Dynamic mixed attention-based radar image correction method for wind field inversion

Abstract

The invention provides a dynamic mixed attention-based radar image correction method for wind field inversion, and solves the technical problem that a large error still exists after a radar image obtained in a complex environment is corrected in the prior art. The method comprises the steps of obtaining a radar rain pollution image, and performing shallow feature extraction to obtain an initial feature map; inputting the initial feature map into a dynamic convolution module to obtain a dynamic convolution kernel, and carrying out convolution on the dynamic convolution kernel and the initial feature map to obtain an output feature map of the dynamic convolution module; inputting the output feature map of the dynamic convolution module into a parallel mixed attention module to obtain a target feature map, and performing local residual learning to obtain a final feature map; and reconstructing the final feature map and then carrying out global residual learning to obtain a corrected radar rain pollution image. The method can be widely applied to the technical field of meteorological monitoring.

Description

Technical Field

[0001] This application belongs to the technical field of meteorological monitoring. More specifically, it relates to a radar image correction method based on dynamic hybrid attention for wind field inversion. Background Art

[0002] X-band radars are widely used in meteorology, hydrology, or urban monitoring due to their high resolution and strong penetration ability. However, in complex environments such as rainy and foggy weather, radar signals are interfered with and need to be corrected specifically. Currently, there is no mature solution for the rain interference correction technology of X-band marine radars. The root cause is that under low incident angle conditions, when the X-band radar beam passes through the rainfall area and irradiates the sea surface, both rainfall and the sea surface will scatter the radar wave, resulting in a rain-stained image obtained by the radar, generating a radar rain-stained image. This scattering is complex and involves the interaction between rainfall and the sea surface, so the dominant characteristics have not been clearly defined, leading to theoretical blind spots in backscattering modeling and signal separation.

[0003] The existing Chinese invention patent with the application number 202410409120.0 proposes a method for inverting the sea surface wind field based on X-band radar images. When processing the obtained radar images, this method uses filters for denoising, correction, and normalization. This method can only simply remove some noise in the radar images, and often there will be large errors in the corrected radar images due to inaccurate feature extraction when facing complex environments. Therefore, there is an urgent need to provide a radar image correction method based on dynamic hybrid attention for wind field inversion to solve the above problems. Summary of the Invention

[0004] The purpose of the embodiments of this application is to provide a radar image correction method based on dynamic hybrid attention for wind field inversion to solve the technical problem that there are still large errors in the corrected radar images obtained in complex environments in the prior art.

[0005] To achieve the above purpose, the embodiments of this application provide a radar image correction method based on dynamic hybrid attention for wind field inversion, including the following steps: Obtain a radar rain-stained image and perform shallow feature extraction to obtain an initial feature map; Input the initial feature map into the dynamic convolution module to obtain a dynamic convolution kernel, and perform convolution with the initial feature map to obtain the output feature map of the dynamic convolution module; Input the output feature map of the dynamic convolution module into the parallel hybrid attention module to obtain a target feature map, and perform local residual learning to obtain the final feature map; Reconstruct the final feature map and perform global residual learning to obtain the corrected radar rain-stained image; The dynamic convolution module is used to generate dynamic weights by performing global average pooling, two fully connected processes, and activation function processing on the initial feature map. The dynamic weights are normalized by using the spline function of the KAN network to adjust hyperparameters dynamically, obtaining the weights of the dynamic convolution kernel, and weighted summing with several expert convolution kernels to obtain the dynamic convolution kernel.

[0006] Preferably, the parallel hybrid attention module is used to batch normalize the output feature map of the dynamic convolution module to obtain the batch-normalized feature map, input it into the channel attention module and the pixel attention module respectively to obtain the output feature map of the channel attention module and the output feature map of the pixel attention module, perform convolution to obtain the output feature map, perform two convolution processes and activation function processing, and then element-wise add it to the output feature map of the dynamic convolution module to obtain the target feature map.

