Automatic segmentation method of F region of return scattering ionization map based on Haar wavelet down-sampling

Abstract

The application discloses a return scattering ionization map F area automatic segmentation method based on Haar wavelet downsampling, to solve the problems of fuzzy F area boundary, low contrast and high frequency information loss in the prior art. The method comprises the following steps: step 1, data acquisition, labeling and pretreatment; step 2, constructing a SegNext segmentation network based on Haar wavelet downsampling; step 3, decoding fusion and segmentation prediction output; step 4, training optimization based on edge weighted focus joint loss. The method retains the high frequency characteristics of the ionization map by using the wavelet downsampling module, and combines the focus loss and the edge weighting mechanism, so that the segmentation result is more accurate and stable in the boundary area.

Description

Technical Field

[0001] This invention belongs to the field of ionospheric image processing and deep learning technology, specifically relating to an automatic segmentation method for the F region of a backscattered ionograph based on Haar wavelet downsampling. Background Technology

[0002] A reflected ionosphere image is an image of the reflected signal obtained by transmitting electromagnetic waves through ground-based ionospheric detectors, which are then scattered in the ionosphere and return to the receiving system. This image reflects the electron density distribution at different altitudes in the ionosphere and is widely used in fields such as space environment monitoring, communication signal propagation research, and radio wave propagation prediction.

[0003] However, due to the complexity of the ionosphere and its susceptibility to various factors such as weather, seasons, and solar activity, the quality of backscattered ionographs fluctuates significantly, especially when there is substantial noise or missing data in the image, posing a significant challenge to segmentation accuracy. The F region of the ionosphere, as the area most significantly affecting radio propagation, typically exhibits blurred boundaries and low contrast, particularly noticeable in low signal-to-noise ratio environments.

[0004] Traditional segmentation methods, such as thresholding and edge detection, often struggle to achieve high accuracy in the F-region segmentation of ionographs due to factors like signal noise, illumination variations, and blurred boundaries. These methods rely on manually set thresholds or rules and cannot automatically adapt to changes in the F-region of the ionograph, resulting in poor stability and accuracy of the segmentation results, especially in complex environments.

[0005] With the rapid development of deep learning technology, methods based on convolutional neural networks (CNNs) have gradually become the mainstream in the field of image segmentation, especially achieving remarkable results in high-precision segmentation tasks such as medical images and remote sensing images. In recent years, deep learning models have also been widely used in ionization image segmentation tasks. Nevertheless, most methods still face the following challenges: (1) Loss of high-frequency edge information: When performing downsampling, the convolution operation in the deep learning model usually loses high-frequency information in the image, especially the boundary information in the ionized image, which is crucial. Therefore, it is necessary to ensure that the model can effectively preserve these high-frequency edge features.

[0006] (2) Low contrast and blurred boundaries: The F region of the ionization map has blurred boundaries and low contrast. Traditional segmentation networks are difficult to effectively cope with these challenges of low contrast and blurred boundaries, resulting in unsatisfactory segmentation results.

[0007] (3) Computational resources and efficiency issues: Deep learning models usually require a lot of computational resources and time, especially when dealing with large-scale ionization graph data, computational efficiency and the real-time performance of the model become key issues.

[0008] To address the aforementioned issues, recent studies have proposed using wavelet transform to handle image downsampling. Due to its excellent frequency resolution and localization properties, wavelet transform exhibits significant advantages in preserving high-frequency edge information in images. In particular, the Haar wavelet is computationally simple and efficient, and is widely used in image compression, denoising, and other fields. However, existing wavelet transform-based downsampling methods mostly focus on image compression and denoising; a systematic approach to effectively integrate wavelet transform into deep neural networks to preserve the boundary features of the F-region of ionization maps remains lacking.

[0009] Furthermore, traditional loss functions (such as cross-entropy loss and Dice loss) do not fully consider the special characteristics of boundary regions when dealing with blurred and difficult-to-separate samples. Although FocalLoss has been applied in object detection, its application in ionization map segmentation is still relatively limited, especially when combined with edge information. How to design an effective loss function to improve attention to the boundaries of the F region of the ionization map remains a technical challenge to be solved.

