A method and system for detecting electric power equipment based on dynamic fusion of multiple feature channels

Abstract

The present invention discloses a method and system for detecting electric equipment based on dynamic fusion of multiple feature channels, comprising: obtaining visible light images and infrared images of electric equipment; extracting features of the visible light image according to a preset first feature extraction backbone network to obtain a first feature map corresponding to the visible light image, and extracting features of the infrared image according to a preset second feature extraction backbone network to obtain a second feature map corresponding to the infrared image; connecting the first feature map and the second feature map to obtain an initial feature map, and using a channel attention mechanism to perform feature channel screening on the initial feature map to obtain a target feature map; inputting the first feature map, the second feature map and the target feature map into a preset detection network, and the detection network outputs the type result of the electric equipment. The accuracy of electric equipment identification is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power equipment detection, and in particular relates to a power equipment detection method and system based on dynamic fusion of multiple feature channels. Background Art

[0002] In the monitoring of power equipment status, infrared images and visible light images have good complementarity. However, due to the different imaging mechanisms of infrared images and visible light images, the feature selection mode is not fixed when the two are fused and complemented. This is mainly because the imaging state of the target in the infrared image is not only related to its own surface temperature, but also affected by the external environment, resulting in uncertainty in the grayscale change and texture change of the target in the infrared image. In addition, the target feature information that can be presented by the visible light image in a weak or no light environment is limited, and at this time it is more necessary to rely on the target feature information of the infrared image.

[0003] The existing multi-source image target detection based on deep learning mainly includes RCNN series models, YOLO series models, Transformer series models, etc. There are two modes for deep learning models to process multi-source image inputs. One is to extract the features of single-source images separately, and then fuse the features extracted from each source to obtain the mapping relationship between features and targets; the other is to directly superimpose multi-source images, and then perform feature extraction to obtain the mapping relationship between features and targets. Since the states of power equipment in infrared images and visible light images are quite different, the former mode is more conducive to fully exploring the features of power equipment in infrared images.

[0004] The feature extraction model in the YOLOv8 model has strong feature extraction capabilities and computational efficiency, which is conducive to extracting multi-scale texture, structure and other features of power equipment in visible light images. Relatively speaking, the structural features of power equipment are obvious, but the infrared images can only show a few detailed features of power equipment, but more appearance features, and the infrared images contain more noise. The deep residual shrinkage network can suppress the influence of noise while extracting features. Summary of the invention

[0005] The present invention provides a method and system for detecting electric power equipment based on dynamic fusion of multiple feature channels, which are used to solve the technical problem that electric power equipment cannot be identified by multi-source images.

[0006] In a first aspect, the present invention provides a method for detecting electric power equipment based on dynamic fusion of multiple feature channels, comprising:

[0007] Acquire visible light images and infrared images of power equipment;

[0008] Performing feature extraction on the visible light image according to a preset first feature extraction backbone network to obtain a first feature map corresponding to the visible light image, and performing feature extraction on the infrared image according to a preset second feature extraction backbone network to obtain a second feature map corresponding to the infrared image;

[0009] The first feature map and the second feature map are connected to obtain an initial feature map, and a channel attention mechanism is used to perform feature channel screening on the initial feature map to obtain a target feature map;

[0010] The first characteristic diagram, the second characteristic diagram and the target characteristic diagram are input into a preset detection network, and the detection network outputs a type result of the electric power equipment.

[0011] In a second aspect, the present invention provides a power equipment detection system based on dynamic fusion of multiple feature channels, comprising:

[0012] An acquisition module configured to acquire visible light images and infrared images of the electric power equipment;

[0013] an extraction module, configured to perform feature extraction on the visible light image according to a preset first feature extraction backbone network to obtain a first feature map corresponding to the visible light image, and to perform feature extraction on the infrared image according to a preset second feature extraction backbone network to obtain a second feature map corresponding to the infrared image;

[0014] A screening module is configured to connect the first feature map and the second feature map to obtain an initial feature map, and use a channel attention mechanism to perform feature channel screening on the initial feature map to obtain a target feature map;

[0015] The output module is configured to input the first characteristic diagram, the second characteristic diagram and the target characteristic diagram into a preset detection network, and the detection network outputs the type result of the power equipment.

[0016] According to a third aspect, an electronic device is provided, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can perform the steps of the power equipment detection method based on dynamic fusion of multiple feature channels according to any embodiment of the present invention.