[0007] Preferably, the dynamic convolution module includes a global average pooling layer, two fully connected layers, a ReLU layer, a KAN network, and several expert convolution kernels; The KAN network adaptively fits and calculates the complex relationship of features in the dynamic weights by changing the shape of the spline in the spline function; Each grid point of the spline function is a spline basis function, and the hyperparameters are dynamically adjusted according to the density, shape, and smoothness of the grid to normalize the dynamic weights.

[0008] Preferably, after the global average pooling layer performs global average pooling on the initial feature map, it inputs the first fully connected layer for transformation to obtain a feature vector, inputs the ReLU layer for activation, and then inputs the second fully connected layer for transformation to generate dynamic weights.

[0009] Preferably, the parallel hybrid attention module includes a batch normalization layer, parallel channel attention module and pixel attention module, two convolutional layers, and a GELU layer; The channel attention module is used to obtain channel weights based on the batch-normalized feature map and multiply them element-wise with the batch-normalized feature map to obtain the output feature map of the channel attention module; The formula for obtaining the output feature map of the channel attention module includes: ; ; ; In the formula, is the channel feature, is the c-th channel at the value of , is the global pooling function, p is the pooling type-related parameter in the global pooling operation, here it is the identifier representing the average pooling operation mode, is the feature map after batch normalization, H is the height of the feature map after batch normalization, W is the width of the feature map after batch normalization, i is the coordinate index of the feature map after batch normalization in the height direction, and j is the coordinate index of the feature map after batch normalization in the width direction. is the Sigmod function. is the ReLU function. is the channel weight. is the convolutional layer. is the output feature map of the channel attention module.

[0010] Preferably, the pixel attention module obtains pixel weights according to the feature map after batch normalization, and multiplies the pixel weights element-wise with the feature map after batch normalization to obtain the output feature map of the pixel attention module. The formula for obtaining the output feature map of the pixel attention module includes: ; ; In the formula, PA p is the pixel weight, is the Sigmod function, is the ReLU function, is the feature map after batch normalization, is the convolutional layer, is the output feature map of the pixel attention module.

[0011] Preferably, performing local residual learning includes convolving the target feature map and adding it element-wise to the initial feature map to obtain the final feature map.

[0012] Preferably, reconstruction includes concatenating the obtained multiple final feature maps along the channel dimension to obtain the reconstructed feature map.

[0013] Preferably, global residual learning includes inputting the reconstructed feature map into the channel attention module and the pixel attention module for combination, performing two convolutions, and then adding it element-wise to the radar rain and sewage image to obtain the corrected radar rain and sewage image.

[0014] Preferably, the formula of the spline function includes: ; In the formula, is the spline function, x is the feature of the dynamic weight, C is the coefficient optimized during training, B is the kernel function, and i is the label of the sample point used during training.

[0015] The beneficial effects of this application are as follows: This application provides a radar image correction method for wind field inversion based on dynamic hybrid attention, and proposes a new neural network model - dynamic parallel hybrid attention network. The dynamic convolution module is used to extract features to obtain a dynamic convolution kernel. Among them, the activation function softmax in the traditional dynamic convolution module is replaced by a learnable KAN network, which can capture the complex features of radar rain and sewage images from the initial feature map. The spline function of the KAN network is used to dynamically adjust the hyperparameters to normalize the dynamic weights to obtain the weights of the dynamic convolution kernel, which can simulate complex functions with fewer parameters, enabling the dynamic convolution module combined with the KAN network to be dynamically adjusted according to the characteristics of radar rain and sewage images, improving the accuracy of function approximation and the expression ability of the model, greatly reducing the number of model parameters, and achieving high precision and low redundancy. Aiming at the non-uniformity of rain and sewage components in radar images and the high coupling between rain and sewage and the background, the parallel hybrid attention module is used to parallelly capture important information in radar rain and sewage images, providing multi-dimensional information, reducing the sensitivity of single attention to noise and outliers, reducing the bias in noise suppression and error correction, and improving the accuracy of the obtained target feature map. Finally, local residual learning and global residual learning are performed on the target feature map to fuse multi-scale global features and local features, alleviating the problem of detail blurring caused by information transmission loss. The residual learning enhances the gradient flow through the skip connection structure, alleviates the problem of gradient disappearance in deep networks, improves the training stability, and enables the dynamic parallel hybrid attention network model to effectively suppress the interference of rainfall or foggy weather on radar images. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] To more clearly illustrate the technical solutions in the embodiments of this application, the following will briefly introduce the drawings required for use in the embodiments or the description of the prior art. Obviously, the following drawings are only some embodiments of this application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0017] Figure 1 It is a flowchart of a radar image correction method for wind field inversion based on dynamic hybrid attention provided by an embodiment of this application; Figure 2 It is a schematic structural diagram of a dynamic hybrid attention neural network provided by an embodiment of this application; Figure 3 It is a schematic diagram of the usage process of the KAN network provided by an embodiment of this application; Figure 4 It is an uncorrected radar rain and sewage image provided by an embodiment of this application; Figure 5The corrected radar rain and sewage image provided by an embodiment of the present application; Figure 6 The intensity distribution histogram of the uncorrected radar rain and sewage image provided by an embodiment of the present application; Figure 7 The intensity distribution histogram of the corrected radar rain and sewage image provided by an embodiment of the present application. Detailed implementation manners

