A power line aerial image integrated severe weather removal method based on a multi-scale fourier network

Abstract

The application provides a power line aerial image integrated severe weather removal method based on a multi-scale Fourier network. The method acquires a power line aerial degraded image to be processed; inputs the degraded image into a pre-trained multi-scale Fourier network model; extracts multi-scale low-level features of the degraded image through a multi-scale structure module; processes the features by using an encoder, converts spatial domain features to a frequency domain by using a double-branch Fourier transform module, separately processes amplitude components and phase components, restores the amplitude components and the phase components to the spatial domain through inverse transformation after separate processing, and obtains encoded features; processes the encoded features for feature reconstruction by using a decoder, and selects a result corresponding to an original image resolution as a target clean image. The application can integrally extract and remove meteorological interference features, and effectively realize high-fidelity reconstruction of power line element structure features.

Description

Technical Field

[0001] This invention relates to the fields of image processing and computer vision technology, and in particular to an integrated method for removing severe weather from aerial images of power lines based on multi-scale Fourier networks. Background Technology

[0002] In UAV power line inspection scenarios, the quality of aerial images is easily affected by complex weather conditions such as fog, rain, and snow. These meteorological factors can lead to decreased image contrast and photometric degradation, while the edge features of target power facilities are easily obscured by rain streaks or snow particles. Existing image processing solutions often lack integrated and effective computation and restoration methods when faced with aerial images containing multiple meteorological degradation features, making it difficult to simultaneously remove fog distributions with global diffuse characteristics or rain and snow patterns at different scales. More limitingly, during the filtering operation for removing severe weather features, high-frequency feature information representing key rigid structures such as power lines and insulators is easily distorted and deformed. The residual meteorological degradation features and the loss of internal structural information severely reduce the restored quality of the target image, thus transmitting interfering source data to downstream instance segmentation modules or various automatic defect identification components, affecting the overall accuracy and stability of subsequent power visual inspection calculations. Summary of the Invention

[0003] To address the shortcomings of existing technologies, the present invention aims to provide an integrated method for removing severe weather from aerial images of power lines based on multi-scale Fourier networks. This method solves the problems of reduced contrast and target edge occlusion caused by degradation interference in existing aerial images under complex weather conditions, and overcomes the technical defects of conventional image restoration algorithms that make the rigid structures of key power facilities prone to distortion.

[0004] This invention provides an integrated method for severe weather removal in aerial images of power lines based on multi-scale Fourier networks, comprising: Acquire aerial images of degraded power lines to be processed; The degraded image is input into a pre-trained multi-scale Fourier network model, wherein the multi-scale Fourier network model is constructed with a multi-scale structure module and a two-branch Fourier transform module as its core. The multi-scale low-level features of the degraded image are extracted using the multi-scale structure module. The encoder is used to process the multi-scale low-level features, and the dual-branch Fourier transform module is used to transform the spatial features to the frequency domain during the processing. The amplitude component and the phase component are separated and processed respectively, and then restored to the spatial domain by inverse Fourier transform to achieve frequency domain global feature calibration and local feature extraction, so as to obtain the coded features. The encoded features are reconstructed using a decoder, and the output results at multiple scales are calculated. The result corresponding to the original image resolution in the output results at multiple scales is selected as the target clean image after removing severe weather.

[0005] The present invention has achieved the following beneficial effects: By introducing a multi-scale Fourier network model built around a multi-scale structure module and a two-branch Fourier transform module, this method achieves integrated computation and restoration of degraded images of transmission lines under foggy and rainy / snowy weather conditions. Utilizing frequency domain transformation logic, the method performs customized separation and denoising processing on the amplitude and phase components in the frequency domain. This accurately suppresses anomalously activated frequency bands representing meteorological noise while effectively preventing distortion in the extraction of phase information representing key rigid structures such as power lines and insulators. This scheme balances the removal of meteorological interference features with the reconstruction and fusion of local feature details of power facilities, effectively restoring the true contrast of the image and preserving its spatial texture structure. The final clean target image provides more accurate and reliable source data support for object segmentation and defect classification components in downstream edge computing devices.

[0006] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0007] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0008] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a diagram showing the overall structure of a multi-scale Fourier network (MFNET) model for an integrated severe weather removal method for aerial power line images based on multi-scale Fourier networks, as provided in this embodiment of the invention.

[0009] Figure 2 This is a schematic diagram of the network structure of the two-branch Fourier transform module (DFTB) inside the multi-scale Fourier network model in this embodiment of the invention; wherein, (a) is the external connection structure of the two-branch Fourier transform module (DFTB), and (b) is the flow diagram of the specific processing features of the two-branch Fourier transform network (DFTN). Detailed Implementation

[0010] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0011] In UAV power line inspection scenarios, the quality of aerial images is affected by weather conditions such as fog, rain, and snow, resulting in decreased image contrast, photometric degradation, and obscuring of target edge features by rain lines or snow particles. To achieve integrated computation and restoration of these degraded images, this embodiment provides an integrated severe weather removal method for aerial power line images based on a multi-scale Fourier network. The core network model used in this method is defined as a multi-scale Fourier network (MFNET). The multi-scale Fourier network uses a two-branch Fourier transform module (DFTB) and a multi-scale structure module (MSS) as its core computational architecture. Through the fusion computation of frequency domain feature mapping and multi-scale spatial features, it achieves the extraction and removal of meteorological interference features in degraded images, as well as the reconstruction of the structural features of power line components. Figure 1 As shown, the multi-scale Fourier network model in this embodiment adopts the classic encoder-decoder architecture and includes... , Until The input image is processed at multiple scale levels, including downsampling (D), multi-scale feature fusion (MIFM) and two-branch Fourier transform (DFTB) modules at each level, and then reconstructed layer by layer through upsampling (U) and skip connections (C), finally outputting a clean image result at multiple scales.

[0012] The method provided in this embodiment first acquires aerial images of degraded power lines at the execution front end. The acquired degraded images include RGB three-channel digital image data of power lines collected by a drone in foggy, rainy, or snowy conditions. The degraded images are then used as input variables and fed into a pre-established and trained multi-scale Fourier network model.

