Radio frequency interference region segmentation method based on morphological prior and deep learning fusion

Abstract

The present application relates to the technical field of radio astronomical signal processing, image pixel processing analysis and deep learning segmentation, and particularly relates to a radio frequency interference area segmentation method based on morphological prior and deep learning fusion, comprising: pre-processing a two-dimensional radio frequency observation image to obtain a radio frequency input image; obtaining a morphological prior image based on fine edge contour and overall edge gradient features of the radio frequency input image; establishing a deep segmentation network; using the radio frequency input image, the morphological prior image and a labeled radio frequency interference mask to establish a data set, and establishing a joint loss function; training the deep segmentation network to obtain a trained segmentation network; inputting a radio frequency image to be processed into the trained segmentation network to obtain an interference area prediction result; calculating feature information of the interference area based on the radio frequency interference area segmentation result to form a radio frequency interference area segmentation report; the present application can improve the precision and adaptability of the interference area segmentation.

Description

Technical Field

[0001] This invention relates to the fields of electroastronomical signal processing, image pixel processing and analysis, and deep learning segmentation technology, specifically to a radio frequency interference region segmentation method based on the fusion of morphological priors and deep learning. Background Technology

[0002] After pulsar observation data undergoes dispersion correction, folding integration, or time-frequency transformation, it typically forms two-dimensional images such as frequency-phase maps and dynamic spectrograms. Due to the complex observation environment, radiation from electronic equipment, and the influence of external radio sources, various forms of radio frequency (RF) interference are often mixed into the images. RF interference can manifest in the images as narrow band bright lines continuously distributed along the frequency axis, broadband stripes extending across multiple frequency channels, locally sudden transient bright patches, or composite interference with multiple superimposed forms. This type of interference can obscure the true signal structure of the pulsar, reducing the accuracy of subsequent signal detection, identification, and parameter estimation. Therefore, accurate segmentation of RF interference regions is of great significance.

[0003] Existing radio frequency interference image processing methods mainly fall into two categories: one is based on traditional image processing methods such as threshold segmentation, edge detection, texture features, and clustering. These methods are simple to implement and highly interpretable, but they are highly dependent on features and rules, and have poor adaptability to complex noise backgrounds, different imaging conditions, and diverse interference patterns. The other category is based on deep learning segmentation methods using convolutional neural networks or Transformers. Although these methods can automatically learn the semantic features of the interference region and have a certain degree of robustness, these general-purpose models are mostly designed for natural images. They are still insufficient in modeling thin stripes, highly directional, blurred boundaries, large scale spans, and small-area weak interference targets in radio frequency images, and they fail to fully utilize the morphological prior information of the interference region.

[0004] The existing technologies mainly have the following problems: insufficient segmentation capability for weak interference and small targets, which easily leads to missed detections and false detections; lack of structural modeling for complex interference such as narrowband, broadband, and transient interference, which often results in rough, broken or adhered boundaries; limited generalization capability in multi-scale and multi-type interference scenarios, making it difficult to adapt to changes under different devices and frequency bands; and weak connection between segmentation results and subsequent feature analysis, making it difficult to fully preserve target structural information.

[0005] Existing methods have not yet formed a radio frequency interference segmentation framework that organically integrates morphological structural priors, deep semantic features, multi-scale contextual information, and boundary constraint mechanisms, thus making it difficult to balance sensitivity, accuracy, and boundary integrity in complex scenarios. Summary of the Invention

[0006] In view of the above problems, the present invention provides a radio frequency interference region segmentation method based on the fusion of morphological priors and deep learning, which solves the technical problems of missed detection of weak interference regions, false positives caused by complex background interference, insufficient modeling of shape and topological features of complex morphological interference regions, poor model adaptability and robustness under flexible transformation of multi-scale interference features, blurred interference region segmentation boundaries, and insufficient utilization of morphological information in the prior art.

[0007] This invention provides a radio frequency interference region segmentation method based on the fusion of morphological priors and deep learning, comprising the following steps: Step S1: Acquire a two-dimensional radio frequency observation image, and preprocess the two-dimensional radio frequency observation image to obtain a radio frequency input image; the radio frequency input image includes linear, striped, blocky and burst radio frequency interference; Step S2: Use the Sobel operator adjusted by the adaptive Gaussian filter kernel and the morphological gradient operator to obtain the fine edge contour and overall edge gradient features of the radio frequency input image; A morphological prior map is obtained based on the fine edge contour and overall edge gradient features; Step S3: Establish a deep segmentation network, which can encode the radio frequency input image and the morphological prior image into image semantic features and morphological structural features, respectively; perform cross-attention multi-scale fusion on the image semantic features and morphological structural features, and obtain the interference region prediction result after boundary enhancement and decoding reconstruction; Step S4: Establish a dataset using the radio frequency input image, morphological prior image, and labeled radio frequency interference mask, and establish a joint loss function; train the deep segmentation network based on the dataset and joint loss function to obtain the trained segmentation network; Step S5: Input the radio frequency image to be processed into the trained segmentation network to obtain the interference region prediction result represented by the pixel-level interference probability map; The pixel-level interference probability map is binarized and post-processed according to a preset threshold to obtain the radio frequency interference region segmentation result. The feature information of the interference region is calculated based on the radio frequency interference region segmentation result to form a radio frequency interference region segmentation report.

