Remote sensing image transmission tower detection method based on MFF-YOLO

Abstract

The invention relates to a remote sensing image transmission tower detection method based on MFF-YOLO, and the method comprises the following steps: S1, obtaining a slice data set of a satellite image of a region, screening slice data containing a transmission tower as a data set, and carrying out the image preprocessing; s2, labeling the power transmission tower by utilizing Labelimg software, and obtaining a data set label in an xml (Extensible Markup Language) format; s3, constructing a power transmission tower detection network model based on MFF-YOLO; s4, training and verifying the transmission tower detection network model based on the constructed data set; and S5, on the basis of the trained transmission tower detection network model, detecting a satellite image of a to-be-detected area. According to the invention, the available computing resources can be utilized more efficiently to detect the transmission tower target, and the efficiency and accuracy of transmission tower identification are improved.

Description

Technical Field

[0001] The present invention relates to, and in particular to, a remote sensing image transmission tower detection method based on MFF-YOLO. Background Art

[0002] With the development of remote sensing technology, remote sensing images have become increasingly important in environmental monitoring, urban planning, industrial inspection, etc. Target detection of transmission towers through remote sensing images can provide accurate tower location and layout information for power planning, and optimize grid layout and operation management. However, compared with remote sensing images, tower target objects are too small in size and have limited features. In addition, remote sensing images have the characteristics of large resolution, uneven quality, and complex background. Therefore, there are huge challenges in the detection of transmission towers using remote sensing images. Summary of the invention

[0003] In order to solve the above problems, the purpose of the present invention is to provide a remote sensing image transmission tower detection method based on MFF-YOLO, which can more efficiently utilize available computing resources to detect transmission tower targets, thereby improving the efficiency and accuracy of transmission tower identification.

[0004] To achieve the above object, the present invention adopts the following technical solutions:

[0005] A remote sensing image transmission tower detection method based on MFF-YOLO, comprising the following steps:

[0006] Step S1: Obtain a slice data set of satellite images of a region, select slice data containing transmission towers as a data set, and perform image preprocessing;

[0007] Step S2: Use Labelimg software to label the transmission tower and obtain the dataset label in XML format;

[0008] Step S3: construct a transmission tower detection network model based on MFF-YOLO;

[0009] Step S4: training and verifying the transmission tower detection network model based on the constructed data set;

[0010] Step S5: Based on the trained transmission tower detection network model, the satellite image of the area to be detected is detected.

[0011] Furthermore, a transmission tower detection network model is constructed based on MFF-YOLO, as follows: the MFF-YOLO is a new network established on the basis of the YOLOv9 network by introducing the SCFG convolutional layer and the triple dynamic attention module TDAM, which is divided into three parts: the backbone network, the neck network and the head network; the backbone network is composed of five SCFG convolutional layers, four RepNCSPELAN4 modules, a SPPELAN module and a TDAM module; the neck network is composed of four RepNCSPELAN4 modules, two convolutional layers and two upsampling modules; the head network is composed of the sixth RepNCSPELAN4 module, the fifth RepNCSPELAN4 module and the eighth RepNCSPELAN4 module, which are respectively output to three convolutional modules for processing to obtain output features.

[0012] Furthermore, SCFG includes a spatial refinement unit SFG and a channel reconstruction unit CFG. The main steps include: obtaining the spatial refinement feature X through the SFG module w , and then input it into the CFG module to obtain the channel refined feature Y, and finally input the feature quantity Y into the next layer module.

