Pulsar candidate body recognition model construction method and pulsar candidate body recognition method

Abstract

The invention relates to a pulsar candidate body recognition model construction method and a pulsar candidate body recognition method. Obtaining a preset candidate body data set, and performing sub-graph segmentation processing on each pulsar candidate body sample image in the preset candidate body data set to obtain a target sub-graph data set; performing labeling processing on the target sub-graph data set, and constructing a labeled target sub-graph data set; performing fine tuning training on a preset multi-modal large language model by using the labeled target sub-graph data set to obtain an initial pulsar candidate body recognition model; using a preset reward function and the labeled target sub-graph data set to perform enhanced training on the initial pulsar candidate body recognition model to obtain a pulsar candidate body recognition model; the recognition accuracy and robustness of the model on key categories are improved, and the finally obtained pulsar candidate body recognition model has higher generalization ability and higher recognition precision.

Description

Technical Field

[0001] The present application relates to the technical field of astronomical image recognition, and in particular to a method for constructing a pulsar candidate recognition model and a pulsar candidate recognition method. Background Art

[0002] Pulsars are a type of rapidly rotating, ultra-dense neutron star. Theoretical research and observations of pulsars are crucial for advancing the development of physics and astrophysics, for example, in areas such as the study of neutron star evolution, equations of state for dense matter, and verification of general relativity. Therefore, identifying pulsar candidates is crucial for scientific research in physics and astronomy.

[0003] However, traditional pulsar candidate identification methods are limited by the image information features contained in the training data. Differences in observation equipment, signal processing methods, and time spans among astronomical observation projects result in significant variations in image quality, resolution, and statistical properties among the data generated. This results in poor generalization of models trained on a single dataset to new data sources, significantly reducing recognition accuracy.

[0004] Currently, no effective solution has been proposed to address the problems of poor generalization and low accuracy of traditional pulsar candidate identification methods in related technologies. Summary of the Invention

[0005] Based on this, it is necessary to provide a pulsar candidate identification model construction method and a pulsar candidate identification method to address the above technical problems.

[0006] In a first aspect, the present application provides a method for constructing a pulsar candidate identification model, the method comprising:

[0007] Obtaining a preset candidate data set, performing sub-image segmentation processing on each pulsar candidate sample image in the preset candidate data set to obtain a target sub-image data set;

[0008] Performing labeling processing on the target subgraph dataset to construct a labeled target subgraph dataset;

[0009] Fine-tuning a preset multimodal large language model using the labeled target subgraph dataset to obtain an initial pulsar candidate recognition model;

[0010] The initial pulsar candidate recognition model is subjected to intensive training using a preset reward function and the labeled target subgraph dataset to obtain a pulsar candidate recognition model.

[0011] In one embodiment, the step of obtaining a preset candidate dataset and performing sub-image segmentation processing on each pulsar candidate sample image in the preset candidate dataset to obtain a target sub-image dataset includes:

[0012] For each of the pulsar candidate sample images, according to the target size information of the pulsar candidate sample image, determining a target segmentation template that matches the target size information from a preset segmentation template library;

[0013] Using the target segmentation template, performing sub-image segmentation processing on the pulsar candidate sample image to obtain a corresponding target sub-image;

[0014] A target subgraph dataset is constructed according to the multiple target subgraphs.

[0015] In one embodiment, determining a target segmentation template that matches the target size information from a preset segmentation template library based on the target size information of the pulsar candidate sample image includes:

[0016] determining a target aspect ratio corresponding to the pulsar candidate sample image according to target size information of the pulsar candidate sample image;

[0017] According to the target aspect ratio, a target segmentation template matching the target aspect ratio is determined from the preset segmentation template library.

[0018] In one embodiment, using the target segmentation template to perform sub-image segmentation processing on the pulsar candidate sample image to obtain the corresponding target sub-image includes:

[0019] Obtaining reference size information corresponding to the reference image where the target segmentation template is located;

[0020] determining a scaling factor according to a proportional relationship between the target size information and the reference size information;

[0021] performing scaling mapping processing on the target segmentation template according to the scaling factor to obtain target cropping region coordinates corresponding to the pulsar candidate sample image;

[0022] The target cropping region coordinates are used to perform sub-image segmentation processing on the pulsar candidate sample image to obtain a corresponding target sub-image.