[0008] Preferably, in step S1, the two-dimensional radio frequency observation image is a frequency-phase diagram or a dynamic spectrum diagram; The steps for preprocessing the two-dimensional radio frequency observation image include: grayscale normalization, outlier suppression, noise smoothing, and contrast enhancement.

[0009] Preferably, step S2 specifically includes: Step S2-1: Preprocess the radio frequency input image using an adaptive Gaussian filter kernel; wherein, the adaptive adjustment of the standard deviation parameter of the Gaussian filter kernel based on the standard deviation of pixels within a local window includes: for regions where the local standard deviation is greater than a preset threshold, a first Gaussian kernel standard deviation is used; for regions where the local standard deviation is not greater than the preset threshold, a second Gaussian kernel standard deviation greater than the first Gaussian kernel standard deviation is used. The Sobel operator is applied to the image after adaptive Gaussian filtering, and the Sobel edge intensity map is used as the fine edge contour. Step S2-2: Perform dilation and erosion operations on the radio frequency input image using structuring elements, and determine the morphological gradient based on the difference between dilation and erosion, which is used as the overall edge gradient feature; Step S2-3: Perform a weighted summation of the fine edge contour and the overall edge gradient features to obtain a morphological prior image.

[0010] Preferably, in step S3, the deep segmentation network includes: an original image feature extraction branch, a morphological prior feature extraction branch, a multi-scale feature fusion module, a boundary enhancement module, and a decoding and reconstruction module; The original image feature extraction branch includes multiple coding levels, each coding level containing a convolutional layer, a batch normalization layer, and a non-linear activation function; during the coding process, the spatial resolution of the feature map is gradually reduced through downsampling operations. The network structure of the morphological prior feature extraction branch is consistent with that of the original image feature extraction branch.

[0011] Preferably, the multi-scale feature fusion module adopts a cross-attention fusion mechanism, which takes the features of one layer of the original image branch as the query and the features of the corresponding layer of the morphological prior branch as the key and value, calculates the similarity between the query and the key to obtain the attention weight matrix, and then uses the attention weight to perform a weighted summation of the values; the cross-attention fusion mechanism is executed at multiple levels of the encoder respectively, and the obtained multi-scale features are concatenated to obtain the fused features; The boundary enhancement module includes a boundary detection convolutional layer, which uses multiple convolutional kernels to process the fused features and outputs a boundary probability map. The decoding and reconstruction module adopts a skip connection mechanism to fuse the shallow features of the corresponding layer in the encoder with the features of the current layer in the decoder, and gradually upsamples the fused features to restore them to the original image resolution, and generates the final pixel-level prediction result of the interference region.

[0012] Preferably, in step S4, the step of establishing a dataset using the radio frequency input image, morphological prior image, and labeled radio frequency interference mask includes: The dataset consists of multiple radio frequency observation images and their corresponding interference region labeled masks as samples; The interference region labeling mask is obtained through manual labeling, semi-automatic labeling, or automatic labeling.

[0013] Preferably, in step S4, the joint loss function expression is:

[0014]

[0015]

[0016]

[0017] in, Denotes the joint loss function. , , These are the weighting coefficients. Represents pixel classification loss, Indicates the region overlap loss. This represents the boundary constraint loss, where P and R represent precision and recall, respectively, and A and B represent the predicted region and the ground truth region, respectively. This represents the intersection operation. This represents the union operation. Indicates the area of ​​the region.

[0018] Preferably, in step S5, the step of binarizing and post-processing the pixel-level interference probability map according to a preset threshold to obtain the radio frequency interference region segmentation result specifically includes: Pixels with a probability value greater than a preset threshold are identified as interference regions; pixels with a probability value less than or equal to the threshold are identified as background regions, where the preset threshold is 0.5. After post-processing, the radio frequency interference region segmentation result is obtained in the form of a binary mask.

[0019] Preferably, in step S5, the step of calculating the feature information of the interference region based on the radio frequency interference region segmentation result and forming a radio frequency interference region segmentation report specifically includes: Based on the radio frequency interference region segmentation results, feature information of the interference region is extracted. The feature information includes the location, area, outline, direction, length, width, connectivity, and other information of each independent interference region. The segmentation results and extracted feature information are integrated to form a radio frequency interference region segmentation report. The radio frequency interference region segmentation report includes: a segmentation mask image, the original image with superimposed annotations, and the feature information of each interference region.