[0013] Furthermore, the spatial refinement unit SFG includes two operations: splitting and fusion. The splitting operation includes multiplying the input feature X normalized by the GN group and the renormalized related weight parameters to obtain the importance weight matrix; inputting the importance weight matrix into the Relu function, and mapping it to the range [0,1] to obtain the information weight W1 and the non-information weight W2; then multiplying the input feature X by W1 and W2 respectively to obtain the feature quantity with most feature representation and redundant information The formula for the weight coefficient W is:

[0014] W=Gate(Relu(W τ (GN(X))))

[0015] The GN formula is:

[0016]

[0017] Among them, μ and σ are the mean and standard deviation of X, τ and β are trainable affine transformations;

[0018] The SFG fusion operation will and Each is divided into two sub-parts, and the two sub-parts are cross-added and reconstructed respectively, and the two reconstructed outputs are combined to obtain the spatial refined feature X w , the reconstruction operation is expressed as:

[0019]

[0020] Furthermore, the CFG channel refinement module includes three operations: decomposition, extraction, and fusion. The decomposition operation is to refine the spatial features X output by SFG. w Split into two parts: αC channel and (1-α)C channel, where α (0≤α≤1) is the ratio of the number of upper channels to the output channels, and the default value is Then, the features of the output of the upper and lower channels are compressed by 1×1 convolution to obtain the upper channel feature quantity X high and the lower channel feature X low ;

[0021] The output of the extraction operation corresponding to the split operation can be divided into two parts: the upper and lower layers. The upper extraction module converts X up Input to the global weight convolution GWC of size k×k, and add the output and X high Combined, the merged upper channel representative feature Y1 is obtained. GWC can introduce global information in the channel dimension; the formula for the upper channel representative feature Y1 is:

[0022] Y1=M G X high +X high

[0023] Among them, M G is the learnable weight matrix of GWC, X high and Y1 are the input and output feature representations of the upper-layer conversion module, respectively;

[0024] The lower layer extraction module converts X low Input to the depthwise separable convolution DSC to obtain feature maps with deep hidden details while reducing the number of parameters and computational cost; the output and X low Combined to obtain the combined upper channel representative feature Y 2； The formula for the representative feature Y2 of the lower channel is:

[0025] Y2=M D X low +X low

[0026] in, is the learnable weight matrix of DSC, X low and Y2 are the input and output feature representations of the upper transformation module, respectively;

[0027] The fusion operation concatenates Y1 and Y2 in the channel dimension through the global maximum pooling layer to obtain the merged features S1 and S2 of the upper and lower layers. S1 and S2 are superimposed together, and the probability distribution of each part of the pooled features is calculated through the sigmoid function to obtain weights β1 and β2. The weights β1 and β2 are multiplied by the representative features of the upper and lower channels Y1 and Y2 in the channel direction and then added to obtain the channel refined feature Y.

[0028] Furthermore, TDAM consists of three attention branches and a dynamic weight mechanism, including Fourier frequency attention branch, contextual attention CoA branch and lightweight channel attention LCA branch. The dynamic weight mechanism is used to adaptively adjust the importance of different branches. The three-branch output features X1, X2 and X3 are input into the dynamic weight mechanism to realize the fusion of the feature maps of the three attention outputs to obtain the final output feature Y.

[0029] Furthermore, the Fourier frequency attention branch first performs a Fourier transform on the input feature X along the dimension of the feature vector, and then performs a Fourier transform along the dimension of the sequence, and finally inputs it into the real function, taking only the real number part, and obtains the feature importance weight by transforming the input feature from the spatial domain to the frequency domain, and then multiplies it with the input feature map to obtain the output feature map X1 of the Fourier frequency attention branch; Fourier attention converts the attention calculation to the frequency domain by utilizing the properties of the Fourier transform, thereby improving the computational efficiency. The formula for the Fourier frequency attention mechanism is:

[0030]

[0031] in, It represents Fourier transform, and Real() represents taking the real number in the brackets.

[0032] Furthermore, the CoA branch uses the key of the input feature as static context information, connects the key of the input feature after k×k convolution with the query through concat, and inputs the connected fusion value into two consecutive 1×1 convolutions to learn the dynamic multi-head attention matrix; the learned attention matrix is ​​multiplied by the value of the input feature after 1×1 convolution to realize the dynamic context representation of the input; the dynamic and static context are fused as the output feature of the CoA branch; the formula for CoA is:

[0033]

[0034] Among them, concat() is a connection operation, W1 and W2 represent two 1×1 convolutions respectively. It is a regulating factor used to prevent the gradient from being too small. Q, K, and V are the query, key, and value of the input feature, respectively. Fusion() is a fusion operation.

