain n features from the multi-scale feature extraction. These n features are then concatenated to obtain the output of the multi-scale frequency domain feature fusion extraction module.

[0064] Z = Concat(Y) i extract (i = 1, ..., n).

[0065] Through the above process, multi-scale feature extraction can be performed on the input hazard features. Each branch can contain the scale features extracted by the previous branch, making the extracted features richer and enhancing the edge, texture and shape features of the hazard target, thereby improving the ability to locate the hazard target. It can optimize the problem of different scales in the target detection task. Through multi-scale feature fusion, scale changes can be better handled and the model's adaptability at different scales can be improved.

[0066] Meanwhile, in this embodiment, when extracting features at different scales, a multi-scale frequency domain feature fusion unit (WTCF) is introduced. Convolution and multi-level wavelet transform are performed on the input features respectively. The basic features extracted by the convolution operation and the multi-level frequency domain features extracted by the multi-level wavelet transform are fused to obtain a multi-scale, multi-resolution, and multi-frequency feature map with more features.

[0067] As a specific example, combined with Figure 3 The multi-scale frequency domain feature fusion unit WTCF in this embodiment consists of a first wavelet transform layer W1 and a second wavelet transform layer W2, a first convolutional layer C1, a second convolutional layer C2 and a third convolutional layer C3, as well as a first wavelet inverse transform layer I1, a first wavelet inverse transform layer I2 and a third wavelet inverse transform layer I3, which can realize the extraction of frequency domain feature fusion information of multi-scale low-frequency enhancement.

[0068] The data processing procedure for the fusion unit of multi-scale frequency domain features is as follows:

[0069] (1) Multi-scale frequency domain feature extraction:

[0070] (1-1) For the input feature X', the first low-frequency feature of the input feature is obtained through the first wavelet transform layer W1. and the first high-frequency feature Right now:

[0071]

[0072] Where WT represents wavelet transform.

[0073] (1-2) Regarding the first low-frequency characteristic The second low-frequency feature is obtained through the second wavelet transform layer W2. Second high frequency characteristics Right now:

[0074]

[0075] This embodiment extracts high-frequency and low-frequency features of the input image through a first wavelet transform layer W1, and then further extracts high-frequency and low-frequency features of the low-frequency features through a second wavelet transform layer W2. The features extracted by the two wavelet transforms are then fused, which can enhance the low-frequency texture information in the input image, help to accurately extract low-frequency features with slowly changing gray levels, improve image quality, and solve the problem of missed or false detections caused by poor imaging quality of monitoring equipment.

[0076] (1-3) The second low-frequency feature output by the second wavelet transform layer W2 through the first convolutional layer C1 Second high frequency characteristics Perform the first convolution operation to obtain the low-frequency features of the first convolution. and the high-frequency features of the first convolution Right now:

[0077]

[0078] Where Conv represents the convolution operation, and W represents the convolution kernel of the convolution operation Conv.

[0079] (1-4) The first convolutional low-frequency features output from the first convolutional layer C1 are processed by the first wavelet inverse transform layer I1. and the high-frequency features of the first convolution Perform the first wavelet inverse transform to obtain the feature Z. 1 ;Right now:

[0080]

[0081] Where IWT represents the inverse wavelet transform.

[0082] (1-5) Feature Z extracted from the first wavelet inverse transform layer I1 1 Compared with the first low-frequency features extracted through the first wavelet transform layer W1 and the first high-frequency feature After addition, a second convolution operation is performed through the second convolutional layer C2 to obtain the low-frequency features of the second convolution. Second convolution high frequency features Right now:

[0083]

[0084] (1-6) Perform a second wavelet inverse transform on the output of the second convolutional layer C2 through the second wavelet inverse transform layer I2 to obtain the feature Z extracted by the second wavelet inverse transform layer I2. 2 ,Right now:

[0085]

[0086] (2) Basic feature extraction: Perform a third convolution operation on the input feature X to obtain the third convolution domain feature Y. 3 ;

[0087] (3) Feature fusion: The features Z extracted from the second wavelet inverse transform layer I2 are fused together. 2 and the third convolutional domain feature Y 3 After addition, an inverse wavelet transform is performed through the third inverse wavelet transform layer I3 to obtain the final output feature Z. 3 ;Right now:

[0088] Z 3 =IWT(Y 3 +Z2 ).

[0089] This embodiment utilizes different frequency characteristics to better identify interfering targets, achieve accurate detection of potential hazards, and solve the problem of missed or false detections caused by weather, scene changes, and the form in which potential hazards are presented.

[0090] Considering computational cost, this embodiment selects, but is not limited to, the Haar wavelet transform, which has lower computational cost; the wavelet filter applied to the image is shown below:

[0091]

[0092] Among them, f LL For a low-pass filter, f HL ,f LH ,f HH It is a high-pass filter.