[0020] Compared with the prior art, the present invention has at least the following beneficial effects: (1) Improved segmentation accuracy in complex radio frequency interference scenarios.

[0021] This invention integrates morphological prior information with deep semantic features. It leverages the modeling advantages of morphological operators in the structural features of linear, striped, and blocky target topologies, while also utilizing the modeling capabilities of deep neural networks in the contextual relationships of complex noise backgrounds and radio frequency interference regions. By effectively guiding the fusion of these two aspects with efficient prior information and powerful image feature representation capabilities, the invention can more accurately distinguish between interference regions and background noise, and improve the model's ability to accurately segment target regions.

[0022] (2) Enhanced the ability to detect weak interference, small targets and slender targets.

[0023] By constructing a topological prior map using multi-scale morphological structural features and combining a boundary enhancement mechanism and a joint loss function, this method can improve the robustness of identifying weak interference, narrow-band interference, and local transient interference under low-contrast environments and imaging conditions with unclear and confused boundaries, thereby reducing the false negative rate. In addition, it effectively improves the model's ability to model the boundary of the target region, resulting in clear segmentation boundaries.

[0024] (3) Improved the continuity of the dividing boundary and the integrity of the region.

[0025] This method introduces morphological operators to extract prior information on morphological topology and a boundary supervision mechanism, which helps to explicitly model the contour boundary features of interference regions during network training. This makes the output results superior to segmentation methods that rely solely on general deep networks in terms of boundary localization and topological region connectivity.

[0026] (4) It has good multi-type adaptability and cross-scenario generalization ability.

[0027] To address narrowband interference, broadband stripe interference, transient interference, and mixed interference, this invention utilizes morphological computation operators and multi-branch feature extraction and multi-scale fusion strategies to achieve unified processing. Compared to manually designed rules that rely on specific interference forms, this method can effectively extract interference region features of different scales and shapes, which is beneficial for improving stability under different observation and imaging scenarios and the accuracy of segmenting interference regions of different scales and shapes.

[0028] (5) Effective connection between segmentation results and subsequent identification analysis was achieved.

[0029] While outputting the segmentation mask, this invention can also calculate the position, orientation, area, and boundary of the interference region in the original radio image through the obtained mask results, providing a reliable basis for subsequent interference classification, automatic labeling, observation data cleaning, and astronomical signal recovery.

[0030] (6) It has good engineering application value.

[0031] This invention can be deployed in the preprocessing workflow of radio astronomy observation data, supporting automated processing of massive observation images, reducing manual screening costs, and improving data processing efficiency. Attached Figure Description

[0032] The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of the invention.

[0033] Figure 1 The flowchart shows the radio frequency interference region segmentation method based on the fusion of morphological priors and deep learning provided by this invention.

[0034] Figure 2 A detailed flowchart of the radio frequency interference region segmentation method based on the fusion of morphological priors and deep learning provided by this invention.

[0035] Figure 3 The network model diagram of the deep segmentation network provided by this invention. Detailed Implementation

[0036] To better understand the above-described objectives, features, and advantages of the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other. Furthermore, the present invention can be implemented in other ways different from those described herein; therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0037] This invention aims to construct a morphological prior extraction mechanism for radio frequency interference structural features, and to synergistically fuse the morphological prior information with multi-scale semantic features extracted by deep neural networks. This enables high-precision pixel-level segmentation of multi-scale, multi-morphological interference targets in the frequency-phase diagram, including narrowband interference, broadband interference, transient interference, and their mixed interference regions. Simultaneously, it enhances the perception capability of low-contrast, weak interference targets, improves the model's ability to capture and segment interference regions, and enhances the boundary continuity, regional integrity, and subsequent analyzability of the segmentation results.

[0038] To illustrate the effectiveness of the method proposed in this invention, the following detailed description of the above technical solution is provided through a specific embodiment, such as... Figure 1As shown, a radio frequency interference region segmentation method based on the fusion of morphological priors and deep learning is disclosed. The specific implementation steps are as follows: Step S1: Acquire a two-dimensional radio frequency observation image, and preprocess the two-dimensional radio frequency observation image to obtain a radio frequency input image; The radio frequency input image includes linear, striped, blocky, and bursty radio frequency interference; In this step, a two-dimensional radio frequency observation image is first obtained from pulsar observation data after dispersion removal, folding integration, or time-frequency transformation.