[0010] Therefore, this invention proposes an automatic segmentation method for the F region of the return scattering ionization map based on Haar wavelet downsampling and edge-weighted focus joint loss. Summary of the Invention

[0011] To overcome the shortcomings of existing technologies, the present invention aims to provide an automatic segmentation method for the F region of a return scattering ionization map based on Haar wavelet downsampling, in order to solve the problems of blurred boundaries, low contrast, and loss of high-frequency information in the F region of the return scattering ionization map in existing technologies. By using a wavelet downsampling module to preserve the high-frequency features of the ionization map, and combining focus loss and edge weighting mechanisms, the segmentation results are more accurate and stable in the boundary region.

[0012] The automatic segmentation method for the F region of the return scattering ionization map based on Haar wavelet downsampling includes the following steps: Step 1: Data acquisition, labeling, and preprocessing; Step 2: Construct a SegNext segmentation network based on Haar wavelet downsampling; Step 3: Decode, fuse, and segmentation prediction output; Step 4: Training optimization based on edge-weighted focus joint loss.

[0013] Furthermore, the specific process of step 1 is as follows: Step 1-1: Obtain the original image of the return scattering ionization map. To create a dataset for training, validation, or testing; Steps 1-2 use image annotation tools to manually annotate the F-region boundary in the returned scattering ionization map to obtain a boundary annotation file; the boundary must include the leading edge and trailing edge of the F-region. Steps 1-3, based on the boundary annotation file obtained in Step 1-2, convert the boundary annotations into pixel-level mask annotations using a script program. Mask label It can be a binary mask or a multi-class mask; where the mask pixel values ​​are used to indicate the target area and background area in region F; Steps 1-4 on the image With mask label Perform consistency preprocessing to obtain network input. With training labels Consistency preprocessing includes: size scaling, cropping, normalization, intensity mapping, outlier removal, or data augmentation.

[0014] Furthermore, the specific process of step 2 is as follows: Step 2-1: Using the SegNext semantic segmentation network as the baseline network, construct an encoder-decoder structure; the encoder contains downsampling layers with multi-level scale transformations. Step 2-2 replaces each traditional convolutional downsampling layer in the encoder with a Haar wavelet downsampling module to explicitly encode high-frequency edge information during multi-scale feature extraction. Steps 2-3: Process the input feature map Perform a one-level two-dimensional Haar wavelet transform to obtain the low-frequency subband. With high-frequency subband The spatial resolution of each sub-band is ; Steps 2-4 involve concatenating the four sub-bands along the channel dimension to obtain: ; Steps 2-5 conduct Convolutional fusion and downsampling output: , in, for Convolution kernel parameters, For learnable parameters, BN is batch normalization. For activation functions; Step 2-6: Repeat steps 2-2 to 2-5 to obtain multi-scale encoded features. .

[0015] Furthermore, in steps 2-3, the two-dimensional Haar wavelet transform performs low-pass filtering on the input feature map in both the horizontal and vertical directions. With high-pass filtering And by performing a 2x downsampling, four subbands are obtained: Low-pass to low-pass component; High-pass-low-pass component; Low-pass to high-pass component; Qualcomm-Qualcomm component.

[0016] Furthermore, the specific process of step 3 is as follows: Step 3-1 Encodes features at multiple scales The input decoder performs multi-scale contextual information aggregation and feature fusion. Step 3-2 performs stepwise upsampling on the fused features and outputs the pixel-level prediction probability map of region F. ; Step 3-3 according to Generate segmentation results.

[0017] Furthermore, the specific process of step 4 is as follows: Step 4-1 Based on real labels Extracting a single-pixel wide-border mask ; Step 4-2 Boundary Mask Perform a distance transformation to obtain a distance map. ,in, Represents pixels Distance to the nearest boundary; Step 4-3 Based on distance map Constructing an edge weight graph This maximizes the weight at the boundary and allows it to smoothly decay to the baseline value as the distance increases. Step 4-4: Construct the edge-weighted focus cross-entropy loss: , in, For pixels The predicted probability of the true class. Total number of pixels For class balancing parameters, These are the focus modulation parameters; Steps 4-5: Construct the Dice loss: ,in, For smoothing terms; Steps 4-6: Construct the joint loss function and perform training and optimization. ,in, The balance coefficient is used; The network parameters are iteratively updated to complete model training.