[0018] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present application clearer and more understandable, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0019] The present application proposes a new neural network model - Dynamic Parallel Mixed Attention Network (DPMA-Net). This neural network model obtains the target feature map through the dynamic convolution module and the parallel mixed attention module, and performs local residual learning and global residual learning on the target feature map, thereby correcting the radar rain and sewage image obtained by the radar in a complex environment.

[0020] The dynamic convolution module is used as the feature extraction part. The dynamic convolution module contains several expert convolution kernels, which can ensure that the weights are adjusted in real time according to the differences in the initial feature maps, so that after the model obtains the initial feature map, it can flexibly select the appropriate convolution kernel for feature extraction, improving the accuracy of feature extraction. Moreover, the setting of multiple expert convolution kernels can also capture features diversely, enabling it to adaptively adjust the weights to obtain more comprehensive and accurate features. The present application replaces the activation function softmax in the dynamic convolution module with the learnable KAN network and uses the spline function to replace the fixed weight parameters in the traditional neural network model. The model uses a learnable activation function, which can better capture the complex features of the input radar rain and sewage image, simulate complex functions with fewer parameters, and enable the dynamic convolution combined with the KAN network to be dynamically adjusted according to the features of the input radar rain and sewage image, improving the flexibility of the dynamic convolution module, and thus better meeting the requirements of subsequent processing.

[0021] A parallel hybrid attention module is set up to address issues such as the non-uniformity of rain and sewage components and the high coupling between rain and sewage and the background in radar rain and sewage images. Through the parallel connection of channel attention (CA) and pixel attention (PA), the model's ability to capture important information in the features of radar rain and sewage images is enhanced. By processing information in different dimensions in parallel, the problems of information loss and gradient coupling are reduced. Compared with the serial network, the parallel network can reduce the serial delay, lower the sensitivity of a single attention to noise and outliers, reduce the impact of an abnormality in a certain module on the correction result, reduce the bias in noise suppression and error correction, and improve the accuracy of the obtained target feature map.

[0022] The method combines local residual learning and global residual learning to fuse global and local features at multiple scales, alleviating the problem of detail blurring caused by information transmission loss in traditional methods. Through the skip connection structure, residual learning directly adds the input feature map to the layer output, directly transmits the input information to subsequent levels, allows the gradient to backpropagate, enhances the gradient flow, alleviates the vanishing gradient problem in deep networks, improves the training stability, and enables the model to effectively suppress the interference of rainfall or foggy weather on radar images.

[0023] Please refer to Figure 1 , a radar image correction method based on dynamic hybrid attention for wind field inversion provided by an embodiment of the present application, includes: S1: Perform shallow feature extraction on the radar rain and sewage image obtained by the radar to obtain an initial feature map.

[0024] Input the radar rain and sewage image obtained by the radar into the dynamic parallel hybrid attention network. The radar rain and sewage image is a vector matrix containing angles and distances and is presented as a pseudo-color map after processing. The radar rain and sewage image is subjected to shallow feature extraction through a 3×3 convolutional layer to obtain an initial feature map. The initial feature map includes low-level features of the captured radar rain and sewage image, such as details like edges and textures.