[0039] The two-dimensional radio frequency observation image can be a frequency-phase diagram, a dynamic spectrum diagram, or other forms of radio frequency observation image. Due to the complex observation environment, radiation from electronic equipment, and the influence of external radio emission sources, the original radio frequency observation image is usually mixed with different forms of radio frequency interference. These radio frequency interferences manifest in the image as various morphological features such as lines, stripes, blocks, and bursts.

[0040] For the original two-dimensional radio frequency observation image, this invention first performs grayscale normalization processing. In some embodiments, the image pixel values ​​can be uniformly mapped to the standard range of [0,1] or [0,255], eliminating the differences in grayscale dynamic range under different observation equipment and imaging conditions, and ensuring that the input data of subsequent processing algorithms has uniform numerical distribution characteristics. Specifically, grayscale normalization can adopt a linear transformation method, mapping the minimum grayscale value of the original image to 0, the maximum grayscale value to 1 or 255, and the intermediate values ​​to a linear relationship.

[0041] Outlier suppression is performed on the normalized image. Radio frequency observation images often contain isolated anomalous pixels caused by equipment transient failures, strong pulse interference, or data transmission errors. These outliers can interfere with subsequent edge detection and feature extraction. In some embodiments, outlier suppression can be achieved using statistical thresholding or median filtering.

[0042] After outlier suppression, noise smoothing is performed to reduce the impact of background random noise. Background noise in radio frequency observation images mainly originates from receiver thermal noise, quantization noise, and electromagnetic interference from the observation environment. Noise smoothing can be achieved using methods such as Gaussian filtering, mean filtering, or bilateral filtering. In some embodiments, Gaussian filtering can be used to smooth the image.

[0043] The noise-smoothed image is then subjected to contrast enhancement processing to improve the separability between the radio frequency interference region and the background. Contrast enhancement can employ methods such as histogram equalization, adaptive histogram equalization, contrast-limited adaptive histogram equalization, or local contrast enhancement. In some embodiments, contrast-limited adaptive histogram equalization can be used, dividing the image into multiple local regions, performing histogram equalization on each region separately, and using contrast limiting parameters to prevent excessive noise amplification. This improves the visual salience of low-contrast, weakly interfering targets while suppressing background noise enhancement.

[0044] After the preprocessing operations such as grayscale normalization, outlier suppression, noise smoothing, and contrast enhancement, a standardized radio frequency (RF) input image is obtained. This RF input image exhibits uniform data distribution characteristics, low background noise levels, and high target-background contrast. The RF interference it contains mainly manifests as linear, striped, blocky, and bursty morphological features, providing input data for subsequent morphological prior construction and deep learning segmentation.

[0045] Step S2: Use the Sobel operator adjusted by the adaptive Gaussian filter kernel and the morphological gradient operator to obtain the fine edge contour and overall edge gradient features of the radio frequency input image; A morphological prior map is obtained based on the fine edge contour and overall edge gradient features; In this step, the present invention constructs a morphological prior map that reflects the structural information of the interference region, targeting the diverse morphological features of radio frequency interference in images, such as linear, striped, blocky, and bursty patterns. This morphological prior map construction allows for the explicit extraction of prior information such as the geometric structure, boundary contours, directional continuity, and topological relationships of the interference region before training the segmentation network, providing structural guidance for subsequent deep neural networks.

[0046] This invention integrates the Sobel operator adjusted by an adaptive Gaussian filter kernel and the morphological gradient operator. The two types of operators capture the edge features of the interference region from different angles.

[0047] (1) The fine edge contour of the radio frequency input image is obtained by using the Sobel operator adjusted by the adaptive Gaussian filter kernel.

[0048] In the complex noise background of radio frequency (RF) images, the traditional Sobel operator lacks sufficient sensitivity for detecting weak edges and easily misses low-contrast, weak interference edges. Therefore, this invention employs a Sobel operator with an adaptive Gaussian filter kernel to obtain fine edge contours of the RF input image, including: An adaptive Gaussian filter kernel is used to preprocess the radio frequency input image. The standard deviation of the adaptive Gaussian filter kernel is dynamically adjusted based on the noise level and edge strength of local image regions. Specifically, for regions where the local standard deviation is greater than a preset threshold (indicating potential edges or interference), a smaller Gaussian kernel standard deviation is used to preserve details; for regions where the local standard deviation is not greater than the preset threshold (indicating a flat background), a larger Gaussian kernel standard deviation is used to adequately smooth noise.

[0049] After adaptive Gaussian filtering preprocessing, the horizontal and vertical convolution kernels of the Sobel operator are convolved with the image to obtain horizontal and vertical gradient maps, respectively. The gradient magnitudes are then calculated to obtain the Sobel edge intensity map, which serves as the fine edge contour.

[0050] (2) The morphological gradient operator is used to extract the overall edge gradient features of the radio frequency input image.