[0018] Furthermore, in steps 4-3, the edge weight map The generation process is as follows: First, based on the real tags Extracting a single-pixel wide-border mask Then, for the boundary mask Perform a distance transformation to obtain a distance map from each pixel to its nearest boundary. ,in Represents pixels Distance to the boundary; finally, based on the distance graph Constructing an edge weight graph This maximizes the weight at the boundary and allows it to decrease smoothly as the distance increases.

[0019] Compared with the prior art, the present invention has the following advantages:

[0020] This invention replaces all levels of downsampling structures in the SegNext-based segmentation network encoder with Haar wavelet downsampling modules, enabling features to retain both low-frequency structural and high-frequency edge information during downsampling, thus mitigating the loss of boundary details caused by traditional convolutional downsampling. Simultaneously, this invention constructs an edge-weighted focal joint loss function that integrates edge priors and hard-example mining mechanisms. By weighting the focal cross-entropy loss with an edge weight map generated based on the real label boundaries and distance transformation, the training process generates stronger gradient drives in blurred boundary regions and hard-to-distinguish pixels. Combined with Dice loss, it constrains the overall consistency of the target region, thereby improving the segmentation stability and consistency of key boundaries such as the leading and trailing edges of the F-region of the backscattered ionization map. Furthermore, this invention provides a complete implementation process from manually annotating boundaries to script-generated masks, and then to training and inference. The structure is simple, the implementation path is clear, and it is easy to apply in automatic segmentation tasks of the F-region of backscattered ionization maps. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0023] like Figure 1As shown, the automatic segmentation method for the F region of the return scattering ionization map based on Haar wavelet downsampling includes the following steps: (a) Step 1: Data acquisition, labeling and preprocessing.

[0024] The specific process for this step includes: Step 1-1: Obtain the original image of the return scattering ionization map. Create a dataset for training, validation, or testing. Images can be grayscale or color images, and can be stored in PNG, JPG, TIFF, or other common image formats.

[0025] Steps 1-2 involve manually annotating the F-region boundary in the returned scattering ionization map using an image annotation tool to obtain a boundary annotation file. The boundary must include both the leading edge and trailing edge of the F-region. As a preferred embodiment, Labelme can be used as the image annotation tool to annotate the boundary using line segments and generate an annotation file, such as a JSON file.

[0026] Steps 1-3 use a script to convert the obtained boundary annotation file into pixel-level mask annotations. Specifically, the script reads the set of annotation points and rasterizes them onto a blank canvas of the same size as the original image, following their spatial order: it connects the first annotation point at the leading edge to the first annotation point at the trailing edge, and similarly connects the last annotation point at the leading edge to the last annotation point at the trailing edge, thus obtaining a closed region. This region is then filled to generate the target region. The generated mask annotations... It can be a binary mask (foreground / background) and maintains the same spatial resolution and alignment as the original image.

[0027] Steps 1-4 on the image With mask label Perform consistency preprocessing to obtain network input. With training labels Consistency preprocessing includes: normalization (scaling pixel intensity to a preset range), size scaling and cropping (making the input meet the network's required size), and optional data augmentation (e.g., random flipping, random cropping, slight intensity perturbation). Preprocessing operations must maintain the same geometric transformations for both the image and the mask to ensure the consistency of the supervisory signal.

[0028] (ii) Step 2: Construct a SegNext segmentation network based on Haar wavelet downsampling.

[0029] The specific process for this step includes: Step 2-1 constructs an encoder-decoder structure based on the SegNext semantic segmentation network. The encoder contains multi-level feature extraction stages, each stage including several convolutional / normalization / activation and context aggregation structures, and scale transformation is achieved through downsampling between stages; the decoder is used to fuse multi-scale features and upsample them step by step to restore spatial resolution.

[0030] Step 2-2 replaces the traditional convolutional downsampling layer at each scale transformation level in the encoder with a Haar wavelet downsampling module as an improved feature extraction method of the present invention, so as to explicitly encode high-frequency edge information during multi-scale feature extraction. "Unified replacement" means that all downsampling positions of the encoder adopt the Haar wavelet downsampling module, so that high-frequency edge information can be continuously encoded into the feature channel during multi-scale feature extraction.