[0035] Furthermore, the LCA branch first performs global average pooling on the input feature X, then captures the dependencies between channels through a one-dimensional convolution operation with a kernel size of k, and then obtains the importance weight through the Sigmoid function, and multiplies the importance weight with the input feature X to obtain the output feature X3; the formula for the LCA channel attention mechanism is:

[0036] X3=Sigmoid(Pooling(X)*W k )X

[0037] Among them, Pooling() represents the global average pooling operation, W k represents k×k convolution, Sigmoid() represents the Sigmoid function; the kernel size k is adaptively calculated according to the number of channels, and the formula for k is:

[0038]

[0039] Among them, C is the number of channels of the input feature, and γ and b are hyperparameters. || odd It means taking the absolute value and rounding down to the nearest odd number.

[0040] Furthermore, the three-branch output features X1, X2 and X3 are input into the dynamic weight mechanism to realize the fusion of the feature maps of the three attention outputs and obtain the final output feature Y, as follows:

[0041] The three feature maps are globally averaged and normalized, and then the Softmax function is used to map the weight values ​​to a probability distribution to obtain their respective importance weights γ1, γ2, and γ3. Each attention output feature map is weightedly fused according to their weights to obtain the final output feature map Y. The formula of the dynamic weight module is:

[0042]

[0043] The present invention has the following beneficial effects:

[0044] 1. The present invention adopts MFF-YOLO as the target detection model. Compared with traditional image processing methods, traditional machine learning methods and deep learning methods, the model can more efficiently use available computing resources to detect transmission tower targets, thereby improving the efficiency and accuracy of transmission tower identification;

[0045] 2. The present invention introduces the SCFG convolution layer, which makes better use of spatial and channel correlation to improve the computing performance and detection accuracy of the model. It adopts the TDAM triple dynamic attention module and uses multi-feature fusion to further optimize the target recognition accuracy of the transmission tower, ensuring the accuracy of transmission tower target detection. It can make full use of the existing remote sensing image data set at the feature level, effectively optimize the utilization of computing resources, and provide a more optimized solution for the operation and maintenance of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is a flow chart of the method of the present invention;

[0047] Figure 2 This is a flowchart of the MFF-YOLO workflow in one embodiment of the present invention;

[0048] Figure 3 This is a triple dynamic attention TDAM module in one embodiment of the present invention. DETAILED DESCRIPTION

[0049] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0050] refer to Figure 1 In this embodiment, a remote sensing image transmission tower detection method based on MFF-YOLO is provided, which is characterized by comprising the following steps:

[0051] Step S1: Obtain a slice data set of satellite images of a region, select slice data containing transmission towers as a data set, and perform image preprocessing;

[0052] Step S2: Use Labelimg software to label the transmission tower and obtain the dataset label in XML format;

[0053] Step S3: construct a transmission tower detection network model based on MFF-YOLO;

[0054] Step S4: training and verifying the transmission tower detection network model based on the constructed data set;

[0055] Step S5: Based on the trained transmission tower detection network model, the satellite image of the area to be detected is detected.

[0056] In this embodiment, a transmission tower detection network model is constructed based on MFF-YOLO, as follows: the MFF-YOLO is a new network established by introducing SCFG convolutional layers and triple dynamic attention modules TDAM on the basis of the YOLOv9 network, and is divided into three parts: a backbone network, a neck network and a head network; the backbone network is composed of five SCFG convolutional layers, four RepNCSPELAN4 modules, a SPPELAN module and a TDAM module; the neck network is composed of four RepNCSPELAN4 modules, two convolutional layers and two upsampling modules; the head network is composed of the sixth RepNCSPELAN4 module, the fifth RepNCSPELAN4 module and the eighth RepNCSPELAN4 module, which are respectively output to three convolutional modules for processing to obtain output features.