[0051] In this step, dilation and erosion operations can be performed on the radio frequency input image using structural elements. The shape and size of the structural elements are designed according to the typical morphological characteristics of radio frequency interference, and rectangular, elliptical, or linear structural elements can be used.

[0052] In some embodiments, the morphological gradient can be determined based on the difference between dilation and erosion, serving as the overall edge gradient feature. The morphological gradient operator has a better ability to capture the overall contours of both blocky and bursty disturbances in an image.

[0053] (3) Based on the fine edge contour and the overall edge gradient features, feature fusion is performed to obtain the morphological prior map.

[0054] In some embodiments, the feature fusion can employ a weighted summation, maximum value fusion, or a fusion strategy based on decision rules. For example, a weighted summation method can be used to linearly combine fine edge contours and overall edge gradient features according to preset weight coefficients. The weight coefficients can be determined based on the detection performance of the two types of operators under different interference patterns, and are typically set to have equal weights.

[0055] In some embodiments, the fused feature map can be normalized and optionally subjected to morphological post-processing operations, such as morphological closing operations to fill edge gaps and morphological opening operations to remove isolated noise points, to finally obtain a morphological prior map.

[0056] Morphological prior maps express the structural prior knowledge of interference regions in radio frequency input images in the form of images, including information such as boundary location, shape contour, orientation features and topological relationships, providing explicit and prior structural guidance signals for subsequent deep segmentation networks.

[0057] Step S3: Establish a deep segmentation network, which can encode the radio frequency input image and the morphological prior image into image semantic features and morphological structural features, respectively; perform cross-attention multi-scale fusion on the image semantic features and morphological structural features, and obtain the interference region prediction result after boundary enhancement and decoding reconstruction; In this step, the present invention establishes a deep segmentation network for radio frequency interference region segmentation. The deep segmentation network adopts a dual-branch coding architecture, which encodes features of the radio frequency input image and the morphological prior image respectively, and achieves multi-scale feature fusion through a cross-attention mechanism. Finally, the interference region prediction result is obtained through boundary enhancement and decoding reconstruction.

[0058] The deep segmentation network includes: an original image feature extraction branch, a morphological prior feature extraction branch, a multi-scale feature fusion module, a boundary enhancement module, and a decoding and reconstruction module.

[0059] (1) The original image feature extraction branch is used to encode the image semantic features of the radio frequency input image. The original image feature extraction branch uses a convolutional neural network as the encoder network.

[0060] In some embodiments, the encoder network for the original image feature extraction branch can employ ResNet, VGG, or UNet encoders, etc. The encoder network can consist of multiple encoding layers, each containing convolutional layers, batch normalization layers, and nonlinear activation functions. During encoding, the spatial resolution of the feature map can be progressively reduced through downsampling operations, while simultaneously increasing the number of feature channels and expanding the receptive field, thereby extracting multi-scale feature representations from shallow texture features to deep semantic features. Shallow features preserve high-frequency detail information of the image, such as edges, textures, and local details, while deep features capture high-level semantic information and global contextual relationships of the image.

[0061] The original image feature extraction branch outputs multiple feature maps at different scales, for example, four to five different scales, with the feature map resolution decreasing sequentially and the number of channels increasing sequentially. These multiple feature maps at different scales constitute the image semantic features.

[0062] (2) The morphological prior feature extraction branch is used to encode the morphological structure features of the morphological prior image. The network structure of the morphological prior feature extraction branch can be consistent with the network structure of the original image feature extraction branch. It also adopts the convolutional neural network encoder architecture, has the same number of layers and downsampling ratio, and finally outputs morphological structure features.

[0063] Since the morphological prior map has already explicitly modeled the geometric structure of the interference region through mathematical morphological operators, the morphological structural features extracted by the morphological prior feature extraction branch can better characterize the morphological attributes of the interference region, such as shape, contour, orientation, and connectivity.

[0064] (3) The multi-scale feature fusion module is used to perform semantic alignment and deep fusion of image semantic features and morphological structural features to obtain fused features.

[0065] Specifically, this invention employs a cross-attention fusion mechanism to achieve adaptive fusion of features from two branches. In some embodiments, features from one branch can be used as a query, and features from the other branch as keys and values. Cross-branch feature enhancement is achieved through attention weight calculation. Specifically, for a certain layer of features from the original image branch, it is used as the query, and features from the corresponding layer of the morphological prior branch are used as keys and values. The similarity between the query and the keys is calculated to obtain an attention weight matrix, and then the values ​​are weighted and summed using the attention weights.

[0066] Cross-attention fusion is performed at multiple levels of the encoder to achieve multi-scale feature fusion and obtain the fused features.