[0013] At the level of mathematical definition and data flow, the data matrix of a single input degraded image is defined as follows: Among them, parameters This represents the index value of the input sample sequence in the current batch, and its value ranges from the set of positive integers. Parameters This indicates the set batch size. This image data matrix The data dimension is defined as the real number space. , where parameters The parameter represents the total number of pixels in the height dimension of the input image. Represents the total number of pixels in the width dimension of the input image, a constant. This represents three independent color information channels: red, green, and blue. Before being input into the multi-scale Fourier network model for forward propagation, the image data matrix... The intensity values ​​of each pixel are strictly normalized to a linear value using a linear mapping. The range of floating-point numbers.

[0014] In the feature extraction stage, the multi-scale Fourier network model extracts multi-scale low-level features of the degraded image through its front-end multi-scale structural modules. The computational process of multi-scale feature extraction relies on the sequential combination of downsampling operators and convolution operators at different network layers. For the first layer scale of the network model (i.e., the basic layer corresponding to the highest spatial resolution of the initial original image), the input image matrix... Linear mapping is performed via a feature extraction convolutional layer. Specifically, a kernel size of [size missing] is used. The convolution operation yields the low-level features of the first layer, and this feature tensor is denoted as . Its mathematical formula is: ; In the above formula, the variables This represents the low-level feature tensor output by the first layer of the network, which belongs to the real-valued feature space. .parameter and The definitions are consistent with the input image, corresponding to the spatial height and width of the feature map, respectively. Parameters This represents the total number of basic feature channels defined in the network model. In the network hyperparameter configuration of this embodiment, it is used to specify the network's total number of channels. The default value of the constant is Operator This represents a standard two-dimensional convolution operation. This is to ensure the spatial resolution of the feature map. The stride of this 2D convolution operation is set to a constant, remaining unchanged before and after the convolution calculation. Zero padding at the edges is set to a constant. .

[0015] To obtain feature representations with a wider spatial receptive field, covering haze distributions with global diffusion characteristics or large-scale snow particle morphology, the first multi-scale structural module... Layer (where parameters) To identify the index constant of the current downsampling scale level depth, its value range is defined as greater than... Positive integers, i.e. ,parameter This represents the total number of layers in the network structure, as shown in this embodiment. Set as The network employs a layer-by-layer downsampling processing structure. Specifically, after performing downsampling on the image data corresponding to the previous scale level, it then performs a two-dimensional convolution calculation with the same hyperparameter configuration to obtain the next level. Low-level feature tensors of the layer The formula for calculating their relationship is as follows: ; In the above formula, the variables Indicates the first The scale-level extraction outputs low-level feature data tensors. Operator This indicates a downsampling operation. In the specific underlying implementation of this embodiment, this downsampling operation does not use conventional max pooling, but is configured through a step size parameter. The strided convolution algorithm module is implemented, with its convolution kernel size set to [size missing]. Edge fill is set to .variable This indicates the initial input image matrix. Continuous The intermediate image data generated after subspace downsampling is represented. Based on the downsampling ratio attenuation rule, this variable... The height and width values ​​of the included feature planes are exactly equal to those of the initial input image. Corresponding spatial dimension parameters .

[0016] After obtaining the low-level feature tensors at each scale, the network model processes these multi-scale low-level features using the encoder module. The data flow processing logic within the encoder includes feature transformation from the spatial domain to the frequency domain and feature attention fusion calculation across scale levels. Define the first... After the input feature sequence of the layer encoder is fully processed by the pre-deployed two-branch Fourier transform module, its output deep frequency domain feature data tensor is identified as follows: To achieve strict spatial alignment and interactive computation of cross-scale features between adjacent downsampling levels, this feature data tensor... Input the downsampling operator module to obtain the spatially aligned feature tensor. The mathematical expression corresponding to this calculation process is: ; In the above formula, the variables Represents tensor The feature data object output after performing spatial downsampling computation. At this point, the tensor... Spatial height and width dimension parameters and the first Alignment of layer feature parameters.

[0017] Furthermore, the encoder utilizes an adaptive attention mechanism to aggregate and compute the aligned features from the current adjacent layers, generating a fused feature tensor. This feature fusion mechanism employs a scaled dot-product attention topology framework at the underlying level, consisting of a query matrix, a key matrix, and a value matrix. The corresponding feature fusion calculation formula is strictly defined as follows: ; In the above formula for calculating attention aggregation, the variables are... This represents the output tensor generated after fusing features from two adjacent network layers. (Operator) This represents the normalized exponential function computation operation performed along a specific tensor dimension, for the input vector domain. , its first The specific mathematical mapping output of each element is: This is used to convert the inner product similarity matrix into a matrix whose sum is a constant. The attention weight probability distribution matrix. Variables This represents the query feature matrix generated after linear mapping. (Variable) Key feature matrix The transpose matrix object is generated after performing the matrix transpose operation in linear algebra. Parameters This represents a scaling factor constant used to constrain the numerical upper limit of the inner product calculation to prevent the gradient vanishing phenomenon during backpropagation. In this embodiment, The specific numerical value is defined as the square root of the corresponding feature channel number parameter. (Symbol) Represents the standard matrix multiplication algebraic operation. Variable This represents the value feature matrix generated by the mapping.

[0018] The core variable matrix used to drive the above attention calculation model , as well as The data source is generated based on multi-scale feature extraction from the current layer and downsampled high-level features. The corresponding linear mapping projection calculation formula is defined as: ; ; ; In the above set of matrix generation formulas, the variables Representative by the first Low-level feature tensors extracted by the layer module. Variables Representing the The feature tensor obtained by downsampling the output features of the layer encoder. (Variable matrix) Variable matrix Variable matrix These represent the predefined linear projection layer weight matrices, corresponding to the query, key, and value vector mappings, respectively, embedded in the network computation graph. In the basic operator implementation of this embodiment, the linear projection layer matrix operation is configured by the kernel size. The unbiased convolutional layer is executed. (Symbol) This represents the algebraic operation of matrix multiplication.

[0019] After completing the attention feature fusion calculation, the generated fusion feature tensor It is then directed to the two-branch Fourier transform module's computation node for frequency domain filtering and spatial structure reconstruction calculations targeting meteorological characteristics. The final output data object is defined as the first... Layer Encoding Feature Tensor .