[0017] In a fourth aspect, the present invention further provides a computer-readable storage medium having a computer program stored thereon, wherein when the program instructions are executed by a processor, the processor executes the steps of the power equipment detection method based on dynamic fusion of multiple feature channels of any embodiment of the present invention.

[0018] The electric power equipment detection method and system based on dynamic fusion of multiple feature channels of the present application, for the registered visible light images and infrared images of the electric power equipment, extracts the multi-scale texture, structure and other features of the electric power equipment in the visible light image on the basis of considering the magnification of the input image, adopts the residual shrinkage network in the infrared image to suppress the infrared image noise and extract the structural features of the electric power equipment, adopts the channel attention mechanism dynamic feature fusion method to fuse the features of the same scale, and dynamically selects the appropriate features, thereby improving the accuracy of electric power equipment recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0020] Figure 1 A flow chart of a method for detecting electric power equipment based on dynamic fusion of multiple feature channels provided by one embodiment of the present invention;

[0021] Figure 2 A model framework diagram of power equipment detection based on dynamic fusion of multiple feature channels provided by an embodiment of the present invention;

[0022] Figure 3 A structural block diagram of a power equipment detection system based on dynamic fusion of multiple feature channels provided by an embodiment of the present invention;

[0023] Figure 4 It is a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0025] See also Figure 1 , which shows a flow chart of a method for detecting electric power equipment based on dynamic fusion of multiple feature channels of the present application.

[0026] like Figure 1 As shown, the power equipment detection method based on dynamic fusion of multiple feature channels specifically includes the following steps:

[0027] Step S101, obtaining a visible light image and an infrared image of the power equipment.

[0028] Step S102, performing feature extraction on the visible light image according to a preset first feature extraction backbone network to obtain a first feature map corresponding to the visible light image, and performing feature extraction on the infrared image according to a preset second feature extraction backbone network to obtain a second feature map corresponding to the infrared image.

[0029] In this step, the first feature extraction backbone network includes a first layer, a second layer, a third layer and a fourth layer, the first layer includes a convolution module, the second layer and the third layer both include a convolution module and a C2f module, and the fourth layer includes a convolution module, a C2f module and an SPPF module; wherein the convolution module has a convolution kernel size k=3, a step size s=2, and a padding size p=1, and the number of bottleneck blocks of the C2f module is n=3×d, where d is a depth factor.

[0030] It should be noted that the second feature extraction backbone network includes a convolution module and a two-dimensional residual shrinkage module;

[0031] The two-dimensional residual shrinkage module obtains a set of thresholds from the width and height dimensions of the infrared image through learning, and performs soft thresholding. For the feature map FS_in with an input of C×W×H, the specific processing process includes:

[0032] After two residual module processing, the intermediate result FS_mid is obtained, where the residual module includes batch normalization, rectified linear unit activation function, and convolution layer with convolution kernel k=c;

[0033] In the height direction of the image, take the absolute value of the intermediate result FS_mid, then perform global mean pooling to obtain a feature vector f_gap_h of size C×1×H. The feature vector f_gap_h passes through a two-layer fully connected layer to obtain a scaling parameter z_h. The scaling parameter z_h is regularized to between zero and one using the Sigmoid function to obtain the first scaling parameter α_h. The regularized first scaling parameter α_h is multiplied by the feature vector f_gap_h as the threshold, and the intermediate result FS_mid is processed to obtain the first soft thresholding result f_h.

[0034] In the width direction of the image, take the absolute value of the intermediate result FS_mid, then perform global mean pooling to obtain a feature vector f_gap_w of size C×W×1. The feature vector f_gap_w passes through a two-layer fully connected layer to obtain a scaling parameter z_w. The scaling parameter z_w is regularized to between zero and one using the Sigmoid function to obtain the second scaling parameter α_w. The regularized second scaling parameter α_w is multiplied by the feature vector f_gap_w as the threshold, and the intermediate result FS_mid is processed to obtain the second soft thresholding result f_w.

[0035] The first soft thresholding result f_h, the second soft thresholding result f_w and the identity mapping scaling parameter result of the feature map FS_in are superimposed to obtain a feature map FS_out with an output size of C×W×H of the basic residual shrinkage module, that is, the second feature map.

[0036] Step S103: Connect the first feature map and the second feature map to obtain an initial feature map, and use a channel attention mechanism to screen feature channels of the initial feature map to obtain a target feature map.