[0067] At shallow layers, the fusion operation focuses on the complementarity of detailed features, enabling the textural details and morphological edge information of the image to enhance each other. At deep layers, the fusion operation focuses on the alignment of semantic features, enabling the semantic region features and morphological structural features of the image to work together. Through multi-scale cross-attention fusion, the network can integrate two types of complementary information at different levels of abstraction, preserving the real texture and contextual semantics of the original image while introducing morphological prior structural constraints and boundary guidance, significantly improving the ability to model complex morphological interference regions.

[0068] (4) Boundary enhancement module and decoding reconstruction module The boundary enhancement module extracts boundary-related features based on the fused features. In some embodiments, the boundary enhancement module may include a boundary detection convolutional layer. The boundary detection convolutional layer may use multiple convolutional kernels to process the fused features and output a boundary probability map.

[0069] During the training phase, real boundary labels are obtained by performing morphological edge extraction on the real segmentation mask, and supervised learning can be carried out based on the boundary prediction results and the real boundary labels.

[0070] The decoding and reconstruction module can progressively upsample the fused features to restore them to the original image resolution and generate the final pixel-level prediction results of the interference areas.

[0071] In some embodiments, the decoding and reconstruction module employs an upsampling structure symmetrical to the encoder, comprising multiple decoding levels. Each decoding level doubles the spatial resolution of the feature map through an upsampling operation, while adjusting the number of feature channels through a convolution operation.

[0072] To preserve spatial details lost during encoding, the decoding and reconstruction module employs a skip connection mechanism, fusing shallow features from the corresponding level in the encoder with features from the current level in the decoder. Skip connections can be implemented through feature concatenation or element-wise addition. Since the shallow features of the encoder retain rich spatial details and edge information, skip connections effectively compensate for information loss during upsampling, improving the spatial accuracy and boundary clarity of the segmentation results.

[0073] The decoding and reconstruction module ultimately outputs the probability value of each pixel belonging to the interference region, resulting in a pixel-level interference probability map. The spatial resolution of the probability map is consistent with the input RF image, and the value at each pixel position represents the confidence level that the position belongs to the interference region.

[0074] Step S4: Establish a dataset using the radio frequency input image, morphological prior image, and labeled radio frequency interference mask, and establish a joint loss function; train the deep segmentation network based on the dataset and joint loss function to obtain the trained segmentation network; In this step, the present invention trains the constructed deep segmentation network so that the network learns the mapping relationship from the radio frequency input image and morphological prior image to the interference region segmentation mask.

[0075] The dataset of this invention consists of radio frequency observation images and their corresponding interference region annotation masks.

[0076] Interference region annotation masks are obtained through manual, semi-automatic, or automatic annotation. Pixels in interference regions are labeled with a value of 1 or the foreground category, while pixels in background regions are labeled with a value of 0 or the background category. Since annotating interference regions in radio frequency images requires specialized knowledge in radio astronomy, the annotation process can be completed by domain experts using visualization tools. Semi-automatic annotation methods can first generate an initial mask using traditional thresholding or clustering methods, which is then refined by experts.

[0077] In some embodiments, the dataset can be divided into a training set, a validation set, and a test set according to a preset ratio. The training set is used for optimizing network parameters, the validation set is used for hyperparameter tuning and model selection, and the test set is used for final performance evaluation. To improve the model's generalization ability and robustness to different imaging conditions, data augmentation techniques can be used during training, including random horizontal flipping, random vertical flipping, random rotation (90 degrees, 180 degrees, 270 degrees), random brightness adjustment, and random contrast adjustment. Data augmentation can expand the diversity of training samples and mitigate the risk of overfitting.

[0078] This invention establishes a joint loss function. Radio frequency interference region segmentation tasks face problems such as class imbalance (the number of pixels in the interference region is usually much less than the number of background pixels), blurred boundaries, and easy missed detection of small targets. A single loss function is insufficient to comprehensively optimize these problems. This invention establishes a joint loss function that comprehensively considers three aspects: pixel classification accuracy, region overlap, and boundary localization accuracy. The expression is:

[0079]

[0080]

[0081]

[0082] in, Denotes the joint loss function. , , These are the weighting coefficients. Represents pixel classification loss, Indicates the region overlap loss. This represents the boundary constraint loss, where P and R represent precision and recall, respectively, and A and B represent the predicted region and the ground truth region, respectively. This represents the intersection operation. This represents the union operation. Indicates the area of ​​the region.

[0083] This invention trains the deep segmentation network based on the aforementioned dataset and joint loss function. The network training uses the Adam or AdamW optimizer, with an initial learning rate set to... A learning rate decay strategy, such as cosine annealing or step decay, is employed to improve convergence stability. The training batch size is set according to the GPU memory, typically ranging from 4 to 16. The number of training epochs is set to 100 to 200, and early stopping is performed by monitoring the validation set performance to prevent overfitting. During training, both the main segmentation output and the boundary prediction output are optimized simultaneously, enabling the network to explicitly learn boundary localization capabilities while learning semantic segmentation of interference regions.