[0031] Steps 2-3: Process the input feature map Performing a one-level two-dimensional Haar wavelet transform yields four subbands: the low-frequency subband. With high-frequency subband And the spatial resolution of each sub-band is The two-dimensional Haar wavelet transform can be achieved by using a fixed orthogonal filter bank to perform low-pass filtering in the horizontal and high-pass filtering directions respectively, and then performing a 2x downsampling to obtain four sub-bands; or it can be achieved using an equivalent wavelet decomposition operator.

[0032] Specifically, the two-dimensional Haar wavelet transform performs low-pass filtering on the input feature map in both the horizontal and vertical directions. With high-pass filtering And by performing a 2x downsampling, four subbands are obtained: Low-pass to low-pass component; High-pass-low-pass component; Low-pass to high-pass component; Qualcomm-Qualcomm component.

[0033] Steps 2-4 involve concatenating the four sub-bands along the channel dimension to obtain: .

[0034] Steps 2-5 conduct Convolutional fusion, combined with normalization and activation functions, outputs downsampled features: , in, for Convolution kernel parameters, For learnable parameters, BN is batch normalization. is the activation function (ReLU is preferred).

[0035] Step 2-6: Repeat steps 2-2 to 2-5 to obtain the multi-scale encoded feature set. .

[0036] The Haar wavelet downsampling module is used only as a downsampling replacement unit and does not have an additional high-frequency information backflow bypass structure; at the same time, the original multi-scale context information aggregation mechanism of the SegNext baseline network is retained for subsequent decoding and segmentation prediction.

[0037] (III) Step 3: Decoding, fusion and segmentation prediction output.

[0038] The specific process for this step includes: Step 3-1: Process the multi-scale features output by the encoder The input decoder performs multi-scale contextual information aggregation and feature fusion. The fusion method can be convolutional fusion after feature concatenation, stepwise additive fusion, or attention fusion, etc. It is preferred to adopt a fusion strategy consistent with SegNext to ensure structural stability.

[0039] Step 3-2: Upsample the fused features step by step to restore them to the same spatial resolution as the input image or a preset resolution, and output a pixel-level prediction probability map. When it is a binary classification segmentation, This represents the probability of the foreground (F region); in multi-class segmentation, This represents the probability distribution of each category.

[0040] Step 3-3: In the inference phase, the predicted probability map is generated. A prediction mask is generated by thresholding or taking the class with the highest probability. This serves as the automatic segmentation result of the F region in the returned scattering ionization map.

[0041] (iv) Step 4: Training optimization based on edge-weighted focus joint loss.

[0042] The specific process for this step includes: Step 4-1: Based on real labels Extracting a single-pixel wide-border mask Boundary extraction can be achieved using morphological gradients, contour extraction, or dilation-erosion difference methods, making... This represents the set of boundary pixels of the target region.

[0043] Step 4-2: Boundary Mask Perform a distance transformation to obtain a distance map. ,in, Represents pixels Distance to the nearest boundary pixel.

[0044] Step 4-3: Based on distance map Constructing an edge weight graph This maximizes the weight at the boundary and smoothly decays to a baseline value of 1 as the distance increases. In a preferred embodiment, the edge weight map satisfies any of the following forms: exponential decay: , Or linear decay: , in, , or These are preset parameters used to control the intensity and decay rate of edge attention.

[0045] Step 4-4: Construct the edge-weighted focal cross-entropy loss: , in, For pixels The predicted probability of the true class. For class balancing parameters, For focus modulation parameters, The total number of pixels. This loss function allows the model to focus on both indistinguishable pixels and boundary regions during training.

[0046] Steps 4-5: Construct the Dice loss: , in, This is a smoothing term used to avoid zero denominators and improve numerical stability.

[0047] Steps 4-6: Construct the joint loss function and perform training and optimization: , in, This is the balancing coefficient. During training, stochastic gradient descent or adaptive optimization algorithms are used to iteratively update the network parameters until a preset stopping condition is met (e.g., reaching a preset number of rounds or a loss convergence threshold), resulting in a trained segmentation model.

[0048] This invention integrates wavelet transform into a deep learning network, effectively preserving high-frequency edge information in ionization images. Furthermore, by combining an edge-weighted focus loss function, it enhances attention to boundary regions during training, thereby improving the accuracy and stability of F-region segmentation. This method not only effectively solves the problems of low contrast and blurred boundaries in ionization image segmentation but also improves computational efficiency, meeting the processing needs of large-scale ionization image data.