[0093] The high- and low-frequency features extracted from the input features using wavelet transform are as follows:

[0094] [X LL ,X LH ,X HL ,X HH ] = Conv([f LL ,f LH ,f HL ,f HH ],X');

[0095] Among them, X LL X represents the low-frequency characteristics of slowly changing gray levels. LH ,X HL ,X HH These are high-frequency features characterized by abrupt changes in grayscale along the horizontal, vertical, and diagonal directions, respectively. These three elements together constitute the overall high-frequency feature X. H .

[0096] Each wavelet transform reduces the resolution of the feature map by half, thus enabling multi-resolution feature extraction of the input features.

[0097] As an optional example, this embodiment connects a 1x1 convolutional unit before the input and after the output of the Multi-Scale Frequency Domain Feature Fusion Unit (WTCF). By adjusting the number of convolutional kernels in the convolutional unit, the number of input and output channels of the WTCF can be adjusted, allowing the WTCF to execute only on a certain proportion (r) of the input channels while keeping the remaining channels unchanged. This comprehensively considers computational cost and feature redundancy, obtaining richer and more differentiated features in the channel dimension without increasing computational load. Specifically, when r = 1, it is executed on all channels; as a specific example, this embodiment chooses r = 0.5; of course, it can also be adjusted according to actual needs.

[0098] The foregoing has provided a detailed description of the multi-scale feature extraction module in this embodiment. As a specific implementation, in the hazard detection model of this embodiment, the backbone network is based on an existing backbone network structure, with the residual network module replaced by the multi-scale feature extraction module of this embodiment. For example, the backbone network can be a ResNext101 model, with the residual network module replaced by the multi-scale feature extraction module to form the backbone network of this embodiment.

[0099] Similarly, the feature fusion network is based on the existing feature fusion network structure, replacing the residual network module with the multi-scale feature extraction module of this embodiment. For example, the feature fusion network can use PAFPN (PathAggregation Feature Pyramid Network), which can fully integrate low-level location information and high-level semantic information; by replacing the residual network module with the multi-scale feature extraction module, the feature fusion network of this embodiment is formed to capture richer multi-scale frequency domain fusion features, and to achieve feature fusion of multi-scale frequency domain features extracted from different levels of the backbone network.

[0100] It should be noted that the training of the hazard detection model can be completed offline. The dataset required for model training consists of historical images of power transmission lines, labeled with hazard types. After training and evaluation, the hazard detection model is trained to achieve metrics such as accuracy (evaluating false positive rate) and recall (evaluating false negative rate). As an optional example, this embodiment uses WIoUv3 as the regression loss and employs NMS to remove redundant bounding boxes, effectively reducing the large harmful gradients generated by low-quality examples in the training data, ultimately resulting in a well-trained hazard detection model. Inputting the acquired power transmission line images into the trained hazard detection model will output the hazard detection results for the power transmission line.

[0101] This embodiment introduces a multi-scale feature extraction module into the backbone network and the feature fusion network. At the same time, it introduces a multi-scale frequency domain feature fusion unit into the feature extraction branches at different scales. This can enhance the extraction of low-frequency texture information in the input image and obtain multi-scale, multi-resolution, and different frequency hazard feature maps with more feature quantity. This enables the hazard detection model to have a strong ability to distinguish between hazard targets and similar interference targets, and achieve accurate detection of hazard in power transmission channels.

[0102] Example 2

[0103] In one or more embodiments, a power transmission channel hazard detection system is disclosed, specifically including:

[0104] The data acquisition module is used to acquire images of the power transmission channel;

[0105] The hazard detection module is used to input the acquired images of the power transmission channel into the trained hazard detection model to obtain the hazard detection results of the power transmission channel;

[0106] The hazard detection model comprises a backbone network, a feature fusion network, and a detection head connected in sequence. A multi-scale feature extraction module is included in the backbone network and / or the feature fusion network. This multi-scale feature extraction module includes n feature extraction branches, where n ≥ 2. The output of the i-th feature extraction branch is added to the input of the (i+1)-th feature extraction branch, and then fused with multi-scale frequency domain features to obtain the output of the (i+1)-th feature extraction branch; i = 1, 2, ..., n. The outputs of all feature extraction branches are concatenated to obtain the output of the multi-scale feature extraction module.

[0107] It should be noted that the specific implementation methods of the above modules are the same as those in Example 1, and will not be described in detail again.

[0108] Example 3

[0109] In one or more embodiments, a terminal device is disclosed, which includes a processor and a memory, wherein the processor is used to implement instructions; the memory is used to store multiple instructions adapted to be loaded by the processor and executed by the power transmission channel hazard detection method in Embodiment 1.

[0110] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0111] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0112] In the implementation process, each step of the above method can be completed by the integrated logic circuits in the processor hardware or by software instructions.

[0113] Example 4

[0114] In one or more embodiments, a computer-readable storage medium is disclosed, which stores a plurality of instructions adapted to be loaded by a processor of a terminal device and executed by the transmission channel hazard detection method of Embodiment 1.