[0084] Through the aforementioned training strategy based on the joint loss function, the deep segmentation network not only learns the image semantic features of radio frequency interference regions but also learns morphological structure rules and boundary localization patterns, thereby achieving high-precision segmentation capabilities for interference regions of different shapes, scales, and intensities. After training, the model weights with the best performance on the validation set are saved to obtain the trained segmentation network, which can then be used for subsequent inference applications.

[0085] Step S5: Input the radio frequency image to be processed into the trained segmentation network to obtain the interference region prediction result represented by the pixel-level interference probability map; The pixel-level interference probability map is binarized and post-processed according to a preset threshold to obtain the radio frequency interference region segmentation result. The feature information of the interference region is calculated based on the radio frequency interference region segmentation result to form a radio frequency interference region segmentation report.

[0086] In this step, the radio frequency image to be processed is first input into the trained segmentation network to obtain the interference region prediction result represented by the pixel-level interference probability map.

[0087] The pixel-level interference probability map is binarized. In some embodiments, binarization employs a threshold discrimination method, classifying pixels with probability values ​​greater than a preset threshold as interference regions and marking them as 1; pixels with probability values ​​less than or equal to the threshold are classified as background regions and marked as 0. The preset threshold is determined based on application requirements and segmentation accuracy requirements, with a typical value of 0.5.

[0088] The post-processing may include morphological opening operations to remove isolated noise points, morphological closing operations to fill holes and break connections, and connected component analysis to remove false detection regions with excessively small areas. After post-processing, the final radio frequency interference region segmentation result is obtained, represented in the form of a binary mask.

[0089] This invention extracts feature information of the interference region based on the radio frequency interference region segmentation results. The feature information includes the location, area, outline, direction, length, width, connectivity, and other information of each independent interference region.

[0090] Finally, the segmentation results and extracted feature information are integrated to form a structured radio frequency interference (RFI) region segmentation report. The report includes a segmentation mask image, the original image with overlaid annotations, and the aforementioned feature information for each interference region. The RFI region segmentation report can be used by radio astronomy researchers to assess the quality of observational data, identify interference sources, and develop data cleaning strategies.

[0091] While the specific embodiments of the present invention depict actions or steps in a particular order, this should be understood as requiring such actions or steps to be performed in the specific order shown or in sequential order, or requiring all illustrated actions or steps to be performed to achieve the desired result. In certain environments, multitasking and parallel processing may be advantageous. Similarly, although several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of this disclosure. Certain features described in the context of individual embodiments may also be implemented in combination in a single implementation. Conversely, various features described in the context of a single implementation may also be implemented individually or in any suitable sub-combination in multiple implementations.

[0092] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A radio frequency interference region segmentation method based on the fusion of morphological priors and deep learning, characterized in that, Includes the following steps: Step S1: Acquire a two-dimensional radio frequency observation image, and preprocess the two-dimensional radio frequency observation image to obtain a radio frequency input image; the radio frequency input image includes linear, striped, blocky and burst radio frequency interference; Step S2: Use the Sobel operator adjusted by the adaptive Gaussian filter kernel and the morphological gradient operator to obtain the fine edge contour and overall edge gradient features of the radio frequency input image; A morphological prior image is obtained based on the fine edge contour and overall edge gradient features; Step S3: Establish a deep segmentation network, which can encode the radio frequency input image and the morphological prior image into image semantic features and morphological structural features, respectively; Cross-attention multi-scale fusion is performed on the semantic features and morphological structural features of the image, and the interference region prediction result is obtained after boundary enhancement and decoding reconstruction; Step S4: Establish a dataset using the radio frequency input image, morphological prior image, and labeled radio frequency interference mask, and establish a joint loss function; train the deep segmentation network based on the dataset and joint loss function to obtain the trained segmentation network; Step S5: Input the radio frequency image to be processed into the trained segmentation network to obtain the interference region prediction result represented by the pixel-level interference probability map; The pixel-level interference probability map is binarized and post-processed according to a preset threshold to obtain the radio frequency interference region segmentation result. The feature information of the interference region is calculated based on the radio frequency interference region segmentation result to form a radio frequency interference region segmentation report.

2. The radio frequency interference region segmentation method based on the fusion of morphological priors and deep learning as described in claim 1, characterized in that, In step S1, the two-dimensional radio frequency observation image is a frequency-phase diagram or a dynamic spectrum diagram; The steps for preprocessing the two-dimensional radio frequency observation image include: grayscale normalization, outlier suppression, noise smoothing, and contrast enhancement.