[0020] After the encoder completes feature extraction from the forward computation graph, the network model calls the decoder module to perform feature reconstruction and resolution restoration on the encoded features. The computational logic goal of the decoder is to map the low spatial resolution deep semantic features back to the initial high-resolution image pixel spatial domain step by step, and to compute and output various result sets at multiple scales. During the data flow execution of the decoder, the first... After the feature data of the layer decoder undergoes depth processing by its corresponding two-branch Fourier transform module, the output data object is represented as follows: To achieve progressive scaling and restoration of spatial resolution, a spatial upsampling operation is performed on the feature data tensor to obtain the amplified upsampled feature tensor. The calculation formula is expressed as follows: ; In the above formula, the variables Represents the original tensor The target feature tensor generated after performing the upsampling calculation step. Operator This represents an upsampling operation used to improve the dimensionality resolution of the feature map. In the specific algorithm configuration of this embodiment, this upsampling operation is performed by the Transposed Convolution function, whose kernel parameter configuration includes: the convolution kernel size is set to... The sliding step size is set to The edge fill amount is set to This operation has the capability to precisely expand the spatial height and width dimensions of the input tensor to the input size. The mathematical properties of multiples. The obtained amplified feature tensor. The data is then input again to the next-level associated two-branch Fourier transform module node to further eliminate high-frequency artifacts that may be introduced in the upsampling difference calculation and remove residual meteorological feature data. After processing, the result is the [number of]th [item / module]. Layer decoding output feature tensor .

[0021] Based on the feature tensor data output from each level of the decoder, the multi-scale Fourier network model employs a residual addition mechanism to generate and output prediction result sets at multiple scales. This is specifically for the first level, which is the highest resolution base scale in the network structure. The layer output (i.e., the layer corresponding to the resolution of the initial original image), and its corresponding output image matrix. The calculation steps are as follows: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would The feature tensor output by the layer decoder The image is projected into the standard image color channel space through a predefined feature reduction convolution operation, and then compared with the initial degraded image data matrix initially fed into the network. Perform pixel-by-pixel summation between tensors. The calculation formula for this logic is as follows: ; In the above calculation relationship, variables The network model is represented by the first... The final multi-scale output image matrix. Operator Indicates the core size configuration as The feature reduction dimensionality of two-dimensional convolution operation is to reduce the number of channels from a high-dimensional constant. Direct mapping compression to output dimension constant (Strictly corresponding to RGB channels) to meet the data dimension consistency constraint of residual addition. Variables Representing the The layer decoder node computes the deep feature tensor output in the forward pass. Variables The corresponding input number A degraded image data matrix. Arithmetic symbols. This indicates that an element-wise addition algebraic operation is performed on a tensor with two identical dimensions, adding the spatial coordinates to the corresponding channels pixel by pixel.

[0022] Following the same residual network design criteria, for other depth downsampling levels in the network model, i.e., the corresponding 1st... Layer (where the constant constraint is) Output matrix The calculation formula for its prediction is strictly defined as: ; In this definition formula, variables Indicates the network number Hierarchical downsampling multi-scale output image matrix. Variables Indicates the first The feature tensor is generated by the layer decoder after processing by its internal two-branch Fourier transform module. (Variables) This indicates the preprocessing stage for the initial input image matrix. implement The image data matrix generated after subspatial downsampling operation.

[0023] By performing the residual matrix summation process described above, the model constructs an output data set containing multiple spatial resolution scales. In the final calculation stage, the system executes selection logic to precisely extract the first set of output results from the multi-level scales whose spatial resolution is completely equal to the initial degraded image resolution. Hierarchical Result Matrix It defines and outputs the target clean image matrix after the severe weather removal operation is completed, which is used to provide processing source data to downstream instance segmentation modules or various defect recognition components.

[0024] The dual-branch Fourier transform module (DFTB), repeatedly invoked in this embodiment, is the core processing unit for performing frequency domain filtering of meteorological interference characteristics and detailed reconstruction of airspace power facilities. For example... Figure 2 As shown in (a), the Two-Branch Fourier Transform (DFTB) module adopts a residual structure. The input features are batch normalized (BN) or layer normalized before entering the Two-Branch Fourier Transform (DFTN) network, and then the residuals are added to the input features to obtain the output. When the module receives the operation trigger command, it strictly follows the calculation sequence including spatial domain feature mapping, frequency domain decomposition, independent filtering, and inverse reconstruction to perform the following processing steps.

[0025] First, define the input feature tensor parameters sent to the computation node port of this module as follows: To adjust the probability distribution of the feature tensor to the optimal range for network learning and eliminate covariate bias between channels, the system adjusts the input feature tensor... Layer normalization is performed. The normalized data vector sequence then passes sequentially through a set of specifically configured convolutional and activation operators along pre-defined channels in the network, forming a serial feature extraction pipeline. This pipeline's specific processing sequence includes: performing pixel-level convolution operations (i.e., kernel size...). The convolution is used for channel-dimensional mapping expansion; subsequently, a kernel with a more concentrated parameter size is performed. Depth-wise convolution operations are used to extract local spatial responses; data is processed through a system with smooth derivative properties. The (GaussianError Linear Unit) activation function layer introduces nonlinear computation capabilities; this is followed by another round of execution with consistent hyperparameter configuration. Depthwise separable convolution operation; again through An activation function layer is applied; at the end of the sequence, another pixel-level convolution operation is used to compress the number of channels in the feature tensor to the set target dimension. After the above rigorously defined spatial operator sequence coherent processing, the final output is a spatial basic feature tensor containing spatial neighborhood context information. .

[0026] like Figure 2 As shown in (b) above, the spatial fundamental feature tensor is obtained. Subsequently, in order to completely remove weather degradation interference information such as global scattering haze, large-area snow particles, and dense rain lines from the global receptive field, the module will... The discrete two-dimensional spatial image coordinate system data is completely transformed and projected into the frequency domain coordinate system through a 2D Fast Fourier Transform (2D-FFT) algorithm node. Mathematically, the two-dimensional spatial feature channel function input to the frequency domain transformation node is defined as... Its matrix space size is The formula for the two-dimensional discrete Fourier transform is: In this transformation formula, the parameters Represents the frequency domain complex characteristic matrix generated after the transformation; parameter With parameters These represent the horizontal and vertical frequency indices in the frequency domain coordinate system, respectively, and their value ranges are constrained by the set of constants. and ; parameters With parameters These correspond to the horizontal and vertical indices in the spatial coordinate system, respectively; variables Characterizing spatial location points The characteristic intensity value at the location; constant The basic imaginary unit in mathematical definition, satisfying the equation ;constant Pi; a constant is the base of the natural logarithm.