[0115] Example 5

[0116] In one or more embodiments, a computer program product is disclosed, including a computer program / instruction that, when executed by a processor, implements the power transmission channel hazard detection method in Embodiment 1.

[0117] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for detecting hidden dangers in power transmission channels, characterized in that, include: The acquired images of the power transmission channel are input into the trained hazard detection model to obtain the hazard detection results of the power transmission channel; The hazard detection model includes a backbone network, a feature fusion network, and a detection head connected in sequence. A multi-scale feature extraction module is provided in the backbone network and / or the feature fusion network. The multi-scale feature extraction module includes n feature extraction branches, where n ≥ 2. The output of the i-th feature extraction branch is added to the input of the (i+1)-th feature extraction branch, and then fused with multi-scale frequency domain features to obtain the output of the (i+1)-th feature extraction branch; i = 1, 2, ..., n. The outputs of all feature extraction branches are concatenated to form the output of the multi-scale feature extraction module.

2. The method for detecting hidden dangers in power transmission channels as described in claim 1, characterized in that, The backbone network is used to extract multi-scale frequency domain fusion features at different levels of the power transmission channel image, and feature fusion... The network is used to further fuse multi-scale frequency domain fusion features at different levels, and the detection head is used to output the detection results of potential hazards in the power transmission channel.

3. The method for detecting hidden dangers in power transmission channels as described in claim 1, characterized in that, The specific processing steps of the multi-scale feature extraction module include: The input features are expanded by convolution, dividing them into n independent features. n independent features are input into n feature extraction branches respectively; The first feature extraction branch performs an identity mapping operation on the independent input features to obtain the output of the first feature extraction branch; The output of the i-th feature extraction branch is added to the independent features input to the (i+1)-th feature extraction branch, and then the output of the (i+1)-th feature extraction branch is obtained by fusing multi-scale frequency domain features. The outputs of all feature extraction branches are concatenated to form the output of the multi-scale feature extraction module.

4. The method for detecting hidden dangers in power transmission channels as described in claim 1, characterized in that, The fusion process of the multi-scale frequency domain features is as follows: The basic features are obtained by performing a convolution operation on the input features; The input features are subjected to at least two levels of wavelet transform to obtain the frequency domain features; The basic features and frequency domain features are fused to obtain the final output.

5. The method for detecting hidden dangers in power transmission channels as described in claim 4, characterized in that, The fusion of the multi-scale frequency domain features is carried out as follows: For input feature X, the first low-frequency feature of input feature X is obtained by the first wavelet transform. and the first high-frequency feature By analyzing the first low-frequency feature Perform a second wavelet transform to obtain the second low-frequency feature. Second high frequency characteristics Regarding the second low-frequency feature Second high frequency characteristics Perform the first convolution operation and inverse wavelet transform to obtain the feature Z extracted by the first inverse wavelet transform. 1 ; Feature Z extracted from the first wavelet inverse transform 1 With the first low-frequency feature and the first high-frequency feature After addition, a second convolution operation and inverse wavelet transform are performed to obtain the feature Z extracted by the second inverse wavelet transform. 2 ; Perform a third convolution operation on the input feature X to obtain the third convolution domain feature Y. 3 ; The feature Z extracted by the inverse wavelet transform of the second wavelet 2 and the third convolutional domain feature Y 3 After addition, an inverse wavelet transform is performed to obtain the final output feature Z. 3 .

6. The method for detecting hidden dangers in power transmission channels as described in claim 1, characterized in that, The backbone network is specifically defined as follows: based on the conventional backbone network, the residual network module is replaced with a multi-scale frequency domain feature fusion and extraction module; or, The feature fusion network is specifically designed to replace the residual network module with a multi-scale frequency domain feature fusion extraction module, based on a conventional feature fusion network.

7. A power transmission channel hidden danger detection system, characterized in that, include: The hazard detection module is used to input the acquired images of the power transmission channel into the trained hazard detection model to obtain the hazard detection results of the power transmission channel; The hazard detection model includes a backbone network, a feature fusion network, and a detection head connected in sequence. A multi-scale feature extraction module is provided in the backbone network and / or the feature fusion network. The multi-scale feature extraction module includes n feature extraction branches, where n ≥ 2. The output of the i-th feature extraction branch is added to the input of the (i+1)-th feature extraction branch, and then fused with multi-scale frequency domain features to obtain the output of the (i+1)-th feature extraction branch; i = 1, 2, ..., n. The outputs of all feature extraction branches are concatenated to form the output of the multi-scale feature extraction module.

8. A terminal device comprising a processor and a memory, the processor for implementing instructions; the memory for storing multiple instructions, characterized in that, The instructions are adapted to be loaded by a processor and executed as the power transmission channel hazard detection method according to any one of claims 1-6.

9. A computer-readable storage medium storing a plurality of instructions, characterized in that, The instructions are adapted to be loaded by the processor of the terminal device and executed as described in any one of claims 1-6, the method for detecting hidden dangers in power transmission channels.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the method for detecting potential hazards in power transmission channels as described in any one of claims 1-6.

Images