[0025] S2: Input the initial feature map into the dynamic convolution module to obtain a dynamic convolution kernel, and perform convolution with the initial feature map to obtain the output feature map of the dynamic convolution module.

[0026] Please refer to Figure 2, The dynamic parallel hybrid attention network includes several groups of structural parts. Each group of structural parts includes several blocks of structures. Each block of structure includes a dynamic convolution module and a parallel attention module, and the dynamic convolution module is used as the feature extraction part of the parallel attention module. The weights of each expert convolution kernel are dynamically calculated according to the features of the input initial feature map through a gating mechanism, and weighted summation is performed to obtain the final dynamic convolution kernel. In this way, the neural network can dynamically generate convolution kernel parameters according to different image regions and different time spans, so as to more flexibly adapt to the changes of the input initial feature map and capture more complex features. Among them, the dynamic convolution module includes a weight generation network (WGN) and dynamic filter generation.

[0027] Specifically, the initial feature map is input into the weight generation network of the dynamic convolution module. The initial feature map with dimensions of H×W×C is pooled through a global average pooling layer (GPA), and the initial feature map is transformed into a numerical value of length C, that is, a channel feature vector (C-dimensional), so as to effectively reduce model parameters, prevent overfitting, and improve the generalization ability of the model. Then, it sequentially passes through a first fully connected layer (FC1) with a height and width of 1×1, a ReLu layer, and a second fully connected layer (FC2), and through a learnable Kolmogorov–Arnold network (KAN network) to generate the weights of the dynamic convolution kernel.

[0028] Specifically, the initial feature map with the shape of H×W×C is input into the global average pooling layer, and global average pooling is performed independently for each channel to calculate the mean value, and a feature vector Z of length C is output, that is, the channel feature vector. The specific formula is as follows: ; In the formula, is the channel feature vector z of length C, H is the height of the initial feature map, W is the width of the initial feature map, is the feature value at the position of coordinates (i,j) in the c-th channel, where i corresponds to the height dimension index and j corresponds to the width dimension index.

[0029] Then, the channel feature vector is input into the first fully connected layer FC1 to calculate the hidden feature vector, that is, a feature vector s of length d. The specific formula is: ; In the formula, s is a D-dimensional feature vector, is a weight matrix of C×d dimensions, d is the dimension of the hidden layer, z is a feature vector of length z, b 1is the bias term.

[0030] To introduce non - linear transformation, the ReLU layer is used to activate the feature vector s of length d. The specific formula is as follows: ; In the formula, h is the activated feature vector, and s is the D - dimensional feature vector of length d.

[0031] The activated feature vector after the ReLU layer transformation is input into the second fully - connected layer FC2 to generate K dynamic weights. The specific formula is as follows: ; In the formula, is the weight matrix of K×d dimension, d is the dimension of the hidden layer, b 2 is the bias term, and a is the unnormalized dynamic weight.

[0032] To ensure that the sum of the dynamic weights of all dynamic convolution kernels is 1, the KAN network is used for normalization. The KAN network is used to replace the Softmax function in the dynamic convolution module.

[0033] Specifically, the KAN network can decompose any multivariate continuous function into several univariate continuous functions. Compared with the fixed activation function used in the traditional multi - layer perceptron (MLP), this application uses a trainable univariate function to replace the traditional weight parameters in the KAN network, fundamentally improving the accuracy of function approximation and the expressive ability of the model, and greatly reducing the number of model parameters. The shallow network structure of the KAN network in this application is as Figure 3 shown. Each edge in the figure corresponds to a learnable one - dimensional spline function, and the node only performs linear summation. The formula is as follows: ; In the formula, is the L - th spline function in the KAN network, L is the network level, and x is the feature of the dynamic weight.