[0037] In this step, the initial feature map is compressed , we get a vector Z of size 1×1×Cse, is the number of channels after the compression operation, and the expression is:

[0038] ,

[0039] In the formula, is a channel of the initial feature map, is the height of the initial feature map, is the width of the initial feature map, for The value at position (i, j), is the vector obtained after compression, For sum operation;

[0040] Activate the vector Z to get a weight vector s of size 1×1×Cse, expressed as:

[0041] ,

[0042] In the formula, , They are ReLU activation function and Sigmod activation function respectively. is the weight matrix, , are all decompositions of the weight matrix W, , , For and the activation function of W, for The calculation shows that represents the matrix, is the number of hidden nodes in the middle layer, Indicates channels;

[0043] The initial feature map and weight vector s are scaled, and the expression is:

[0044] ,

[0045] In the formula, is the weight corresponding to a channel in the weight vector s, is the scaling function;

[0046] Will Merge with the initial feature map to get a new feature map, namely the target feature map.

[0047] Step S104: input the first characteristic graph, the second characteristic graph and the target characteristic graph into a preset detection network, and the detection network outputs a type result of the electric power equipment.

[0048] In this step, the detection network includes a first branch and a second branch, each of which includes a target convolution module, an occlusion-aware attention network, and a 2D convolution module connected in sequence, wherein the parameters of the target convolution module are a convolution kernel size k=3, a step size =1, padding size p=1, the parameters of the 2D convolution are convolution kernel size k=1, step size =1, padding size p=0;

[0049] Among them, the first feature map, the second feature map and the target feature map are output through the target convolution module to obtain a feature F_SEAM_in of size Wp×Hp×Cp.

[0050] It should be noted that the specific steps performed by the occlusion-aware attention network include:

[0051] Input the feature F_SEAM_in into the first convolution, and through the depth convolution, Gaussian error linear unit activation function and batch normalization operation in the first convolution, and add it to the input through the residual connection to obtain the target result, and input the target result into the second convolution, and through point-by-point convolution, GeLU activation function and batch normalization operation, obtain the result F_mid;

[0052] The result F_mid passes through the global average pooling layer, compressing the feature map of each channel into a scalar value, and obtains a tensor of shape C×1×1, denoted as F_gap;

[0053] F_gap is processed by the first fully connected layer, the second fully connected layer, the ReLU activation function and the Sigmoid activation function to learn the dependencies between channels and generate channel attention weights, denoted as F_ca, where the first fully connected layer compresses the number of channels and the second fully connected layer restores the number of channels. The Sigmoid activation function limits the weight value to between 0 and 1;

[0054] After the exponential operation, F_ca is multiplied element-wise with the original input feature F_SEAM_in, and the output is recorded as F_out.

[0055] In summary, the method of the present application, the visible light image feature extraction is composed of a convolution module and a C2f module to extract multi-scale texture, structure and other features of the power equipment, the infrared image feature extraction is achieved by stacking a two-dimensional residual shrinkage module to better obtain the structural characteristics of the power equipment while suppressing the infrared image noise, the feature fusion network adopts a channel attention mechanism dynamic feature fusion method to dynamically select appropriate features from visible light and infrared images, and the prediction network judges the type of power equipment based on the features, thereby improving the accuracy of power equipment identification.

[0056] In a specific embodiment, the model framework of the method of the present application is as follows Figure 2 As shown in the figure, it consists of a feature extraction backbone network, a feature fusion network, and a prediction network. The input of the model framework is infrared images and visible light images. The feature extraction backbone network extracts the features of infrared images and visible light images respectively. The feature fusion network dynamically selects appropriate features from the features of infrared images and visible light images. The prediction network determines the type of power equipment based on the features.

[0057] The visible light image feature extraction network in the model feature extraction backbone network adopts a multi-stage convolution structure, and the infrared image feature extraction network adopts a multi-level two-dimensional residual shrinkage module; at the same time, the number of stages in the visible light image feature extraction network in the model feature extraction backbone network and the number of two-dimensional residual shrinkage modules in the infrared image feature extraction network are determined by the resolution ratio of the visible light image to the infrared image. For example, when the resolution ratio of the visible light image to the infrared image is 3, it includes three stages of feature extraction modules and three stacked two-dimensional residual shrinkage modules.

[0058] See also Figure 3 , which shows a structural block diagram of an electric power equipment detection system based on dynamic fusion of multiple feature channels of the present application.