[0027] The complex characteristic matrix obtained by the above frequency domain transformation Internally contains a matrix with a real part And imaginary part matrix The system algorithm executes a decoupling mechanism, utilizing formulas. With formula Calculations were performed to derive the amplitude component matrix that characterizes the macroscopic illumination and overall periodic intensity distribution of the image. And the phase component matrix that characterizes the high-frequency edge texture and structural location information of the image. Among them, symbols This represents the square root operation in mathematics, and the operator is... This indicates the operation of calculating the arctangent trigonometric function.

[0028] To address the different interference mechanisms that weather degradation phenomena produce on the amplitude and phase spectra, the method of this invention separately applies the amplitude component matrix... With phase component matrix Establish independent channels and perform customized separation and noise reduction processing.

[0029] For the amplitude data processing branch, its core objective is to suppress specific high-energy frequency band activation anomalies that characterize haze or meteorological noise in the amplitude matrix. The amplitude component matrix... First, spatial information compression and statistics are performed using a global average pooling layer operator, and then input into the Squeeze-and-Excitation Network (SENet) computation component. Internally, this SENet component runs a multilayer perceptron computation process, adaptively learning to extract the nonlinear dependencies between channels, generating corresponding dynamic weight coefficients for the amplitude response of each channel, and outputting a recalibrated and corrected amplitude feature weight matrix after computation. Furthermore, the magnitude component matrix of the original input branch is then... The calculated magnitude feature weight matrix By performing element-wise matrix multiplication, the final output feature tensor of the magnitude branch is calculated and derived. Its corresponding mathematical expression is: ; In the above expression, the variable This represents the output feature tensor generated after the amplitude branch completes the filtering process. (Variable) This represents the initial magnitude component matrix object obtained directly through a two-dimensional Fast Fourier Transform operation. (Variable) This represents the magnitude feature weight matrix generated by the SENet network. (Symbol) Here, it strictly refers to the pixel-wise matrix multiplication (Hadamard Product) algebraic operation between tensors of the same dimension.

[0030] For the parallel phase data processing branch, the core computational constraint lies in minimizing the distortion of phase information representing critical rigid structures such as power lines and insulators during the severe weather removal filtering process. The phase component matrix... The data is sequentially passed through a chain-like network sequence composed of specific deep learning operators. Specifically, the sequential data flow passes through: a standard two-dimensional convolutional layer operator, an instance normalization layer operator to eliminate contrast fluctuations in individual sample channels, and a layer that introduces non-linear computational properties. The data stream consists of a (RectifiedLinearUnit) activation function layer operator, a second convolutional layer operator responsible for feature mapping, and a final second instance normalization layer operator. After the data stream is processed through the above sequence nodes, the output generates a purified and corrected phase feature matrix. This branch network employs a residual output structure with skip connections, specifically outputting feature tensors. The formula for generating it is: ; In this output formula, the variables This represents the output feature tensor generated at the end of the phase processing branch. (Variable) Refers to the initial primitive phase component matrix generated by the two-dimensional fast Fourier transform. Variable This represents the phase-restored feature matrix generated after processing by the sequence operator chain. (Symbol) Explicitly represent the pixel-by-pixel algebraic computation operation corresponding to the tensor.

[0031] After the two independent branches complete the information purification and separation process in the frequency domain, the network architecture initiates the inverse mapping sequence, using inverse Fourier transform to restore the data to the pixel-level spatial coordinate system. The specific transformation and restoration process is as follows: the system reads the feature tensor output by the amplitude branch. and the phase branch output feature tensor Using the fundamental modulus and angular parameters for the inverse construction of the complex matrix domain, a two-dimensional inverse fast Fourier transform (2D-IFFT) algorithm is executed to obtain the feature tensor of the spatial residual result generated by the mapping and projection after frequency domain feature processing. After obtaining the spatial residual features, the spatial output feature tensor of the complete single-cycle computation of the two-branch Fourier transform module is calculated using a cross-layer skip connection mechanism. The characteristic integration arithmetic formula followed in this step is strictly defined as follows: ; In this computational relationship, the parameter This represents the spatial output feature tensor generated by the algorithm module after completing a single logical iteration. Parameter This refers to the spatial fundamental feature tensor initially input to the front end of this two-branch Fourier transform frequency domain processing module. (Parameter) This represents the feature tensor of the residual result after frequency domain processing operations are completed and the nodes are restored to the physical spatial domain using the inverse Fourier transform algorithm. (Operator) This refers to the tensor pixel-by-pixel addition calculation instruction. To ensure that degradation interference can be completely filtered out from deep feature manifolds, the network controller will automatically schedule and repeat the above-mentioned complete processing steps, including spatial domain projection to frequency domain, dual-branch parallel extraction, and inverse transformation back to spatial domain, several times. In the deterministic hyperparameter configuration of this embodiment, the number of loop executions is explicitly set to... When the set iteration count constant boundary is reached, the system outputs the final iterative feature tensor generated by the two-branch Fourier transform module, and records it as a constant identifier. This feature tensor It will be used as the feedforward input parameter for subsequent network modules.

[0032] Before applying multi-scale Fourier network models to inference prediction on real-world edge computing devices, a closed-loop parameter pre-training step must be performed on a hardware platform equipped with tensor computing units (GPUs / TPUs). The multiple core steps of the pre-training model and their corresponding mathematical processes are detailed below: First, the system executes the dataset construction and loading module. It integrates publicly available database samples and field-acquired images to construct a massive sample set consisting of power transmission line image data entities under foggy, rainy, and snowy conditions. During each independent iteration of the training process, the data loader extracts a mini-batch of data entities from the constructed sample set using a uniformly distributed random sampling mechanism. This data is then fed into the network model's in-memory processing queue. Within this extracted dataset, the variable parameters... This represents the single-frame degraded image data matrix to be processed and analyzed by the model; variable parameters Refers to the non-degradable, clean, real-world image label matrix that has an absolute physical scene correspondence with the degraded image; parameters This represents the batch size defined by the system hyperparameter configuration, i.e., the total number of independent sample entities included in a single forward propagation computation. In this embodiment, its value is set to [value missing]. .

[0033] Based on the multi-scale prediction output matrix generated after the network forward propagation computation and the stored and loaded clean image data matrix and ground truth labels, the computation engine executes a mathematical function to generate a constant value for the loss function. After obtaining the loss metric, the automatic differentiation module calculates the gradient vector of the loss with respect to each model weight, and uses the standard gradient descent method optimizer (in this embodiment, the AdamW optimization algorithm engine is specifically called) to perform backpropagation training and update the model's global learnable parameter matrix.