[0049] Verification Example In this embodiment, the specific implementation process of the method of the present invention is illustrated using a backscattering ionization map dataset as an example. The dataset contains 1200 backscattering ionization map images, which are RGB images with an original resolution of (1024×768) pixels and stored in PNG format. This embodiment is a binary classification segmentation task, with only two categories: F-region targets and background. The dataset is divided into a training set, a validation set, and a test set in a 7:2:1 ratio, containing 840, 240, and 120 images respectively.

[0050] (1) Labeling and mask generation Step 1-1: Manually annotate the images in the training and validation sets using an image annotation tool. The annotation targets are the leading and trailing edges of the F region in the returned scattering ionization map. The annotation method is polyline annotation, forming a boundary annotation file. Labelme is used as the annotation tool, and the output is a JSON format annotation file.

[0051] Steps 1-2: Convert the JSON annotation file into pixel-level mask annotations using a script. Specifically, the script reads the sequence of marker points for the leading edge and the trailing edge respectively. Let the sequence of marker points for the leading edge be... The sequence of trailing markers is The script constructs the target closed region using the following closure rule: the first marked point of the leading edge... The first annotation point at the trailing edge Connect, and mark the last point of the leading edge. The last annotation point at the trailing edge The connection allows the leading edge curve, the trailing edge curve, and the two connecting lines to form a closed region. Then, the closed region... The execution region filling generates the target region, resulting in a binary mask. The mask encoding rules stipulate that the pixel value of area F is 1, and the pixel value of the background is 0.

[0052] Steps 1-3: Perform consistency preprocessing on the image and mask. Normalize the image intensity so that pixel values ​​are mapped to... The interval is used to obtain the network input. With training labels Meanwhile, data augmentation strategies were employed during the training phase, including random horizontal flipping (probability 0.5) and random pruning.

[0053] (2) Network structure configuration Step 2-1: Using SegNext as the baseline segmentation network, construct an encoder-decoder structure and set the output category to binary classification (F region / background). The encoder is set with 4 scale stages, with the number of output channels in each stage being 64, 128, 320, and 512, respectively.

[0054] Step 2-2: Replace each downsampling layer in the encoder with a Haar wavelet downsampling module. For any input feature... Perform a one-level two-dimensional Haar wavelet decomposition to obtain The resolution of each sub-band is The four sub-bands are spliced ​​together in the channel dimension. Then through Convolutional fusion yields output features .

[0055] (3) Output and Reasoning Step 3-1: The decoder fuses multi-scale features and upsamples them step by step, outputting an F-region prediction probability map of the same size as the input. .

[0056] Step 3-2: Reasoning Stage Thresholding is performed to obtain the prediction mask. The threshold is set to 0.5; when Pixel determination If it is zone F, otherwise it is the background.

[0057] (4) Loss function, marginal weights and training parameters Step 4-1: From the real mask Extracting a single-pixel wide-border mask and to Perform distance transformation to obtain a distance map .

[0058] Step 4-2: Construct the edge weight graph It adopts an exponential decay form: , Among them, take , .

[0059] Step 4-3: Construct edge-weighted focal cross-entropy loss Among them, take , Constructing the Dice loss Smoothing term The combined loss is: , in, .

[0060] Step 4-4: Training settings are as follows: AdamW is selected as the optimizer. , The learning rate strategy employs polynomial decay. , Update by iteration ( Training uses IterBasedRunner, with a maximum number of iterations. Model checkpoint save interval (Save by iteration,) Model evaluation interval The evaluation indicators are configured as follows (This embodiment does not limit the specific evaluation results.)

[0061] Steps 4-5: The training hardware environment is a single NVIDIA GeForce RTX 4090 GPU or equivalent computing resources.

[0062] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An automatic segmentation method for the F region of a return scattering ionization map based on Haar wavelet downsampling, characterized in that, Includes the following steps: Step 1: Data acquisition, labeling, and preprocessing; Step 2: Construct a SegNext segmentation network based on Haar wavelet downsampling; Step 3: Decode, fuse, and segmentation prediction output; Step 4: Training optimization based on edge-weighted focus joint loss.