[0059] like Figure 3As shown, the power equipment detection system 200 includes an acquisition module 210 , an extraction module 220 , a screening module 230 and an output module 240 .

[0060] Among them, the acquisition module 210 is configured to acquire a visible light image and an infrared image of the power equipment; the extraction module 220 is configured to perform feature extraction on the visible light image according to a preset first feature extraction backbone network to obtain a first feature map corresponding to the visible light image, and to perform feature extraction on the infrared image according to a preset second feature extraction backbone network to obtain a second feature map corresponding to the infrared image; the screening module 230 is configured to connect the first feature map and the second feature map to obtain an initial feature map, and use a channel attention mechanism to perform feature channel screening on the initial feature map to obtain a target feature map; the output module 240 is configured to input the first feature map, the second feature map and the target feature map into a preset detection network, and the detection network outputs the type result of the power equipment.

[0061] It should be understood that Figure 3 Modules and references documented in Figure 1 Therefore, the operations and features described above for the method and the corresponding technical effects are also applicable to Figure 3 The modules in it will not be described in detail here.

[0062] In some other embodiments, the embodiments of the present invention further provide a computer-readable storage medium having a computer program stored thereon, wherein when the program instructions are executed by a processor, the processor is caused to execute the power equipment detection method based on dynamic fusion of multiple feature channels in any of the above method embodiments;

[0063] As an implementation mode, the computer-readable storage medium of the present invention stores computer-executable instructions, and the computer-executable instructions are configured as follows:

[0064] Acquire visible light images and infrared images of power equipment;

[0065] Performing feature extraction on the visible light image according to a preset first feature extraction backbone network to obtain a first feature map corresponding to the visible light image, and performing feature extraction on the infrared image according to a preset second feature extraction backbone network to obtain a second feature map corresponding to the infrared image;

[0066] The first feature map and the second feature map are connected to obtain an initial feature map, and a channel attention mechanism is used to perform feature channel screening on the initial feature map to obtain a target feature map;

[0067] The first characteristic diagram, the second characteristic diagram and the target characteristic diagram are input into a preset detection network, and the detection network outputs a type result of the electric power equipment.

[0068] The computer-readable storage medium may include a program storage area and a data storage area, wherein the program storage area may store an operating system and application programs required for at least one function; the data storage area may store data created based on the use of the power equipment detection system based on the dynamic fusion of multiple feature channels, etc. In addition, the computer-readable storage medium may include a high-speed random access memory, and may also include a memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the computer-readable storage medium may optionally include a memory remotely arranged relative to the processor, and these remote memories may be connected to the power equipment detection system based on the dynamic fusion of multiple feature channels via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0069] Figure 4 is a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention, such as Figure 4 As shown, the device includes: a processor 310 and a memory 320. The electronic device may also include: an input device 330 and an output device 340. The processor 310, the memory 320, the input device 330 and the output device 340 may be connected via a bus or other means. Figure 4 In the example, the connection through the bus is taken as an example. The memory 320 is the above-mentioned computer-readable storage medium. The processor 310 executes various functional applications and data processing of the server by running the non-volatile software programs, instructions and modules stored in the memory 320, that is, the power equipment detection method based on the dynamic fusion of multiple feature channels in the above-mentioned method embodiment is realized. The input device 330 can receive input digital or character information, and generate key signal input related to user settings and function control of the power equipment detection system based on the dynamic fusion of multiple feature channels. The output device 340 may include display devices such as display screens.

[0070] The electronic device can execute the method provided by the embodiment of the present invention, and has the functional modules and beneficial effects corresponding to the execution method. For technical details not described in detail in this embodiment, please refer to the method provided by the embodiment of the present invention.

[0071] As an implementation mode, the electronic device is applied to a power equipment detection system based on dynamic fusion of multiple feature channels, and is used for a client, and includes: at least one processor; and a memory connected to the at least one processor in communication; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can:

[0072] Acquire visible light images and infrared images of power equipment;

[0073] Performing feature extraction on the visible light image according to a preset first feature extraction backbone network to obtain a first feature map corresponding to the visible light image, and performing feature extraction on the infrared image according to a preset second feature extraction backbone network to obtain a second feature map corresponding to the infrared image;

[0074] The first feature map and the second feature map are connected to obtain an initial feature map, and a channel attention mechanism is used to perform feature channel screening on the initial feature map to obtain a target feature map;

[0075] The first characteristic diagram, the second characteristic diagram and the target characteristic diagram are input into a preset detection network, and the detection network outputs a type result of the electric power equipment.