[0034] Define the total loss function for calculating and measuring the model's prediction error. The mathematical formula structure is a weighted summation polynomial of discrete error components calculated from output results at different spatial scales. The formula structure is set as follows: ; Among them, the first one is for the highest spatial resolution of the benchmark (i.e., the original input size). The absolute bias loss metric corresponding to the prediction results of the layer network The specific scalar calculation formula is as follows: ; For other network prediction architectures that include downsampling resolution output, ... The results at the scale level correspond to the loss metric. The formula for calculating its deviation assessment has been confirmed as follows: ; In the presented system of formulas that include global supervision loss, parameters are defined. Explicitly represent the scalar value of the total loss function that monitors the entire multi-scale computational training process of the network. Define parameters. This indicates that the evaluation network is in the uncompressed first stage. The loss deviation between the model's predicted output and the label at the original image scale. Sequence parameters. These represent the validation loss bias calculated independently for each downsampling scale level corresponding to other decreasing resolutions. Constant parameters The calibration system sets the total number of downsampling layers and depth scale. Configuration parameter sequence. This explicitly states that during the system's weighted summation operation, a set of real-valued weight coefficients are assigned to the specific loss at each scale level. In the configuration specifications set in this embodiment, each value exhibits a proportional decay characteristic based on the decreasing resolution. (Action symbols) This represents all items within a single batch loading queue in the mathematical rules. Perform a cumulative summation algebraic operation on each sample entity one by one. Parameter The value of this parameter directly corresponds to the total number of image samples loaded in a single batch training queue. This is a sequence index flag representing the corresponding individual image sample currently being accessed by the system. (Variable) Indicates the end of the model at the 1st Output the image matrix derived from the prediction calculation within the layer-scale channel. Variables Indicates that the system loads internally, and is related to the first A baseline of a clean, undegraded image data label matrix directly associated with each input degraded sample entity. (Symbol) It is used to define and calculate the distance between two specific equal-dimensional matrices. The norm (i.e., Manhattan distance vector calculation) instruction code operation, its underlying execution logic is to calculate the sum of the absolute values ​​of the scalar differences at corresponding coordinates after subtracting two matrices. Variables The reference model framework in the first Single-branch validation loss parameters for the scaling downsampling channel. Variables Refers to the model node at the th The matrix of multi-scale down-resolution output image data generated by inference within the scale channel. Variables The pre-evaluation module pre-processes a clean image label matrix of the baseline target. Execute the settings The corresponding target verification data matrix generated after the bilinear interpolation downsampling algebraic operation.

[0035] To comprehensively and objectively quantify and verify the execution effect and performance parameters of the integrated severe weather removal method described in this invention embodiment, the system was subjected to comparative measurement and verification experiments in multiple specially designed benchmark dataset environments. The quantitative measurement parameters for testing and evaluation include the calculated peak signal-to-noise ratio (PSNR, in dB, derived by dividing the square of the maximum pixel value by the mean square error and then taking the logarithm to base 10; the larger the value, the lower the image restoration distortion) and the structural similarity index (SSIM) which measures the similarity between the predicted image and the benchmark label image in terms of brightness features, contrast features, and structural covariance characteristics; its value range is strictly between [0,1], and a value close to 1 indicates a higher degree of preservation of structural texture information. To ensure the relative validity of the evaluation, the list of known technical algorithm components selected for isoparametric deployment and used as a reference benchmark includes: SwinIR, Uformer, Restormer, CAT, Stoformer, AirNet, Dehazeformer, DRSformer, LMQFormer, ShuffleFormer, CODE, ART, GRL, promptIR, AST, VmambaIR, MambaIRv2, and DSwinIR.The following are the calculation output results based on the set script calls and statistics of various relevant datasets: In the HazeCPLID data verification test task targeting defogging calculation capabilities: SwinIR results are PSNR 26.11, SSIM 0.9527; Uformer results are PSNR 26.04, SSIM 0.9514; Restormer results are PSNR 27.25, SSIM 0.9641; CAT results are PSNR 26.33, SSIM 0.9537; Stoformer results are PSNR 25.44, SSIM 0.9555; AirNet results are PSNR 26.09, SSIM 0.9510; Dehazeformer results are PSNR 26.00, SSIM 0.9611; DRSformer results are PSNR 20.89, SSIM 0. .8218; LMQFormer results: PSNR 25.67, SSIM 0.9519; ShuffleFormer results: PSNR 26.11, SSIM 0.9633; CODE results: PSNR 25.27, SSIM 0.9529; ART results: PSNR 26.32, SSIM 0.9302; GRL results: PSNR 24.15, SSIM 0.9252; promptIR results: PSNR 26.56, SSIM 0.9637; AST results: PSNR 20.67, SSIM 0.8295; VmambaIR results: PSNR 27.84, SSIM 0.9585; MambaIRv2 results: PSNR 26.93, SSIM 0.9597; DSwinIR results: PSNR 27.50, SSIM 0.9615. The model results of the multi-scale Fourier network correspondence method in this embodiment are PSNR 28.15 and SSIM 0.9714.

[0036] When deploying the validation task environment using the HazeTTPLA subset dataset containing target instance occlusion conditions: SwinIR results are PSNR 20.21, SSIM 0.8520; Uformer results are PSNR 20.23, SSIM 0.8575; Restormer results are PSNR 21.64, SSIM 0.8460; CAT results are PSNR 22.19, SSIM 0.8807; Stoformer results are PSNR 20.92, SSIM 0.9010; AirNet results are PSNR 20.20, SSIM 0.8830; Dehazeformer results are PSNR 23.35, SSIM 0.9133; DRSformer results are PSNR 19.01, SSIM 0.7919; LMQFormer results are PSNR 21.74, SSIM 0.8813; ShuffleFormer results are also high. The results are as follows: PSNR 21.92, SSIM 0.9032; CODE results: PSNR 19.84, SSIM 0.8012; ART results: PSNR 22.33, SSIM 0.8557; GRL results: PSNR 19.24, SSIM 0.8024; promptIR results: PSNR 21.71, SSIM 0.8681; AST results: PSNR 18.81, SSIM 0.8124; VmambaIR results: PSNR 22.11, SSIM 0.9099; MambaIRv2 results: PSNR 20.77, SSIM 0.8479; DSwinIR results: PSNR 21.72, SSIM 0.8646. The multi-scale Fourier network correspondence method model results in this embodiment are PSNR 22.81, SSIM 0.9173.