2. The automatic segmentation method for F-region of return scattering ionization map based on Haar wavelet downsampling according to claim 1, characterized in that, The specific process of step 1 is as follows: Step 1-1: Obtain the original image of the return scattering ionization map. To create a dataset for training, validation, or testing; Steps 1-2 use image annotation tools to manually annotate the F-region boundary in the returned scattering ionization map to obtain a boundary annotation file; the boundary must include the leading edge and trailing edge of the F-region. Steps 1-3, based on the boundary annotation file obtained in Step 1-2, convert the boundary annotations into pixel-level mask annotations using a script program. Mask label It can be a binary mask or a multi-class mask; where the mask pixel values ​​are used to indicate the target area and background area in region F; Steps 1-4 on the image With mask label Perform consistency preprocessing to obtain network input. With training labels Consistency preprocessing includes: size scaling, cropping, normalization, intensity mapping, outlier removal, or data augmentation.

3. The automatic segmentation method for F-region of return scattering ionization map based on Haar wavelet downsampling according to claim 1, characterized in that, The specific process of step 2 is as follows: Step 2-1: Using the SegNext semantic segmentation network as the baseline network, construct an encoder-decoder structure; the encoder contains downsampling layers with multi-level scale transformations. Step 2-2 replaces each traditional convolutional downsampling layer in the encoder with a Haar wavelet downsampling module to explicitly encode high-frequency edge information during multi-scale feature extraction. Steps 2-3: Process the input feature map Perform a one-level two-dimensional Haar wavelet transform to obtain the low-frequency subband. With high-frequency subband The spatial resolution of each sub-band is ; Steps 2-4 involve concatenating the four sub-bands along the channel dimension to obtain: ； Steps 2-5 conduct Convolutional fusion and downsampling output: ， in, for Convolution kernel parameters, For learnable parameters, BN is batch normalization. For activation functions; Step 2-6: Repeat steps 2-2 to 2-5 to obtain multi-scale encoded features. .

4. The automatic segmentation method for F-region of return scattering ionization map based on Haar wavelet downsampling according to claim 3, characterized in that, In steps 2-3, the two-dimensional Haar wavelet transform performs low-pass filtering on the input feature map in both the horizontal and vertical directions. With high-pass filtering And by performing a 2x downsampling, four subbands are obtained: Low-pass to low-pass component; High-pass-low-pass component; Low-pass to high-pass component; Qualcomm-Qualcomm component.

5. The automatic segmentation method for F-region of return scattering ionization map based on Haar wavelet downsampling according to claim 1, characterized in that, The specific process of step 3 is as follows: Step 3-1 Encodes features at multiple scales The input decoder performs multi-scale contextual information aggregation and feature fusion. Step 3-2 performs stepwise upsampling on the fused features and outputs the pixel-level prediction probability map of region F. ; Step 3-3 according to Generate segmentation results.

6. The automatic segmentation method for F-region of return scattering ionization map based on Haar wavelet downsampling according to claim 1, characterized in that, The specific process of step 4 is as follows: Step 4-1 Based on real labels Extracting a single-pixel wide-border mask ; Step 4-2 Boundary Mask Perform a distance transformation to obtain a distance map. ,in, Represents pixels Distance to the nearest boundary; Step 4-3 Based on distance map Constructing an edge weight graph This maximizes the weight at the boundary and allows it to smoothly decay to the baseline value as the distance increases. Step 4-4: Construct the edge-weighted focus cross-entropy loss: ， in, For pixels The predicted probability of the true class. Total number of pixels For class balancing parameters, These are the focus modulation parameters; Steps 4-5: Construct the Dice loss: ,in, For smoothing terms; Steps 4-6: Construct the joint loss function and perform training and optimization. ,in, The balance coefficient is used; The network parameters are iteratively updated to complete model training.

7. The automatic segmentation method for F-region of return scattering ionization map based on Haar wavelet downsampling according to claim 6, characterized in that, In step 4-3, the edge weight map The generation process is as follows: First, based on the real tags Extracting a single-pixel wide-border mask ; Then, for the boundary mask Perform a distance transformation to obtain a distance map from each pixel to its nearest boundary. ,in Represents pixels Distance to the boundary; finally, based on the distance graph Constructing an edge weight graph This maximizes the weight at the boundary and allows it to decrease smoothly as the distance increases.

Images