[0076] Through the description of the above implementation modes, those skilled in the art can clearly understand that each implementation mode can be implemented by means of software plus a necessary general hardware platform, or of course by hardware. Based on such an understanding, the above technical solution can essentially or in other words be embodied in the form of a software product that contributes to the prior art. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods of each embodiment or some parts of the embodiment.

[0077] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A power equipment detection method based on dynamic fusion of multiple feature channels, characterized in that: include: Acquire visible light images and infrared images of power equipment; Performing feature extraction on the visible light image according to a preset first feature extraction backbone network to obtain a first feature map corresponding to the visible light image, and performing feature extraction on the infrared image according to a preset second feature extraction backbone network to obtain a second feature map corresponding to the infrared image, wherein the second feature extraction backbone network includes a convolution module and a two-dimensional residual shrinkage module; The two-dimensional residual shrinkage module obtains a set of thresholds from the width and height dimensions of the infrared image through learning, and performs soft thresholding. For the feature map FS_in with an input of C×W×H, the specific processing process includes: After two residual module processing, the intermediate result FS_mid is obtained, where the residual module includes batch normalization, rectified linear unit activation function, and convolution layer with convolution kernel k=c; In the height direction of the image, take the absolute value of the intermediate result FS_mid, then perform global mean pooling to obtain a feature vector f_gap_h of size C×1×H. The feature vector f_gap_h passes through a two-layer fully connected layer to obtain a scaling parameter z_h. The scaling parameter z_h is regularized to between zero and one using the Sigmoid function to obtain the first scaling parameter α_h. The regularized first scaling parameter α_h is multiplied by the feature vector f_gap_h as the threshold, and the intermediate result FS_mid is processed to obtain the first soft thresholding result f_h. In the width direction of the image, take the absolute value of the intermediate result FS_mid, then perform global mean pooling to obtain a feature vector f_gap_w of size C×W×1. The feature vector f_gap_w passes through a two-layer fully connected layer to obtain a scaling parameter z_w. The scaling parameter z_w is regularized to between zero and one using the Sigmoid function to obtain the second scaling parameter α_w. The regularized second scaling parameter α_w is multiplied by the feature vector f_gap_w as the threshold, and the intermediate result FS_mid is processed to obtain the second soft thresholding result f_w. The first soft thresholding result f_h, the second soft thresholding result f_w and the identity mapping scaling parameter result of the feature map FS_in are superimposed to obtain a feature map FS_out with an output size of C×W×H of the basic residual shrinkage module, i.e., a second feature map; The first feature map and the second feature map are connected to obtain an initial feature map, and a channel attention mechanism is used to perform feature channel screening on the initial feature map to obtain a target feature map; The first characteristic diagram, the second characteristic diagram and the target characteristic diagram are input into a preset detection network, and the detection network outputs a type result of the electric power equipment.

2. According to claim 1, a power equipment detection method based on dynamic fusion of multiple feature channels is characterized in that: The first feature extraction backbone network includes a first layer, a second layer, a third layer and a fourth layer, the first layer includes a convolution module, the second layer and the third layer both include a convolution module and a C2f module, and the fourth layer includes a convolution module, a C2f module and an SPPF module; Among them, the convolution kernel size of the convolution module is k=3, the step size is s=2, and the padding size is p=1. The number of bottleneck blocks of the C2f module is n=3×d, where d is the depth factor.

3. The method for detecting electric power equipment based on dynamic fusion of multiple feature channels according to claim 1 is characterized in that: The channel attention mechanism is used to perform feature channel screening on the initial feature map to obtain a target feature map, including: Compress the initial feature map , we get a vector Z of size 1×1×Cse, is the number of channels after the compression operation, and the expression is: ， In the formula, is a channel of the initial feature map, is the height of the initial feature map, is the width of the initial feature map, for The value at position (i, j), is the vector obtained after compression, For sum operation; Activate the vector Z to get a weight vector s of size 1×1×Cse, where the height is 1, the width is 1, and the number of channels is Cse. The expression is: ， In the formula, , They are ReLU activation function and Sigmod activation function respectively. is the weight matrix, , are all decompositions of the weight matrix W, , , For and the activation function of W, for The calculation shows that represents the matrix, is the number of hidden nodes in the middle layer, Indicates channels; The initial feature map and weight vector s are scaled, and the expression is: ， In the formula, is the weight corresponding to a channel in the weight vector s, is the scaling function; Will Merge with the initial feature map to get a new feature map, namely the target feature map.