[0057] In this embodiment, SCFG includes a spatial refinement unit SFG and a channel reconstruction unit CFG. The main steps include: obtaining the spatial refinement feature X by passing the feature output of the previous layer through the SFG module w , and then input it into the CFG module to obtain the channel refined feature Y, and finally input the feature quantity Y into the next layer module. Using the combined module of the SFG module and the CFG module in series to replace the fixed convolution can reduce the computational redundancy generated by the feature mapping.

[0058] In this embodiment, the spatial refinement unit SFG includes two operations: splitting and fusion. The splitting operation includes multiplying the input feature X normalized by the GN group and the renormalized related weight parameters to obtain an importance weight matrix; inputting the importance weight matrix into the Relu function, and mapping it to the range [0,1] to obtain the information weight W1 and the non-information weight W2; then multiplying the input feature X by W1 and W2 respectively to obtain the feature quantity with most feature representation and redundant information The separation operation decomposes the input features into parts with important features and parts with redundant information, which helps to extract and retain important features useful for the task and reduce spatial redundant information in the model. The formula for the weight coefficient W is:

[0059] W=Gate(Relu(W τ (GN(X))))

[0060] The GN formula is:

[0061]

[0062] Among them, μ and σ are the mean and standard deviation of X, τ and β are trainable affine transformations;

[0063] The SFG fusion operation will and Each is divided into two sub-parts, and the two sub-parts are cross-added and reconstructed respectively, and the two reconstructed outputs are combined to obtain the spatial refined feature X w , the reconstruction operation is expressed as:

[0064]

[0065] In this embodiment, the CFG channel refinement module includes three operations: decomposition, extraction and fusion. The decomposition operation is to refine the spatial features X output by SFG. w Split into two parts: αC channel and (1-α)C channel, where α (0≤α≤1) is the ratio of the number of upper channels to the output channels, and the default value is Then, the features of the output of the upper and lower channels are compressed by 1×1 convolution to obtain the upper channel feature quantity X high and the lower channel feature X low ;

[0066] The output of the extraction operation corresponding to the split operation can be divided into two parts: the upper and lower layers. The upper extraction module converts X up Input to the global weight convolution GWC of size k×k, and add the output and X high Combined, the merged upper channel representative feature Y1 is obtained. GWC can introduce global information in the channel dimension; the formula for the upper channel representative feature Y1 is:

[0067] Y1=M G X high +X high

[0068] Among them, M G is the learnable weight matrix of GWC, X high and Y1 are the input and output feature representations of the upper-layer conversion module, respectively;

[0069] The lower layer extraction module converts X low Input to the depthwise separable convolution DSC to obtain feature maps with deep hidden details while reducing the number of parameters and computational cost; the output and X low Combined to obtain the combined upper channel representative feature Y 2； The formula for the representative feature Y2 of the lower channel is:

[0070] Y2=M D X low +X low

[0071] Among them, M P2 is the learnable weight matrix of DSC, X lowand Y2 are the input and output feature representations of the upper transformation module, respectively;

[0072] The fusion operation concatenates Y1 and Y2 in the channel dimension through the global maximum pooling layer to obtain the merged features S1 and S2 of the upper and lower layers. S1 and S2 are superimposed together, and the probability distribution of each part of the pooled features is calculated through the sigmoid function to obtain weights β1 and β2. The weights β1 and β2 are multiplied by the representative features Y1 and Y2 of the upper and lower channels in the channel direction and then added to obtain the channel refined feature Y. CFGU can simplify the spatial fine feature mapping X w The channel features of the network are used to extract rich representative features through lightweight convolution operations, while redundant features are processed through low-cost operation steps to further reduce the computational cost.