[0034] This application uses the spline function in the KAN network to replace the fixed weight parameters in the traditional neural network model. This design allows the neural network model to adaptively fit and calculate the complex relationship of the features in the dynamic weight by adjusting the shape of the spline in the KAN network, reducing the approximation error and enhancing the ability to learn subtle patterns from high - dimensional data. Among them, the features in the dynamic weight include the initial feature map, the channel feature vector (C - dimensional) extracted by the global average pooling layer, and the weight distribution of the KAN network. The channel feature vector is used as the input for the KAN network to calculate the dynamic weight, and the weight distribution of the KAN network can be dynamically adjusted based on the input features and is used to control the combination of different expert convolution kernels.

[0035] During the training process, the coefficients of the spline functions in the KAN network are optimized, and these spline functions are constructed based on spline basis functions defined on a grid. Among them, the positions of the grid points in the grid determine the active intervals of each spline basis function, significantly affecting the shape and smoothness of the spline. Therefore, the density of the grid can be regarded as a hyperparameter affecting the accuracy of the neural network model. By increasing the number of grid points and aggregating the parameters, the neural network model can have higher control precision. This application does not limit the number of grid points, which can be adjusted according to specific application scenarios to achieve optimal performance. Among them, the general formula of the spline is: ; In the formula, is the spline function, x is the feature of the dynamic weight, C is the coefficient optimized during training, B is the kernel function, and i is the label of the sample point used during the training process.

[0036] Multiply the weights of the dynamic convolution kernel by several expert convolution kernels and then sum them to obtain the dynamic convolution kernel, and then convolve the dynamic convolution kernel with the initial feature map to obtain the output feature map of the dynamic convolution module. The specific formula is as follows: ; ; In the formula, is the weight of the expert convolution kernel, is the weight generation network, is the output feature map of the dynamic convolution module, is the dynamic convolution kernel, i is the convolution kernel number, is the expert convolution kernel.

[0037] S3: Input the output feature map of the dynamic convolution module into the parallel hybrid attention module to obtain the target feature map, and perform local residual learning to obtain the final feature map.

[0038] Specifically, perform batch normalization on the output feature map of the dynamic convolution module to obtain the batch-normalized feature map, and input it into the channel attention module and the pixel attention module respectively to obtain the output feature map of the channel attention module and the output feature map of the pixel attention module.

[0039] In the channel attention module, perform global average pooling on the batch-normalized feature map to generate channel features, perform convolution and activation to generate channel weights, and convolve the channel weights with the batch-normalized feature map to obtain the output feature map of the channel attention module.

[0040] In the pixel attention module, perform convolution and activation on the batch-normalized feature map to generate pixel weights, and convolve the pixel weights with the batch-normalized feature map to obtain the output feature map of the pixel attention module.

[0041] The output feature map of the dynamic convolution module is input into the parallel attention module. The parallel attention module of this application integrates a channel attention module and a pixel attention module to enhance the multi-dimensional capture ability of the neural network model. Moreover, in this application, the channel attention and pixel attention are set in parallel to ensure that the neural network can extract local information and global information simultaneously, enhancing the attention to the rain and sewage pollution areas in the rain and sewage images. Combining the above Group Architecture and parallel hybrid attention, it is possible to generate an adaptive weight using channel attention after concatenating the output feature maps of the dynamic convolution modules of each hierarchical group structure in the channel dimension, and dynamically adjust the contributions of features at various scales, where the multi-scale features include shallow features and deep features. Finally, the shallow features (including edges and textures) and deep features (including semantic information) are weighted and fused to avoid the blindness of simple addition and improve the structural integrity of the corrected rain and sewage pollution results.

[0042] Specifically, the output feature map of the dynamic convolution module is passed through a Batch Norm (BN) layer to accelerate training and improve model stability. The calculation formula is as follows: ; In the formula, is the mean of the current batch, is the variance of the current batch, is the learnable scaling parameter, is the learnable offset parameter, x is the output feature map of the dynamic convolution module, is the feature map after batch normalization, is a very small positive number to prevent the denominator from being zero.