[0037] During environmental metric validation on the fine-classification dataset RainCPLID-L, which simulates light precipitation disturbance: SwinIR results were PSNR 28.38, SSIM 0.9307; Uformer results were PSNR 27.64, SSIM 0.9295; Restormer results were PSNR 31.36, SSIM 0.9584; CAT results were PSNR 31.30, SSIM 0.9418; Stoformer results were PSNR 29.67, SSIM 0.9268; AirNet results were PSNR 30.26, SSIM 0.9445; Dehazeformer results were PSNR 31.61, SSIM 0.9541; DRSformer results were PSNR 23.74, SSIM 0.9231; LMQFormer results were PSNR 29.81, SSIM 0.9316; ShuffleFormer results were also satisfactory. The results are as follows: PSNR 30.25, SSIM 0.9391; CODE: PSNR 29.38, SSIM 0.9356; ART: PSNR 31.22, SSIM 0.9360; GRL: PSNR 27.69, SSIM 0.9137; promptIR: PSNR 33.44, SSIM 0.9709; AST: PSNR 28.52, SSIM 0.9038; VmambaIR: PSNR 27.83, SSIM 0.9034; MambaIRv2: PSNR 26.66, SSIM 0.8663; DSwinIR: PSNR 27.54, SSIM 0.8784. The multi-scale Fourier network correspondence method model in this embodiment has a PSNR of 33.41 and an SSIM of 0.9719.

[0038] In the RainCPLID-H test validation sequence for environmental datasets facing stronger precipitation interference: SwinIR results were PSNR 27.19, SSIM 0.9031; Uformer results were PSNR 26.46, SSIM 0.9088; Restormer results were PSNR 30.41, SSIM 0.9499; CAT results were PSNR 29.45, SSIM 0.9312; Stoformer results were PSNR 26.99, SSIM 0.8844; AirNet results were PSNR 28.90, SSIM 0.9270; Dehazeformer results were PSNR 30.23, SSIM 0.9393; DRSformer results were PSNR 23.60, SSIM 0.8965; LMQFormer results were PSNR 28.01, SSIM 0.9178; ShuffleFormer results were also satisfactory. The results are as follows: PSNR 29.38, SSIM 0.9180; CODE: PSNR 28.90, SSIM 0.9323; ART: PSNR 29.61, SSIM 0.9222; GRL: PSNR 27.39, SSIM 0.8990; promptIR: PSNR 31.97, SSIM 0.9654; AST: PSNR 25.49, SSIM 0.8746; VmambaIR: PSNR 27.19, SSIM 0.8975; MambaIRv2: PSNR 26.36, SSIM 0.8706; DSwinIR: PSNR 26.85, SSIM 0.8939. The multi-scale Fourier network correspondence method model in this embodiment has a PSNR of 32.01 and an SSIM of 0.9661.

[0039] The SnowTTPLA-S fine-classification dataset was introduced to perform comparative validation of the ability to remove features of small-morphological snow particles within images: SwinIR results were PSNR 29.59, SSIM 0.9275; Uformer results were PSNR 35.26, SSIM 0.9637; Restormer results were PSNR 33.55, SSIM 0.9648; CAT results were PSNR 33.23, SSIM 0.9659; Stoformer results were PSNR 31.33, SSIM 0.9538; AirNet results were PSNR 33.98, SSIM 0.9647; Dehazeformer results were PSNR 34.67, SSIM 0.9632; DRSformer results were PSNR 35.36, SSIM 0.9607; LMQFormer results were PSNR 34.52, SSIM 0.9632. The results for ShuffleFormer are: PSNR 33.49, SSIM 0.9434; CODE: PSNR 32.34, SSIM 0.9557; ART: PSNR 33.63, SSIM 0.9625; GRL: PSNR 31.43, SSIM 0.9493; promptIR: PSNR 33.92, SSIM 0.9686; AST: PSNR 34.80, SSIM 0.9702; VmambaIR: PSNR 31.75, SSIM 0.9634; MambaIRv2: PSNR 34.06, SSIM 0.9697; DSwinIR: PSNR 32.36, SSIM 0.9639. The model results of the multi-scale Fourier network correspondence method in this embodiment are PSNR 35.50 and SSIM 0.9696.

[0040] When performing image computation restoration tests and result metric capture extraction on the SnowTTPLA-M subset containing intermediate occlusion levels, the results are as follows: SwinIR: PSNR 29.42, SSIM 0.9308; Uformer: PSNR 34.91, SSIM 0.9653; Restormer: PSNR 34.86, SSIM 0.9720; CAT: PSNR 33.21, SSIM 0.9628; Stoformer: PSNR 31.34, SSIM 0.9546; AirNet: PSNR 33.42, SSIM 0.9624; Dehazeformer: PSNR 33.99, SSIM 0.9648; DRSformer: PSNR 34.87, SSIM 0.9518; LMQFormer: PSNR 33.84, SSIM 0.9628. The results for ShuffleFormer are: PSNR 33.41, SSIM 0.9498; CODE: PSNR 32.03, SSIM 0.9540; ART: PSNR 33.33, SSIM 0.9591; GRL: PSNR 31.02, SSIM 0.9463; promptIR: PSNR 33.40, SSIM 0.9704; AST: PSNR 34.05, SSIM 0.9621; VmambaIR: PSNR 31.70, SSIM 0.9619; MambaIRv2: PSNR 32.90, SSIM 0.9604; DSwinIR: PSNR 32.27, SSIM 0.9580. The model results of the multi-scale Fourier network correspondence method in this embodiment are PSNR 34.89 and SSIM 0.9710.

[0041] When evaluating the results of SnowTTPLA-L in a test deployment with a feature subset containing a high degree of visual information attenuation due to large-scale physical occlusion: SwinIR results were PSNR 24.36, SSIM 0.8895; Uformer results were PSNR 27.21, SSIM 0.9214; Restormer results were PSNR 29.90, SSIM 0.9433; CAT results were PSNR 27.97, SSIM 0.9236; Stoformer results were PSNR 25.39, SSIM 0.8996; AirNet results were PSNR 28.30, SSIM 0.9300; Dehazeformer results were PSNR 28.75, SSIM 0.9248; DRSformer results were PSNR 27.86, SSIM 0.9221; LMQFormer results were PSNR 27.78, SSIM 0.9221. The results for ShuffleFormer are: PSNR 26.86, SSIM 0.9101; CODE: PSNR 27.81, SSIM 0.9272; ART: PSNR 27.95, SSIM 0.9249; GRL: PSNR 25.61, SSIM 0.9009; promptIR: PSNR 29.91, SSIM 0.9428; AST: PSNR 26.28, SSIM 0.9018; VmambaIR: PSNR 28.03, SSIM 0.9245; MambaIRv2: PSNR 27.62, SSIM 0.9202; DSwinIR: PSNR 28.44, SSIM 0.9310. The model results of the multi-scale Fourier network correspondence method in this embodiment are PSNR 31.02 and SSIM 0.9528.