[0073] In this embodiment, TDAM consists of three attention branches and a dynamic weight mechanism, including a Fourier frequency attention branch, a contextual attention CoA branch, and a lightweight channel attention LCA branch. The dynamic weight mechanism is used to adaptively adjust the importance of different branches. The three-branch output features X1, X2, and X3 are input into the dynamic weight mechanism to realize the fusion of the feature maps of the three attention outputs to obtain the final output feature Y.

[0074] In this embodiment, the Fourier frequency attention branch first performs Fourier transform on the input feature X along the dimension of the feature vector, and then performs a Fourier transform along the dimension of the sequence, and finally inputs it into the real function, taking only the real number part, and obtains the feature importance weight by transforming the input feature from the spatial domain to the frequency domain, and then multiplies it with the input feature map to obtain the output feature map X1 of the Fourier frequency attention branch; Fourier attention converts the attention calculation to the frequency domain by utilizing the properties of the Fourier transform, thereby improving the computational efficiency. The formula for the Fourier frequency attention mechanism is:

[0075]

[0076] in, It represents Fourier transform, and Real() represents taking the real number in the brackets.

[0077] In this embodiment, the CoA branch uses the key of the input feature as static context information, connects the key of the input feature after k×k convolution with the query through concat, and inputs the connected fusion value into two consecutive 1×1 convolutions to learn the dynamic multi-head attention matrix; the learned attention matrix is ​​multiplied by the value of the input feature after 1×1 convolution to realize the dynamic context representation of the input; the dynamic and static context are fused as the output feature of the CoA branch; CoA reduces the computational complexity compared to the traditional self-attention mechanism by introducing continuous 1×1 convolution operations. The formula for CoA is:

[0078]

[0079] Among them, concat() is a connection operation, W1 and W2 represent two 1×1 convolutions respectively. It is a regulating factor used to prevent the gradient from being too small. Q, K, and V are the query, key, and value of the input feature, respectively. Fusion() is a fusion operation.

[0080] In this embodiment, the LCA branch first performs global average pooling on the input feature X, then uses a one-dimensional convolution operation with a kernel size of k to capture the dependencies between channels, and then obtains the importance weight through the Sigmoid function, and multiplies the importance weight with the input feature X to obtain the output feature X3; the main advantage of LCA is that it only focuses on the feature interaction in the channel dimension, and does not involve the interaction in the spatial dimension, which reduces the computational complexity and memory consumption. The formula for the LCA channel attention mechanism is:

[0081] X3=Sigmoid(Pooling(X)*W k )X

[0082] Among them, Pooling() represents the global average pooling operation, W k represents k×k convolution, Sigmoid() represents the Sigmoid function; the kernel size k is adaptively calculated according to the number of channels, and the formula for k is:

[0083]

[0084] Among them, C is the number of channels of the input feature, and γ and b are hyperparameters. || odd It means taking the absolute value and rounding down to the nearest odd number.

[0085] In this embodiment, the three-branch output features X1, X2 and X3 are input into the dynamic weight mechanism to realize the fusion of the feature maps of the three attention outputs to obtain the final output feature Y, as follows:

[0086] The three feature maps are globally averaged and normalized, and then the Softmax function is used to map the weight values ​​to a probability distribution to obtain their respective importance weights γ1, γ2, and γ3. Each attention output feature map is weightedly fused according to their weights to obtain the final output feature map Y. The formula of the dynamic weight module is:

[0087]

[0088] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0089] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0090] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0091] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0092] The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any technician familiar with the profession may use the above disclosed technical content to change or modify it into an equivalent embodiment with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still belongs to the protection scope of the technical solution of the present invention.