[0043] After passing through the batch normalization layer, the batch-normalized feature map is input into the channel attention and pixel attention modules in parallel, Among them, the Channel Attention (CA) module enhances the responsiveness of different channels by learning the importance of different channels of the batch-normalized feature map. The channel attention module generates channel features through global average pooling , and then generates channel weights through convolution and activation. Among them, the formula for the channel feature is: ; In the formula, is the value of the c-th channel at , is the global pooling function, p is the parameter related to the pooling type in the global pooling operation, and here it is an identifier representing the operation mode of average pooling, is the feature map after batch normalization, H is the height of the feature map after batch normalization, W is the width of the feature map after batch normalization, i is the coordinate index of the feature map after batch normalization in the height direction, and j is the coordinate index of the feature map after batch normalization in the width direction.

[0044] In this way, the length and width information of the feature map after batch normalization is compressed, the channel information is extracted, and finally the channel weights are obtained. Subsequently, the channel features will pass through two convolutional layers and two activation functions sigmoid and ReLu to obtain the channel weights The formula is as follows: ; In the formula, is the Sigmod function, is the ReLU function, is the channel weight, is the convolutional layer.

[0045] Finally, the channel weights output by the channel attention module are multiplied element-wise with the feature map after batch normalization to obtain the output feature map of the channel attention module. The formula is as follows: ; In the formula, is the output feature map of the channel attention module, is the feature map after batch normalization, is the channel weight.

[0046] Specifically, the Pixel Attention (PA) module enhances the features of key regions by dynamically allocating pixel weights. Using the Sigmoid activation function and the ReLU function in combination with two convolutions, the weights of the feature map after batch normalization are restricted to [0,1] to obtain the pixel weights, and the pixel weights are multiplied element-wise with the feature map after batch normalization to obtain the output feature map of the pixel attention module.

[0047] The formula for obtaining the output feature map of the pixel attention module includes: ; ; In the formula, PA p is the pixel weight, is the Sigmod function, is the ReLU function, is the feature map after batch normalization, is the convolutional layer, is the output feature map of the pixel attention module.

[0048] The output feature map of the channel attention module and the output feature map of the pixel attention module are convolved to obtain an output feature map. After convolution and activation, it is element-wise added to the output feature map of the dynamic convolution module to obtain a target feature map, and local residual learning is performed to obtain the final feature map.

[0049] In an optional embodiment, the output feature map of the channel attention module and the output feature map of the pixel attention module are convolved to obtain an output feature map, which then passes through a 1×1 convolutional layer, a GELU function, and a 1×1 convolutional layer in sequence, and then is element-wise added to the output feature map of the dynamic convolution module to obtain the output of the block structure to obtain the target feature map. After continuously passing through 10 block structures, local residual learning is performed. Local residual learning includes convolving the target feature map and then element-wise adding it to the initial feature map to obtain the final feature map. In this application, there are 3 group structures, so it is necessary to continuously pass through 3 identical group structures to obtain the final feature map.

[0050] Specifically, the convolved feature map is convolved through a 1×1 convolutional layer. The kernel size of the 1×1 convolutional layer is 1×1. During the operation, the 1×1 convolutional layer performs a linear combination of all its channels for each pixel point of the output feature map, that is, matrix multiplication is performed on the channel vector corresponding to each output feature map pixel point. When the number of channels of the output feature map is Cin and the number of output channels is Cout, for each pixel point, the Cin-dimensional vector is transformed into a Cout-dimensional vector through the weight matrix. The feature map after the first 1×1 convolution is input into the GELU function for activation, and the formula is: ; In the formula, GELU() is the GELU function, and z is the feature map after the first 1×1 convolution.

[0051] Finally, the activated feature map is convolved through a 1×1 convolutional layer again to obtain the final feature map. This application uses a 1×1 convolutional layer to adjust the number of channels and perform feature fusion on the convolved feature map, then introduces non-linearity through the GELU function, and finally uses another 1×1 convolutional layer to further adjust the number of channels of the feature map and perform feature transformation. This combined operation can flexibly adjust the representation form of the feature map, increase the non-linear expression ability of the model, help learn more complex patterns, and improve the performance of the model in various tasks.

[0052] S4: The final feature map is reconstructed and then global residual learning is performed to obtain the corrected radar rain and sewage image.