3. The radio frequency interference region segmentation method based on the fusion of morphological priors and deep learning according to claim 2, characterized in that, Step S2 specifically includes: Step S2-1: Preprocess the radio frequency input image using an adaptive Gaussian filter kernel; wherein, the adaptive adjustment of the standard deviation parameter of the Gaussian filter kernel based on the standard deviation of pixels within a local window includes: for regions where the local standard deviation is greater than a preset threshold, a first Gaussian kernel standard deviation is used; for regions where the local standard deviation is not greater than the preset threshold, a second Gaussian kernel standard deviation greater than the first Gaussian kernel standard deviation is used. The Sobel operator is applied to the image after adaptive Gaussian filtering, and the Sobel edge intensity map is used as the fine edge contour. Step S2-2: Perform dilation and erosion operations on the radio frequency input image using structuring elements, and determine the morphological gradient based on the difference between dilation and erosion, which is used as the overall edge gradient feature; Step S2-3: Perform a weighted summation of the fine edge contour and the overall edge gradient features to obtain a morphological prior image.

4. The radio frequency interference region segmentation method based on the fusion of morphological priors and deep learning according to claim 3, characterized in that, In step S3, the deep segmentation network includes: an original image feature extraction branch, a morphological prior feature extraction branch, a multi-scale feature fusion module, a boundary enhancement module, and a decoding and reconstruction module; The original image feature extraction branch includes multiple coding levels, each coding level containing a convolutional layer, a batch normalization layer, and a non-linear activation function; during the coding process, the spatial resolution of the feature map is gradually reduced through downsampling operations. The network structure of the morphological prior feature extraction branch is consistent with that of the original image feature extraction branch.

5. The radio frequency interference region segmentation method based on the fusion of morphological priors and deep learning according to claim 4, characterized in that, The multi-scale feature fusion module employs a cross-attention fusion mechanism, using one layer of features from the original image branch as the query and features from the corresponding layer of the morphological prior branch as the key and value. It calculates the similarity between the query and the key to obtain an attention weight matrix, and then uses the attention weights to perform a weighted summation of the values. The cross-attention fusion mechanism is executed at multiple levels of the encoder, and the obtained multi-scale features are concatenated to obtain the fused features. The boundary enhancement module includes a boundary detection convolutional layer, which uses multiple convolutional kernels to process the fused features and outputs a boundary probability map. The decoding and reconstruction module adopts a skip connection mechanism to fuse the shallow features of the corresponding layer in the encoder with the features of the current layer in the decoder, and gradually upsamples the fused features to restore them to the original image resolution, and generates the final pixel-level prediction result of the interference region.

6. The radio frequency interference region segmentation method based on the fusion of morphological priors and deep learning according to claim 5, characterized in that, Step S4, the step of establishing a dataset using the radio frequency input image, morphological prior image, and labeled radio frequency interference mask, includes: The dataset consists of multiple radio frequency observation images and their corresponding interference region labeled masks as samples; The interference region labeling mask is obtained through manual labeling, semi-automatic labeling, or automatic labeling.

7. The radio frequency interference region segmentation method based on the fusion of morphological priors and deep learning according to claim 6, characterized in that, In step S4, the joint loss function expression is: in, Denotes the joint loss function. , , These are the weighting coefficients. Represents pixel classification loss, Indicates the region overlap loss. This represents the boundary constraint loss, where P and R represent precision and recall, respectively, and A and B represent the predicted region and the ground truth region, respectively. This represents the intersection operation. This represents the union operation. Indicates the area of ​​the region.

8. The radio frequency interference region segmentation method based on the fusion of morphological priors and deep learning according to claim 7, characterized in that, In step S5, the step of binarizing and post-processing the pixel-level interference probability map according to a preset threshold to obtain the radio frequency interference region segmentation result specifically includes: Pixels with a probability value greater than a preset threshold are identified as interference regions; pixels with a probability value less than or equal to the threshold are identified as background regions, where the preset threshold is 0.

5. After post-processing, the radio frequency interference region segmentation result is obtained in the form of a binary mask.

9. The radio frequency interference region segmentation method based on the fusion of morphological priors and deep learning according to claim 8, characterized in that, In step S5, the step of calculating the feature information of the interference region based on the radio frequency interference region segmentation result and forming a radio frequency interference region segmentation report specifically includes: Based on the radio frequency interference region segmentation results, feature information of the interference region is extracted. The feature information includes the location, area, outline, direction, length, width, connectivity, and other information of each independent interference region. The segmentation results and extracted feature information are integrated to form a radio frequency interference region segmentation report. The radio frequency interference region segmentation report includes: a segmentation mask image, the original image with superimposed annotations, and the feature information of each interference region.

Images