[0023] In one embodiment, the target subgraph dataset includes multiple target subgraphs; the target subgraphs include a frequency phase sample graph and a dispersion curve sample graph; and the labeling processing of the target subgraph dataset to construct the labeled target subgraph dataset includes:

[0024] For each of the frequency-phase sample graphs, authenticity annotation processing is performed on the frequency-phase sample graph according to the frequency-phase characteristics of a real pulsar to obtain an annotated frequency-phase sample graph;

[0025] For each dispersion curve sample graph, performing curve morphology matching annotation processing on the dispersion curve sample graph according to the dispersion curve characteristics of a real pulsar to obtain an annotated dispersion curve sample graph;

[0026] A labeled target subgraph data set is constructed according to the plurality of labeled frequency-phase sample graphs and the plurality of labeled dispersion curve sample graphs.

[0027] In one embodiment, the authenticity labeling processing of the frequency phase sample graph according to the frequency phase characteristics of a real pulsar to obtain the labeled frequency phase sample graph includes:

[0028] Performing feature recognition on the frequency-phase sample image to obtain corresponding candidate frequency-phase features;

[0029] Comparing the frequency and phase characteristics of the candidate with those of the real pulsar to determine the target comparison result;

[0030] Determining the target comparison result as a first target label corresponding to the frequency-phase sample graph;

[0031] A labeled frequency-phase sample graph is obtained according to the first target label and the frequency-phase sample graph.

[0032] In one embodiment, the step of performing curve morphology matching annotation processing on the dispersion curve sample graph according to the dispersion curve characteristics of a real pulsar to obtain the annotated dispersion curve sample graph includes:

[0033] Performing feature recognition on the dispersion curve sample graph to obtain corresponding candidate dispersion curve features;

[0034] Comparing the dispersion curve characteristics of the candidate object with the dispersion curve characteristics of the real pulsar to obtain the target matching degree;

[0035] Determining the target matching degree as a second target label corresponding to the dispersion curve sample graph;

[0036] A labeled dispersion curve sample graph is obtained according to the second target label and the dispersion curve sample graph.

[0037] In one embodiment, fine-tuning a preset multimodal large language model using the labeled target subgraph dataset to obtain an initial pulsar candidate recognition model includes:

[0038] Obtaining a preset multimodal large language model, performing fine-tuning training on the preset multimodal large language model using the labeled target subgraph dataset, and saving a plurality of first checkpoint models according to a first preset period during the fine-tuning training process until the number of iterations of the fine-tuning training meets a first preset iteration threshold;

[0039] The first checkpoint model with the smallest loss function value among the multiple first checkpoint models is determined as the initial pulsar candidate identification model.

[0040] In one embodiment, the use of a preset reward function and the labeled target subgraph dataset to perform intensive training on the initial pulsar candidate recognition model to obtain the pulsar candidate recognition model includes:

[0041] Performing intensive training on the initial pulsar candidate recognition model using a preset reward function and the labeled target subgraph dataset, and saving a plurality of second checkpoint models according to a second preset period during the intensive training process until the number of iterations of the intensive training meets a second preset iteration threshold;

[0042] The second checkpoint model with the smallest loss function value among the plurality of second checkpoint models is determined as the pulsar candidate identification model.

[0043] In one embodiment, the intensive training includes:

[0044] In each of the second preset periods, using the preset reward function, a reward calculation is performed on the initial recognition result output by the current initial pulsar candidate recognition model and the label corresponding to the initial recognition result to obtain a reward result corresponding to the initial recognition result;

[0045] According to the reward result, the model weight of the current initial pulsar candidate identification model is updated, and the updated current initial pulsar candidate identification model is saved as a second checkpoint model.