[0042] To further quantify and calculate the image data vector matrix output by the preprocessed model of the multi-scale Fourier network corresponding method, and to assess the impact of the relevant feature transmission accuracy values ​​on the data feature distribution when inputting to the functional operator module components deployed in the downstream detection and recognition stage, the system adopted a benchmark testing framework algorithm for relevant functions, which is a standard component in academic and engineering application evaluation and analysis frameworks. Specifically, it used a general network architecture library (specifically, the YOLACT series feature object segmentation verification algorithm entity execution architecture library object module configuration group based on a unified framework) to perform serial interactive call-style data fusion testing. The benchmark configuration standard system for this functional stage test measurement was uniformly set to calculate, compare, and determine the Intersection over Union (IoU) ratio for different calculation interfaces, as well as the judgment constant, constraint threshold, constraint condition set configuration state, record extraction, target prediction data structure determination, relevant annotation execution standard box feature range overlap detection, and average accuracy calculation of various items. The judgment item record item index parameter (Average Precision, abbreviated character item label is the corresponding parameter notation code identifier referring to the letter AP) was also considered. The test and evaluation system structure for this phase is defined as follows: It will generate and return the target detection features corresponding to the bounding box mapping to the system memory stack at the algorithm's output, which will be used for target mapping boundary data frame selection, selection of coordinates, matrix structure (Bounding Box), geometric region range parameters, and calculated image space specific corresponding feature coordinates, corresponding recognition pixel groups, entity feature pixel sets, detection mask identifiers, associated object region parameter array set (Instance Mask) structure. It will then generate and export the corresponding average precision calculation, statistical parameter record items, and numerical results. According to the calculation classification and calibration protocol, the corresponding description is as follows: Specifically, it refers to the generation of the corresponding related parameter data associated with the application calculation boundary recognition, corresponding range, benchmark feature box selection coordinate judgment, execution, average precision, specific item abbreviation, mark definition, record form, corresponding symbol definition, and identifier as parameter code and variable name. ; Separately assigned to define the expression based on the mapping mask matrix overlap algorithm, extracting the corresponding average precision record item, generating abbreviated parameter item marker symbol code, and confirming it as the name variable code. The accuracy measurement system's indicator configuration is based on defined selection of specific feature objects, testing intersection-union comparison, comparison constants, overlap judgment thresholds, and covers the established parameters. The quantitative constant values ​​are defined as follows: Conditions and constant value setting parameters condition level is determined as follows Condition definition: two types of execution operations are associated, evaluated, overlap judged, extracted, calculated, and the calculation branches are set to set execution limit threshold conditions.

[0043] The system's internal data interaction defines the measurement generation logic, which derives and calculates four discrete, independent corresponding classification items. Features are extracted to determine test parameters and evaluate project record variables. These variables are mapped and labeled using a symbolic subscript format, defining the record as a name-specific referential quantity. Record numerical values ​​and parameter definitions with specific names Extract parameters, associate them with specific test conditions, and define variable identifiers. Data generation and parameter definition, measurement, reference, association, extraction, variable set items The corresponding record is determined. After the system evaluation calculation unit collects and extracts the data index record of the four discrete distribution independent parameter accuracy calculation record item data index record extraction group numerical result set content group, the corresponding instruction module schedules the operator to perform relevant numerical statistical summary operations on the record, and implements the corresponding execution for the aforementioned four fine distribution classes to refine the obtained numerical set item extraction value record column element calculation, and performs the numerical mean algorithm operation, that is, the arithmetic in the standard numerical definition to perform mean calculation, value calculation and statistical analysis to obtain the final corresponding extracted statistical evaluation representative system comprehensive evaluation characterization ability score value calculation evaluation system feature parameter score numerical extraction result definition record variable, and defines the overall result mapping identifier name as parameter code name. .

[0044] When performing instance segmentation validation based on the HazeTTPLA dataset: using clear images for benchmarking, the results are APb50% 57.63, APm50% 42.26, APb75% 33.78, APm75% 21.74, and overall APavg 38.85. When testing directly with unprocessed, degraded images, the accuracy results are APb50% 53.98, APm50% 33.56, APb75% 30.47, APm75% 18.14, and overall APavg 34.03.

[0045] The instance segmentation accuracy after preprocessing using each model is as follows: SwinIR: APb50% 57.25, APm50% 40.87, APb75% 33.33, APm75% 20.79, overall APavg 38.06; Uformer: overall APavg 38.23; Restormer: overall APavg 36.51; CAT: overall APavg 37.31; Stoformer: overall APavg 37.94; AirNet: overall APavg 34.88; Dehazeformer: overall APavg 37.63; DRSformer: overall APavg 31.45; LMQFormer: overall APavg 32.63; ShuffleFormer: overall APavg 34.82; CODE: overall APavg 37.78; ART: overall APavg 37.78. The results are as follows: APavg 36.83; GRL overall APavg 36.27; promptIR overall APavg 35.84; AST overall APavg 33.83; VmambaIR overall APavg 37.03; MambaIRv2 overall APavg 35.99; DSwinIR overall APavg 38.00. The multi-scale Fourier network corresponding method model results in this embodiment are: APb50% 57.30, APm50% 42.78, APb75% 34.06, APm75% 21.65, and overall APavg 38.95.