Claims

1. A remote sensing image transmission tower detection method based on MFF-YOLO, characterized in that: The following steps are involved: Step S1: Obtain a slice data set of satellite images of the region, select slice data containing transmission towers as the data set, and perform image preprocessing; Step S2: Use Labelimg software to label the transmission tower and obtain the dataset label in XML format; Step S3: construct a transmission tower detection network model based on MFF-YOLO; Step S4: training and verifying the transmission tower detection network model based on the constructed data set; Step S5: Based on the trained transmission tower detection network model, the satellite image of the area to be detected is detected.

2. The remote sensing image transmission tower detection method based on MFF-YOLO according to claim 1 is characterized in that: The transmission tower detection network model based on MFF-YOLO is specifically as follows: the MFF-YOLO is a new network established by introducing SCFG convolutional layer and triple dynamic attention module TDAM on the basis of YOLOv9 network, and is divided into three parts: backbone network, neck network and head network; the backbone network is composed of five SCFG convolutional layers, four RepNCSPELAN4 modules, SPPELAN module and TDAM module; the neck network is composed of four RepNCSPELAN4 modules, two convolutional layers and two upsampling modules; the head network is composed of the sixth RepNCSPELAN4 module, the fifth RepNCSPELAN4 module and the eighth RepNCSPELAN4 module, which are respectively output to three convolutional modules for processing to obtain output features.

3. The remote sensing image transmission tower detection method based on MFF-YOLO according to claim 2 is characterized in that: The SCFG includes a spatial refinement unit SFG and a channel reconstruction unit CFG. The main steps include: obtaining the spatial refinement feature X by passing the features output by the previous layer through the SFG module w , and then input it into the CFG module to obtain the channel refined feature Y, and finally input the feature quantity Y into the next layer module.

4. The remote sensing image transmission tower detection method based on MFF-YOLO according to claim 2 is characterized in that: The spatial refinement unit SFG includes two operations: splitting and fusion. The splitting operation includes multiplying the input feature X normalized by the GN group and the renormalized related weight parameters to obtain an importance weight matrix; inputting the importance weight matrix into the Relu function, and mapping it to the range [0,1] to obtain the information weight W1 and the non-information weight W2; then multiplying the input feature X by W1 and W2 respectively to obtain the feature quantity with most feature representation and redundant information The formula for the weight coefficient W is: W=Gate(Relu(W τ (GN(X)))) The GN formula is: Among them, μ and σ are the mean and standard deviation of X, τ and β are trainable affine transformations; The SFG fusion operation will and Each is divided into two sub-parts, and the two sub-parts are cross-added and reconstructed respectively, and the two reconstructed outputs are combined to obtain the spatial refined feature X w , the reconstruction operation is expressed as:

5. The remote sensing image transmission tower detection method based on MFF-YOLO according to claim 4 is characterized in that: The CFG channel refinement module includes three operations: decomposition, extraction and fusion. The decomposition operation is to refine the spatial features X output by SFG. w Split into two parts: αC channel and (1-α)C channel, where α (0≤α≤1) is the ratio of the number of upper channels to the output channels, and the default value is Then, the features of the output of the upper and lower channels are compressed by 1×1 convolution to obtain the upper channel feature quantity X high and the lower channel feature X low； The output of the extraction operation corresponding to the split operation can be divided into two parts: the upper and lower layers. The upper extraction module converts X up Input to the global weight convolution GWC of size k×k, and add the output and X high Combined, the merged upper channel representative feature Y1 is obtained. GWC can introduce global information in the channel dimension; the formula for the upper channel representative feature Y1 is: Y1=M G X high +X high Among them, M G is the learnable weight matrix of GWC, X high and Y1 are the input and output feature representations of the upper-layer conversion module, respectively; The lower layer extraction module converts X low Input to the depthwise separable convolution DSC to obtain feature maps with deep hidden details while reducing the number of parameters and computational cost; the output and X low Combined to obtain the combined upper channel representative feature Y 2； The formula for the representative feature Y2 of the lower channel is: Y2=M D X low +X low in, is the learnable weight matrix of DSC, X low and Y2 are the input and output feature representations of the upper transformation module, respectively; The fusion operation concatenates Y1 and Y2 in the channel dimension through the global maximum pooling layer to obtain the merged features S1 and S2 of the upper and lower layers. S1 and S2 are superimposed together, and the probability distribution of each part of the pooled features is calculated through the sigmoid function to obtain weights β1 and β2. The weights β1 and β2 are multiplied by the representative features of the upper and lower channels Y1 and Y2 in the channel direction and then added to obtain the channel refined feature Y.