[0046] In a second aspect, the present application further provides a method for identifying pulsar candidates, the method comprising:

[0047] Obtaining a candidate image corresponding to the candidate pulsar to be identified, and performing sub-image segmentation processing on the candidate image to be identified to obtain a corresponding sub-image to be identified;

[0048] The subgraph to be identified is input into the pulsar candidate identification model described in any one of the embodiments of the first aspect above to perform pulsar candidate identification, and obtain a target identification result corresponding to the pulsar candidate to be identified.

[0049] In one embodiment, the sub-graph to be identified includes a frequency phase graph to be identified and a dispersion curve graph to be identified; the target identification result includes a first identification result corresponding to the frequency phase graph to be identified and a second identification result corresponding to the dispersion curve graph to be identified; the method further includes:

[0050] The pulsar candidate to be identified is classified according to the first recognition result corresponding to the frequency phase diagram to be identified and the second recognition result corresponding to the dispersion curve diagram to be identified, so as to obtain a classification result corresponding to the pulsar candidate to be identified.

[0051] In a third aspect, the present application further provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the method described in any one of the embodiments of the first and second aspects are implemented.

[0052] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in any one of the embodiments of the first and second aspects above.

[0053] In a fifth aspect, the present application further provides a computer program product, comprising a computer program, which, when executed by a processor, implements the steps of the method described in any one of the embodiments of the first and second aspects above.

[0054] The above-mentioned pulsar candidate identification model construction method and pulsar candidate identification method obtain a preset candidate dataset and perform sub-graph segmentation processing on each pulsar candidate sample image in the preset candidate dataset to extract a more representative and learnable target sub-graph dataset, thereby improving data utilization efficiency and model adaptability. The target sub-graph dataset is annotated to obtain a labeled target sub-graph dataset, and the preset multimodal large language model is fine-tuned and trained based on the annotated target sub-graph dataset, so that the model can quickly adapt to the pulsar identification task and obtain an initial pulsar candidate identification model with preliminary recognition capabilities. Further combined with the preset reward function, a reinforcement learning mechanism is introduced to strengthen the training of the initial pulsar candidate identification model. The model decision strategy is continuously adjusted during the training process, which improves the recognition accuracy and robustness of the model in key categories. The resulting pulsar candidate identification model has stronger generalization ability and higher recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following briefly introduces the drawings required for use in the embodiments of the present application or related technical descriptions. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying any creative work.

[0056] Figure 1 A diagram illustrating an application environment of a method for constructing a pulsar candidate identification model in one embodiment;

[0057] Figure 2 1 is a flow chart of a method for constructing a pulsar candidate identification model in one embodiment;

[0058] Figure 3 A schematic diagram of a sample image of a pulsar candidate in one embodiment;

[0059] Figure 4 A schematic diagram of a process for constructing a target subgraph dataset in one embodiment;

[0060] Figure 5 is a schematic diagram of a sample graph of a dispersion curve in one embodiment;

[0061] Figure 6 A schematic diagram of a frequency-phase sample diagram cut out in one embodiment;

[0062] Figure 7 A schematic diagram of a sample graph of a dispersion curve obtained by segmentation in one embodiment;

[0063] Figure 8 Schematic diagram of the process of target subgraph segmentation in one embodiment;

[0064] Figure 9 A schematic flow chart of the steps for constructing an initial pulsar candidate identification model in one embodiment;

[0065] Figure 10 A schematic flow chart of the steps for constructing a pulsar candidate identification model in a specific embodiment;

[0066] Figure 11 1 is a flow chart of a method for identifying pulsar candidates according to an embodiment;

[0067] Figure 12 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0068] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0069] With the construction of advanced radio telescopes, the number of pulsar candidates has exploded. However, due to radio frequency interference and noise, the proportion of true pulsars among these candidates is extremely small, approximately one in ten thousand. Therefore, selecting pulsars based solely on simple metrics such as signal-to-noise ratio is nearly impossible. Pulsar candidate screening is typically performed by human experts examining the candidate's diagnostic charts, but this method is unable to cope with such a large number of candidates. Therefore, the development of an automated, accurate, and efficient method for screening pulsar candidates has become an urgent need.