[0046] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for integrated severe weather removal in aerial images of power lines based on multi-scale Fourier networks, characterized in that, include: Acquire aerial images of degraded power lines to be processed; The degraded image is input into a pre-trained multi-scale Fourier network model, wherein the multi-scale Fourier network model is constructed with a multi-scale structure module and a two-branch Fourier transform module as its core. The multi-scale low-level features of the degraded image are extracted using the multi-scale structure module. The encoder is used to process the multi-scale low-level features, and the dual-branch Fourier transform module is used to transform the spatial features to the frequency domain during the processing. The amplitude component and the phase component are separated and processed respectively, and then restored to the spatial domain by inverse Fourier transform to achieve frequency domain global feature calibration and local feature extraction, so as to obtain the coded features. The encoded features are reconstructed using a decoder, and the output results at multiple scales are calculated. The result corresponding to the original image resolution in the output results at multiple scales is selected as the target clean image after removing severe weather. The extraction of multi-scale low-level features of the degraded image through the multi-scale structure module specifically includes: for the first layer, the input image... pass Convolution yields the low-level features of the first layer. : ； ； in, Indicates the input number of the first... One degraded image data; This represents the low-level features extracted in the first layer; Indicates the kernel size as Convolution operations; The dimension is The real number feature space, , , The input images are respectively represented. Height, width, and number of channels; for the first Layer, in which The first one was obtained using the same method. Low-level features of the layer : ； in, Indicates the first Low-level features extracted from layers; Indicates a downsampling operation; Indicates the input image The downsampled image, and the The height and width are both the height and width of the input image. Corresponding size ; The process of using an encoder to process the multi-scale low-level features specifically includes: defining the first... The features of the layer encoder are output after passing through a two-branch Fourier transform module. After downsampling, the result is : ； in, Indicates the first Frequency domain characteristics of the layer encoder output; Indicates to Features obtained after downsampling; This indicates a downsampling operation; it fuses features from adjacent levels to obtain fused features. : ； in, This represents the features after merging adjacent levels; This indicates the normalization exponential function operation; Represents the query feature matrix; Key feature matrix The transpose of the matrix; This represents a scaling factor constant related to the feature dimension; This represents the matrix dot product operation; The fused features represent the feature matrix; After passing through the two-branch Fourier transform module, the first... Layer coding features .

2. The method for integrated severe weather removal of aerial power line images based on multi-scale Fourier networks according to claim 1, characterized in that, The query feature matrix Key feature matrix , value characteristic matrix The calculation formula is: ； ； ； in, Indicates the first Low-level features extracted from layers; Indicates the first The layer encoder outputs downsampled features; , , These represent the linear projection layer matrices corresponding to query, key, and value mappings, respectively. This represents the matrix multiplication operation.

3. The method for integrated severe weather removal of aerial power line images based on multi-scale Fourier networks according to claim 2, characterized in that, The feature reconstruction process using the decoder specifically includes: defining the first... The features of the layer decoder are then processed by a two-branch Fourier transform module to obtain the output. After upsampling, the result is : ； in, Indicates the first Features of the layer decoder output; Indicates to Features after upsampling; Indicates an upsampling operation; the After passing through the two-branch Fourier transform module, the first... Layer characteristics .

4. The method for integrated severe weather removal of aerial power line images based on multi-scale Fourier networks according to claim 3, characterized in that, The calculation yields output results at multiple scales, specifically including: calculating the first... Layer output : ； in, Indicates the first The layer's output image; Indicates the kernel size as Convolution operations; Indicates the first Features of the layer decoder output; Indicates the input number of the first... One degraded image data; This indicates a pixel-by-pixel addition operation; calculates the first pixel. Layer output ,in : ； in, Indicates the first The layer's output image; Indicates the first The features obtained by the layer decoder after passing through the double-branch Fourier transform module; Indicates the input image conduct The image after downsampling.

5. The method for integrated severe weather removal of aerial power line images based on multi-scale Fourier networks according to claim 4, characterized in that, The dual-branch Fourier transform module, when executed, includes the following processing steps: defining the input features as... For the input features Perform layer normalization, and then use pixel-level convolution respectively. Depthwise convolution, GELU activation layer, Spatial basic features are obtained through depthwise convolution, GELU activation layers, and pixel-level convolution. The aforementioned basic spatial characteristics The amplitude components are obtained by using a 2D Fast Fourier Transform. With phase components .

6. The method for integrated severe weather removal of aerial power line images based on multi-scale Fourier networks according to claim 5, characterized in that, For the amplitude component With phase components The specific process for separation is as follows: the amplitude component The result is obtained through average pooling and SENet excitation compression module. Output : ； in, This indicates the output characteristics of the amplitude branch; This represents the magnitude component matrix obtained from the two-dimensional fast Fourier transform; This represents the processed amplitude feature matrix; Represents a matrix multiplication operation; the phase component The result is obtained through processing with convolutional layers, instance normalization layers, ReLU activation layers, convolutional layers, and instance normalization layers. Output : ； in, This indicates the output characteristics of the phase branch; This represents the phase component matrix obtained from the two-dimensional fast Fourier transform; This represents the processed phase feature matrix; This indicates a pixel-by-pixel addition operation.

7. The method for integrated severe weather removal of aerial power line images based on multi-scale Fourier networks according to claim 6, characterized in that, The specific process of restoring the spatial domain through inverse Fourier transform is as follows: the output and The spatial domain results are obtained through inverse Fourier transform. Then, the single-iteration output of the two-branch Fourier transform module is calculated. : ； in, This represents the spatial output features of a single iteration; This represents the fundamental spatial characteristics input to the two-branch Fourier transform module; This represents the characteristics of the residual result recovered to the spatial domain through inverse transformation after frequency domain processing; This represents a pixel-by-pixel addition operation; repeating the above processing steps several times yields the final output of the dual-branch Fourier transform module. ,in This represents the final feature after several iterations.

8. The method for integrated severe weather removal of aerial power line images based on multi-scale Fourier networks according to claim 7, characterized in that, The training steps of the pre-trained multi-scale Fourier network model include: constructing a sample set consisting of images of transmission lines under fog, rain, and snow conditions, and extracting small batches of data from the sample set. The data is fed into the network model, where This represents the image data to be recovered. For the corresponding clean image data, Indicate the batch size; calculate the loss function based on the output results at multiple scales and the clean image data. The model is trained using gradient descent; the loss function is... The calculation formula is: ； ； ； in, This represents the total loss function used in network training. Indicates the first Loss at the layer scale; These represent the losses corresponding to each of the other downsampling scale levels. This indicates the total number of downsampling layers; This represents the weighting coefficients set for the loss at each scale level; Indicates the batch Summation of each sample; This represents the total number of samples within a single mini-batch; Indicates the sample index; The model number is represented by the first... The layer's output image; Indicates the first The true label of the clean image data corresponding to each sample; This indicates the operation of calculating the L1 norm; Indicates the first Layer loss; The model number is represented by the first... Multi-scale output images of layers; Indicates a truly clean image conduct The target result obtained after downsampling.

Images