[0053] The final feature maps output by each group structure are reconstructed to obtain the reconstructed feature maps. Among them, the reconstruction includes splicing the obtained multiple final feature maps along the channel dimension to obtain the reconstructed feature maps, so as to achieve the effective fusion of multi-scale and multi-level features. This method can not only retain the local details and global semantic information captured in the shallow features and deep features respectively, but also integrate them into a more expressive feature representation, which helps to improve the image decontamination effect. The reconstructed feature maps will be passed to the global residual learning module for global residual learning. The global residual learning includes inputting the reconstructed feature maps into the channel attention module and the pixel attention module for combination, and then performing two convolutions and adding them element by element to the radar rain and sewage image, further optimizing the stability of gradient transmission and model training, and finally obtaining the corrected radar rain and sewage image.

[0054] Example 1: Correction of rain and sewage image once Please refer to Figure 4 , in this application, an X-band radar is used to obtain the sea surface grayscale image in rainy weather. It can be seen from the figure that there is rain and sewage blocking the sea surface. This is due to the scattering and reflection of radar signals by raindrops, dirt, etc. When the electromagnetic wave emitted by the radar encounters rain particles or fog particles, these particles will scatter and reflect part of the electromagnetic wave, forming a radar echo. These echoes cause the radar to receive abnormal or interference signals, which are then manifested as noise or artifacts in the image. Among them, the white area is the part blocked by rain and sewage. The image processed by the DPMA-NET of this application is as Figure 5 shown. It can be seen that a large amount of rain and sewage in the figure is removed, and the structure of the image after removing rain and sewage is complete, and the correction effect is obvious. According to the comparison after pulse compression Figure 4 and Figure 5 to calculate the variance ratio of the signal and the noise, and the calculation formula is as follows: ; In the formula, g(i,j) is the grayscale value of the uncorrected radar rain and sewage image at (i,j), f(i,j) is the grayscale value of the corrected radar rain and sewage image at (i,j), M and N are the number of pixels in the length and width of the image respectively. Substitute the grayscale values of Figure 4 and 5 at (i,j) respectively, and the signal-to-noise ratio of the peak value of the corrected radar rain and sewage image is 30.86db, indicating that the quality of the rain and sewage image restoration of this application is good and the distortion is small, and the effect is obvious for the computer to use the image for wind field inversion.

[0055] Compare the intensity histograms of the images before and after correction. Among them, the intensity histogram before correction is as Figure 6 shown, and the intensity histogram after correction is as Figure 7As shown in the figure, by comparing the histogram distribution characteristics before and after the correction of radar rain and sewage images, it can be seen that there are significant distribution differences between the rainless radar images (purple) and the rain and sewage radar images (green) in the zero-value pixel area. After being processed by the model of the present application, the histogram statistical characteristics of the two types of images show a convergent trend, specifically manifested as a significant reduction in the difference degree of the values of each bin. This result indicates that the corrected radar image has approached the rainless state at the pixel intensity probability distribution level, effectively suppressing the interference effect of precipitation elements on the radar echo signal, enhancing the accuracy of feature extraction, and significantly improving the quality of radar image correction for rain and sewage.

[0056] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed in this text can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of this application.

[0057] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them; although this application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of each embodiment of this application, and should all be included in the protection scope of this application.

Claims

1. A radar image correction method based on dynamic hybrid attention for wind field inversion, characterized in that: The following steps are involved: Obtain radar rain pollution images and perform shallow feature extraction to obtain initial feature maps; Inputting the initial feature map into a dynamic convolution module to obtain a dynamic convolution kernel, and convolving the kernel with the initial feature map to obtain an output feature map of the dynamic convolution module; The output feature map of the dynamic convolution module is input into the parallel hybrid attention module to obtain the target feature map, and local residual learning is performed to obtain the final feature map; Reconstructing the final feature map and performing global residual learning to obtain a corrected radar rain pollution image; The dynamic convolution module is used to perform global average pooling, two full connection processes and activation function processing on the initial feature map to generate dynamic weights, and use the spline function of the KAN network to dynamically adjust the hyperparameters to normalize the dynamic weights to obtain the weight of the dynamic convolution kernel, and then perform the weighted summation with several expert convolution kernels to obtain the dynamic convolution kernel.