[0070] To solve the above problems, some previous studies have used the capabilities of image recognition models to automatically screen pulsar candidates. Diagnostic maps of pulsar candidates are important tools for identifying and confirming real pulsars. They are usually generated by the PRESTO (Pulsar Exploration and Search Toolkit) software after processing pulsar observation data. These diagnostic maps mainly include four types: profile maps, time phase maps, frequency phase maps, and dispersion curve maps. These sub-maps show the characteristics of the pulse signal in the dimensions of frequency, time, and phase. These sub-maps can be used to identify real pulsar candidates: (1) The profile map is an intensity-phase curve obtained by averaging the pulse signals of all observation frequencies and time periods. The profile map of most pulsars presents one or more narrow peaks, which is a typical feature of pulsars. (2) The time phase map shows the characteristics of the pulse signal changing with time within a specific frequency bandwidth, presenting the relationship between time and phase in a two-dimensional form, and using color depth to represent the signal amplitude. Real pulsars are usually continuous during the observation time, and appear as one or more vertical stripes in the map. (3) The frequency-phase diagram shows the pulse signals superimposed at different observation frequencies, usually presenting the relationship between frequency and phase in a two-dimensional form, with the depth of color indicating the magnitude of the signal amplitude. Real pulsars usually emit broadband signals, so they appear as continuous bright vertical stripes throughout all frequency subbands in the subband diagram. (4) The dispersion curve diagram represents the electron density integral from the Earth to the pulsar, which is used to determine the best dispersion measurement value. Since real pulsar signals come from outer space, their dispersion curves usually reach a peak at a non-zero dispersion value. This feature is an important basis for verifying pulsar candidates. The sharpness of the dispersion curve is closely related to the duty cycle of the pulsar.

[0071] However, traditional pulsar candidate identification methods are limited by the image information features contained in the training data. Differences in observation equipment, signal processing methods, and time spans among astronomical observation projects result in significant variations in image quality, resolution, and statistical properties among the data generated. This results in poor generalization of models trained on a single dataset to new data sources, significantly reducing recognition accuracy.

[0072] Based on this, in order to solve the problems of poor generalization ability and low accuracy existing in the above-mentioned related technologies, the present application proposes a pulsar candidate identification model construction method and a pulsar candidate identification method.

[0073] The method for constructing a pulsar candidate identification model provided in the embodiment of the present application can be applied to Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store data that server 104 needs to process. The data storage system can be integrated on server 104, or placed on a cloud or other network server. A preset candidate dataset is obtained, and each pulsar candidate sample image in the preset candidate dataset is sub-graph segmented to obtain a target sub-graph dataset; the target sub-graph dataset is annotated to construct an annotated target sub-graph dataset; the annotated target sub-graph dataset is used to fine-tune a preset multimodal large language model to obtain an initial pulsar candidate recognition model; and the initial pulsar candidate recognition model is reinforced with a preset reward function and the annotated target sub-graph dataset to obtain a pulsar candidate recognition model. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and IoT devices. Server 104 can be an independent physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server providing cloud computing services.

[0074] In one embodiment, Figure 2 As shown, Figure 2 This is a flow chart of a method for constructing a pulsar candidate identification model in one embodiment. This embodiment uses the method applied to a terminal as an example. It is understood that the method can also be applied to a server, or to a system including a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0075] Step S201 : obtaining a preset candidate volume data set, performing sub-image segmentation processing on each pulsar candidate sample image in the preset candidate volume data set, and obtaining a target sub-image data set.

[0076] The preset candidate datasets are divided into a training set Dtrain and a test set Dtest. The training set Dtrain includes a pulsar candidate dataset A1 that is already publicly available on the internet and a pulsar candidate dataset B generated based on new astronomical observation data. The test set Dtest includes a pulsar candidate dataset A2 that is homologous to the pulsar candidate dataset A1 but not included in the training set Dtrain, and a pulsar candidate dataset C that is already publicly available on the internet but has a different origin from the pulsar candidate dataset A1.