4. The method for detecting electric power equipment based on dynamic fusion of multiple feature channels according to claim 1 is characterized in that: The detection network includes a first branch and a second branch, each of which includes a target convolution module, an occlusion-aware attention network, and a 2D convolution module connected in sequence, wherein the parameters of the target convolution module are a convolution kernel size k=3, a step size =1, padding size p=1, the parameters of the 2D convolution are convolution kernel size k=1, step size =1, padding size p=0; Among them, the first feature map, the second feature map and the target feature map are output through the target convolution module to obtain a feature F_SEAM_in of size Wp×Hp×Cp.

5. The method for detecting electric power equipment based on dynamic fusion of multiple feature channels according to claim 4 is characterized in that: The steps specifically performed by the occlusion-aware attention network include: Input the feature F_SEAM_in into the first convolution, and through the depth convolution, Gaussian error linear unit activation function and batch normalization operation in the first convolution, and add it to the input through the residual connection to obtain the target result, and input the target result into the second convolution, and through point-by-point convolution, GeLU activation function and batch normalization operation, obtain the result F_mid; The result F_mid passes through the global average pooling layer, compressing the feature map of each channel into a scalar value, and obtains a tensor of shape C×1×1, denoted as F_gap; F_gap is processed by the first fully connected layer, the second fully connected layer, the ReLU activation function and the Sigmoid activation function to learn the dependencies between channels and generate channel attention weights, denoted as F_ca, where the first fully connected layer compresses the number of channels and the second fully connected layer restores the number of channels; The Sigmoid activation function limits the weight value to between 0 and 1; After the exponential operation, F_ca is multiplied element-wise with the original input feature F_SEAM_in, and the output is recorded as F_out.

6. A power equipment detection system based on dynamic fusion of multiple feature channels, characterized in that: include: An acquisition module configured to acquire visible light images and infrared images of the electric power equipment; an extraction module, configured to perform feature extraction on the visible light image according to a preset first feature extraction backbone network to obtain a first feature map corresponding to the visible light image, and to perform feature extraction on the infrared image according to a preset second feature extraction backbone network to obtain a second feature map corresponding to the infrared image, wherein the second feature extraction backbone network includes a convolution module and a two-dimensional residual shrinkage module; The two-dimensional residual shrinkage module obtains a set of thresholds from the width and height dimensions of the infrared image through learning, and performs soft thresholding. For the feature map FS_in with an input of C×W×H, the specific processing process includes: After two residual module processing, the intermediate result FS_mid is obtained, where the residual module includes batch normalization, rectified linear unit activation function, and convolution layer with convolution kernel k=c; In the height direction of the image, take the absolute value of the intermediate result FS_mid, then perform global mean pooling to obtain a feature vector f_gap_h of size C×1×H. The feature vector f_gap_h passes through a two-layer fully connected layer to obtain a scaling parameter z_h. The scaling parameter z_h is regularized to between zero and one using the Sigmoid function to obtain the first scaling parameter α_h. The regularized first scaling parameter α_h is multiplied by the feature vector f_gap_h as the threshold, and the intermediate result FS_mid is processed to obtain the first soft thresholding result f_h. In the width direction of the image, take the absolute value of the intermediate result FS_mid, then perform global mean pooling to obtain a feature vector f_gap_w of size C×W×1. The feature vector f_gap_w passes through a two-layer fully connected layer to obtain a scaling parameter z_w. The scaling parameter z_w is regularized to between zero and one using the Sigmoid function to obtain the second scaling parameter α_w. The regularized second scaling parameter α_w is multiplied by the feature vector f_gap_w as the threshold, and the intermediate result FS_mid is processed to obtain the second soft thresholding result f_w. The first soft thresholding result f_h, the second soft thresholding result f_w and the identity mapping scaling parameter result of the feature map FS_in are superimposed to obtain a feature map FS_out with an output size of C×W×H of the basic residual shrinkage module, i.e., a second feature map; A screening module is configured to connect the first feature map and the second feature map to obtain an initial feature map, and use a channel attention mechanism to perform feature channel screening on the initial feature map to obtain a target feature map; The output module is configured to input the first characteristic diagram, the second characteristic diagram and the target characteristic diagram into a preset detection network, and the detection network outputs the type result of the power equipment.

7. An electronic device, characterized in that: include: At least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method described in any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

Images