6. The remote sensing image transmission tower detection method based on MFF-YOLO according to claim 1 is characterized in that: The TDAM consists of three attention branches and a dynamic weight mechanism, including a Fourier frequency attention branch, a contextual attention CoA branch and a lightweight channel attention LCA branch. The dynamic weight mechanism is used to adaptively adjust the importance of different branches. The three-branch output features X1, X2 and X3 are input into the dynamic weight mechanism to realize the fusion of the feature maps of the three attention outputs to obtain the final output feature Y.

7. The remote sensing image transmission tower detection method based on MFF-YOLO according to claim 6 is characterized in that: The Fourier frequency attention branch first performs a Fourier transform on the input feature X along the dimension of the feature vector, and then performs a Fourier transform along the dimension of the sequence, and finally inputs it into the real function, taking only the real number part, and obtains the feature importance weight by transforming the input feature from the spatial domain to the frequency domain, and then multiplies it with the input feature map to obtain the output feature map X1 of the Fourier frequency attention branch; Fourier attention converts the attention calculation to the frequency domain by utilizing the properties of the Fourier transform, thereby improving the computational efficiency. The formula for the Fourier frequency attention mechanism is: in, It represents Fourier transform, and Real() represents taking the real number in the brackets.

8. The remote sensing image transmission tower detection method based on MFF-YOLO according to claim 6 is characterized in that: The CoA branch uses the key of the input feature as static context information, concatenates the key of the input feature after k×k convolution with the query through concat, and inputs the concatenated fusion value into two consecutive 1×1 convolutions to learn the dynamic multi-head attention matrix; The learned attention matrix is ​​multiplied by the value of the input feature after 1×1 convolution to realize the dynamic context representation of the input; the dynamic and static contexts are fused as the output features of the CoA branch; the formula for CoA is: Among them, concat() is a connection operation, W1 and W2 represent two 1×1 convolutions respectively. It is a regulating factor used to prevent the gradient from being too small. Q, K, and V are the query, key, and value of the input feature, respectively. Fusion() is a fusion operation.

9. The remote sensing image transmission tower detection method based on MFF-YOLO according to claim 6 is characterized in that: The LCA branch first performs global average pooling on the input feature X, then captures the dependencies between channels through a one-dimensional convolution operation with a kernel size of k, and then obtains the importance weight through the Sigmoid function, and multiplies the importance weight with the input feature X to obtain the output feature X3; the formula for the LCA channel attention mechanism is: X3=Sigmoid(Pooling(X)*W k )X Among them, Pooling() represents the global average pooling operation, W k represents k×k convolution, Sigmoid() represents the Sigmoid function; the kernel size k is adaptively calculated according to the number of channels, and the formula for k is: Among them, C is the number of channels of the input feature, and γ and b are hyperparameters. || odd It means taking the absolute value and rounding down to the nearest odd number.

10. The remote sensing image transmission tower detection method based on MFF-YOLO according to claim 6, characterized in that: The three-branch output features X1, X2 and X3 are input into the dynamic weight mechanism to realize the fusion of the feature maps of the three attention outputs to obtain the final output feature Y, as follows: The three feature maps are globally averaged and normalized, and then the Softmax function is used to map the weight values ​​to a probability distribution to obtain their respective importance weights γ1, γ2, and γ3. Each attention output feature map is weightedly fused according to their weights to obtain the final output feature map Y. The formula of the dynamic weight module is:

Images