[0077] It can be understood that the preset candidate data set includes multiple pulsar candidate sample images; for example, Figure 3 As shown, Figure 3 Figure 1 is a schematic diagram of a sample image of a pulsar candidate in one embodiment. The sample image of a pulsar candidate includes multiple sub-images, such as a dispersion curve sample image, a cumulative profile sample image, a frequency phase sample image, and a time phase sample image. The horizontal axis of the dispersion curve sample image is the dispersion value, in units of pc / cm. 3 ; The vertical axis is the dimensionality reduction chi-square value (i.e. ); In the frequency phase sample graph, the left vertical axis is the subband, the right vertical axis is the frequency (unit: MHz), and the horizontal axis is the phase; in the time phase sample graph, the left horizontal axis is the phase, and the right horizontal axis is the dimensionality reduction chi-square value (i.e. ), the vertical axis of the time phase sample diagram is time, in seconds. It should be noted that the specific meanings of other sub-diagrams can be understood by referring to relevant technologies in this field and are not listed here one by one.

[0078] The target subgraph data set includes multiple target subgraphs; the target subgraphs may include, but are not limited to, dispersion curve sample graphs and frequency phase sample graphs.

[0079] It should be noted that the sample images of each pulsar candidate in the preset candidate data set are sub-image segmented, and some sub-images with unclear feature distinctions (such as the cumulative contour sample image and the time phase sample image) are selectively discarded. Only target sub-images with clear pulsar feature expressions (such as the frequency phase sample image and the dispersion curve sample image) are selected for training and recognition. On the one hand, this can effectively improve the speed of candidate recognition, and on the other hand, reduce the interference of sub-images with unclear feature expressions on the recognition results, thereby improving the overall recognition accuracy.

[0080] Step S202 : performing labeling processing on the target subgraph dataset to construct a labeled target subgraph dataset.

[0081] It should be noted that when labeling the target sub-image dataset, it is necessary to identify features of each target sub-image in the dataset. The identified features are then compared with the features corresponding to real pulsars to determine the labels corresponding to each target sub-image. Labels can be categories, attributes, or other information, and need to be set according to actual image processing requirements. This is not a specific limitation here.

[0082] Step S203 : fine-tune the preset multimodal large language model using the labeled target subgraph dataset to obtain an initial pulsar candidate recognition model.

[0083] Among them, the preset multimodal large language model is a pre-trained base model with powerful generation capabilities and cross-modal understanding capabilities.

[0084] As can be seen, the powerful capabilities of the pre-built multimodal large language model enable efficient extraction of numerical features from images. By accurately describing the characteristics of pulsar sub-images, even when new observational data contains morphologies not previously encountered by the trained model, the language model's language understanding capabilities enable excellent candidate identification results, significantly improving the generalization of pulsar candidate identification.

[0085] In an exemplary embodiment, when fine-tuning the preset multimodal large language model, the low-rank adaptation (LoRA) mode can be selected to adjust the low-rank adaptation layer of the model to inject task-specific information, while optimizing the collaborative training of the visual transformer (ViT) backbone, adapter module, and large language model (LLM) parameters to fully adapt to the pulsar recognition task.

[0086] Step S204 : Using the preset reward function and the labeled target subgraph dataset, the initial pulsar candidate recognition model is subjected to intensive training to obtain a pulsar candidate recognition model.

[0087] The preset reward function is used to evaluate the model's recognition quality of the target subgraph and provides numerical feedback (e.g., a score) to guide the model's learning direction. The preset reward function can be, but is not limited to, set by domain experts based on actual task requirements and experience, or constructed based on data statistical characteristics. This is not specifically limited here. The core design concept of the preset reward function is to assign different penalty weights to different types of recognition errors in different tasks to reflect the importance and impact of each misjudgment within the task.