2. The radar image correction method based on dynamic hybrid attention for wind field inversion according to claim 1, characterized in that: The parallel hybrid attention module is used to batch normalize the output feature map of the dynamic convolution module to obtain a batch-normalized feature map, input the output feature map of the channel attention module and the output feature map of the pixel attention module respectively, perform convolution to obtain the output feature map, perform two convolution processes and after activation function processing, perform element-by-element addition with the output feature map of the dynamic convolution module to obtain the target feature map.

3. The radar image correction method based on dynamic hybrid attention for wind field inversion according to claim 2, characterized in that: The dynamic convolution module includes a global average pooling layer, two fully connected layers, a ReLU layer, a KAN network and several expert convolution kernels; The KAN network adaptively fits and calculates the complex relationship of the features in the dynamic weight by changing the shape of the spline in the spline function; Each grid point of the spline function is a spline basis function, and the dynamic weight is normalized by dynamically adjusting the hyperparameters according to the density, shape and smoothness of the grid.

4. The radar image correction method based on dynamic hybrid attention for wind field inversion according to claim 3, characterized in that: After the global average pooling layer performs global average pooling on the initial feature map, the map is input into the first fully connected layer for conversion to obtain a feature vector, which is then input into the ReLU layer for activation and then input into the second fully connected layer for conversion to generate dynamic weights.

5. The radar image correction method based on dynamic hybrid attention for wind field inversion according to claim 4, characterized in that: The parallel hybrid attention module includes a batch normalization layer, the channel attention module and the pixel attention module in parallel, two convolutional layers and a GELU layer; The channel attention module is used to obtain the channel weight according to the batch-normalized feature map and then multiply the channel weight element by element with the batch-normalized feature map to obtain the output feature map of the channel attention module; The formula for obtaining the output feature map of the channel attention module includes: ； ； ； In the formula, is the channel characteristic, For the cth channel exist The value at is the global pooling function, p is the pooling type related parameter in the global pooling operation, and here is the identifier that characterizes the average pooling operation mode. is the feature map after batch normalization, H is the height of the feature map after batch normalization, W is the width of the feature map after batch normalization, i is the coordinate index of the feature map after batch normalization in the height direction, j is the coordinate index of the feature map after batch normalization in the width direction, is the Sigmod function, is the ReLU function, is the channel weight, is the convolutional layer, It is the output feature map of the channel attention module.

6. The radar image correction method based on dynamic hybrid attention for wind field inversion according to claim 5, characterized in that: The pixel attention module obtains pixel weights according to the batch-normalized feature map, and multiplies the pixel weights by the batch-normalized feature map element by element to obtain an output feature map of the pixel attention module; The formula for obtaining the output feature map of the pixel attention module includes: ； ； Where, PA p is the pixel weight, is the Sigmod function, is the ReLU function, is the feature map after batch normalization, is the convolutional layer, is the output feature map of the pixel attention module.

7. The radar image correction method based on dynamic hybrid attention for wind field inversion according to claim 6, characterized in that: The local residual learning includes convolving the target feature map and then adding the initial feature map element by element to obtain the final feature map.

8. The radar image correction method based on dynamic hybrid attention for wind field inversion according to claim 7, characterized in that: The reconstruction includes splicing the obtained multiple final feature maps along the channel dimension to obtain a reconstructed feature map.

9. The radar image correction method based on dynamic hybrid attention for wind field inversion according to claim 8, characterized in that: The global residual learning includes inputting the reconstructed feature map into the channel attention module and the pixel attention module for combination and performing two convolutions, and then adding the radar rain pollution image element by element to obtain the corrected radar rain pollution image.

10. The radar image correction method based on dynamic hybrid attention for wind field inversion according to claim 9, characterized in that: The formula of the spline function includes: ； In the formula, is the spline function, x is the feature of the dynamic weight, C is the coefficient optimized during training, B is the kernel function, and i is the label of the sample point used in the training process.

Images