[0088] For example, in a binary classification task, the penalty for misclassifying "yes" as "no" is greater than the penalty for misclassifying "no" as "yes." For another example, in a multi-class classification task, an ordered relationship between categories is introduced, where the farther the misclassification is from the true category, the greater the penalty. Approximate judgments are partially encouraged to improve the robustness of the model under uncertainty. Specifically, for samples labeled "high," if the model predicts "medium," it is considered approximately reasonable and no penalty is imposed; if the prediction is "low," a larger penalty is imposed; for samples labeled "medium," a smaller penalty is imposed when the prediction is "high," and a larger penalty is imposed when the prediction is "low"; for samples labeled "low," the maximum penalty is imposed when the prediction is "high." This overall mechanism encourages the model to stay as close to the true category as possible under uncertainty, avoiding extreme misjudgments.

[0089] The preset reward function constructed in this way not only reflects the actual requirements of the task but also helps balance precision and recall during model training. Understandably, in pulsar identification tasks, a high recall is often more critical, as missing a potential true pulsar signal could mean missing out on important scientific discoveries.

[0090] Therefore, the initial pulsar candidate recognition model is trained intensively using a preset reward function and annotated target subgraph dataset. This can effectively guide the model to learn the expected behavior strategy without explicit supervision labels, and ultimately obtain a pulsar candidate recognition model with better performance and stronger generalization ability.

[0091] In this embodiment, a preset candidate dataset is obtained and sub-graph segmentation processing is performed on each pulsar candidate sample image in the preset candidate dataset to extract a more representative and learnable target sub-graph dataset, thereby improving data utilization efficiency and model adaptability. The target sub-graph dataset is annotated to obtain a labeled target sub-graph dataset, and a preset multimodal large language model is fine-tuned based on the annotated target sub-graph dataset to enable the model to quickly adapt to the pulsar recognition task, thereby obtaining an initial pulsar candidate recognition model with preliminary recognition capabilities. Furthermore, in combination with a preset reward function, a reinforcement learning mechanism is introduced to perform intensive training on the initial pulsar candidate recognition model. The model decision strategy is continuously adjusted during the training process, thereby improving the recognition accuracy and robustness of the model in key categories. The resulting pulsar candidate recognition model has stronger generalization capabilities and higher recognition accuracy.

[0092] In one embodiment, Figure 4 As shown, Figure 4The flowchart of the target sub-graph dataset construction step in one embodiment is as follows: obtaining a preset candidate dataset, performing sub-graph segmentation processing on each pulsar candidate sample image in the preset candidate dataset, and obtaining the target sub-graph dataset includes the following steps:

[0093] Step S401 : for each pulsar candidate sample image, according to the target size information of the pulsar candidate sample image, a target segmentation template that matches the target size information is determined from a preset segmentation template library.

[0094] The target size information includes the height and width of the sample image of the pulsar candidate.

[0095] The preset segmentation template library includes multiple templates. These templates are designed based on the image styles generated by PRESTO, a preprocessing software commonly used in astronomical pulsar searches. Because different versions of PRESTO may produce different image styles when drawing candidate sub-images, it is necessary to design corresponding segmentation templates for various pulsar candidate sample images based on the image styles of different versions, thus constructing a preset segmentation template library.

[0096] Furthermore, the sample images of the pulsar candidates in the pre-set candidate dataset come from diverse sources, resulting in varying resolutions. However, each image contains target size information, which can be used for mapping and matching to determine the most appropriate segmentation template, ensuring consistency and accuracy in subsequent processing.

[0097] Furthermore, when creating the segmentation template, the influence of the sub-image coordinate axis on the image features is also considered. For example, if the frequency-phase sample graph needs to determine whether there is a black vertical line, then the coordinate axis may affect the model's judgment of the result. Therefore, the coordinate axis needs to be set to white when segmenting the target sub-image.

[0098] Furthermore, for the dispersion curve sample graph, since some dispersion peaks are very close to the far left, the scale lines need to be retained, otherwise the peaks may be cut off and the pulsar characteristics may be lost, which may lead to incorrect model judgment. Figure 5 Take the dispersion curve sample graph as an example. The peak of the dispersion curve sample graph is close to the leftmost side. If the scale line is cut off, the remaining area will show a monotonically decreasing trend, which may lead to misjudgment of the model. Figure 5 The horizontal axis is the dispersion value, the unit is pc / cm 3 ; The vertical axis is the dimensionality reduction chi-square value.

[0099] Step S402 : Using the target segmentation template, the pulsar candidate sample image is segmented into sub-images to obtain corresponding target sub-images.

[0100] The target segmentation template is used to segment target sub-images from the pulsar candidate sample image. It is understood that the target segmentation template defines the coordinate locations of the sub-image division regions; target sub-images may include, but are not limited to, frequency-phase sample images and dispersion curve sample images.

[0101] For example, see Figure 6 and Figure 7 , using the target segmentation template, the pulsar candidate sample image is segmented into sub-images to obtain the corresponding frequency phase sample image ( Figure 6 ) and dispersion curve sample diagram ( Figure 7 It should be noted that in order to avoid the influence of the sub-graph coordinate axis outline on the image features, the coordinate axis outline needs to be set to white; at the same time, in order to ensure the integrity of the pulsar characteristics corresponding to the dispersion curve sample graph, the corresponding scale lines are retained.

[0102] Step S403: construct a target subgraph dataset according to the multiple target subgraphs.

[0103] The target sub-image dataset includes a frequency-phase image sub-image dataset and a dispersion curve image sub-image dataset.

[0104] Exemplarily, based on multiple frequency-phase sample graphs and multiple dispersion curve sample graphs, a frequency-phase graph subgraph dataset Dtrainp for training, a dispersion curve graph subgraph dataset Dtraind for training, a frequency-phase graph subgraph dataset Dtestp for testing, and a dispersion curve graph subgraph dataset Dtestd for testing are constructed.

[0105] In this embodiment, for each pulsar candidate sample image, a target segmentation template is adaptively matched from a preset segmentation template library based on its target size information. This method then accurately segments the pulsar candidate sample images into sub-images based on the target segmentation templates. This effectively eliminates interference caused by varying image sources, resolutions, and image style, improving data standardization. The resulting high-quality target sub-image dataset not only enhances the consistency and comparability of image analysis but also provides more efficient and reliable data support for subsequent pulsar signal identification and classification, significantly improving processing efficiency and algorithm generalization.

[0106] In one embodiment, based on the target size information of the pulsar candidate sample image, determining a target segmentation template that matches the target size information from a preset segmentation template library includes the following steps:

[0107] Step 1: Determine the target aspect ratio corresponding to the pulsar candidate sample image according to the target size information of the pulsar candidate sample image.

[0108] The target aspect ratio is the ratio of the width to the height of the sample image of the pulsar candidate.

[0109] Step 2: According to the target aspect ratio, a target segmentation template that matches the target aspect ratio is determined from a preset segmentation template library.

[0110] It should be noted that each aspect ratio has a unique corresponding segmentation template.

[0111] For example, the target size information of the pulsar candidate sample image is parsed to determine the height and width of the pulsar candidate sample image. The target aspect ratio is calculated based on the height and width of the pulsar candidate sample image, and the result is rounded to two decimal places. The target aspect ratio calculation formula is round(width / height, 2). Furthermore, based on the target aspect ratio, a target segmentation template that matches the target aspect ratio is determined from a preset segmentation template library.

[0112] In this embodiment, based on the target aspect ratio corresponding to the pulsar candidate sample image, the target segmentation template that matches the target aspect ratio can be accurately determined from the preset segmentation template library, thereby enhancing the compatibility with images from different sources and different resolutions, and laying the foundation for improving the accuracy and processing efficiency of sub-image segmentation.

[0113] In one embodiment, Figure 8 As shown, Figure 8 FIG. 1 is a flow chart of the target subgraph segmentation step in one embodiment. The target segmentation template is used to perform subgraph segmentation processing on the sample image of the pulsar candidate to obtain the corresponding target subgraph, including the following steps:

[0114] Step S801: Obtain reference size information corresponding to the reference image where the target segmentation template is located.

[0115] The reference image refers to a standard image used as a basis for designing or defining a segmentation template; it provides a reference for scaling when subsequently cropping images of different resolutions. Reference size information includes the height and width of the reference image. It is understood that the target segmentation template is located in a specific area of ​​the reference image.

[0116] For example, assuming that when designing a segmentation template corresponding to an aspect ratio of 1.29, the design can be based on a reference image with a width of 2200 pixels and a height of 1700 pixels.

[0117] Step S802: determining a scaling factor according to a proportional relationship between the target size information and the standard size information.

[0118] The scaling factor is used to determine the degree of enlargement or reduction of the target segmentation template. It is understood that based on the scaling factor, the fixed coordinates defined in the target segmentation template can be denormalized and mapped according to the scaling factor to achieve adaptive cropping of images with different resolutions but specific aspect ratios.

[0119] Exemplarily, according to the proportional relationship between the target size information and the reference size information, a method for determining the scaling factor may be: determining the ratio of the width corresponding to the reference image to the width corresponding to the pulsar candidate sample image as the scaling factor.

[0120] Step S803 : performing scaling mapping processing on the target segmentation template according to the scaling factor to obtain the target cropping region coordinates corresponding to the pulsar candidate sample image.

[0121] The target cropping area coordinates include the target cropping area horizontal coordinate range and the target cropping area vertical coordinate range. In an exemplary embodiment, the target cropping area coordinates are expressed as [y1:y2, x1:x2], where (y1:y2) is the target cropping area vertical coordinate range and (x1:x2) is the target cropping area horizontal coordinate range.

[0122] Step S804 : Using the target cropping region coordinates, the pulsar candidate sample image is sub-divided into sub-images to obtain corresponding target sub-images.

[0123] For example, if the target aspect ratio is 1.29, the scaling factor cc = 2200 / width. Based on the scaling factor cc, the target cropping region coordinates corresponding to the pulsar candidate sample image are [int(1278 / cc):int(1614 / cc), int(861 / cc):int(1440 / cc)]. Where width is the width of the pulsar candidate sample image, 2200 is the width of the reference image where the target segmentation template is located, and the coordinate range [1278:1614,861:1440] defines the coordinate position of the target segmentation template in the reference image.

[0124] If the target aspect ratio is 1.30, the scaling factor cc = 1420 / width. Based on the scaling factor cc, the target cropping region coordinates corresponding to the pulsar candidate sample image are defined as [int(815 / cc):int(1040 / cc), int(545 / cc):int(939 / cc)]. Where width is the width of the pulsar candidate sample image, 1420 is the width of the reference image where the target segmentation template is located, and the coordinate range [815:1040, 545:939] defines the coordinate position of the target segmentation template in the reference image.

[0125] It's understandable that when two pulsar candidate sample images have the same target aspect ratio but different resolutions, sub-image segmentation can be performed based on the same target segmentation template. Since their resolutions are proportional (e.g., 2x), their target cropping region coordinates also follow the same scaling ratio. This approach not only simplifies the processing flow but also ensures consistency and comparability between sub-images extracted at different resolutions. For example, suppose there are two pulsar candidate sample images, both with a target aspect ratio of 1.30 but with a 2x resolution ratio. Assume that the target cropping region coordinates for the lower-resolution pulsar candidate sample image are [y1:y2, x1:x2], and the target cropping region coordinates for the lower-resolution pulsar candidate sample image are [2y1:2y2, 2x1:2x2].

[0126] In this embodiment, adaptive cropping of images of varying resolutions but with specific aspect ratios is achieved by calculating a corresponding scaling factor cc based on the actual image width and denormalizing the fixed coordinates defined in the template according to this scaling factor. This approach ensures that subimages extracted from different images maintain consistency in content layout, relative position, and scale, thereby improving the standardization of dataset construction and the effectiveness of subsequent model training.

[0127] In one embodiment, a target subgraph dataset includes a plurality of target subgraphs; the target subgraphs include a frequency phase sample graph and a dispersion curve sample graph; and labeling the target subgraph dataset to construct the labeled target subgraph dataset includes the following steps:

[0128] Step 1: For each frequency-phase sample graph, authenticity annotation processing is performed on the frequency-phase sample graph according to the frequency-phase characteristics of the real pulsar to obtain the annotated frequency-phase sample graph.

[0129] Among them, the frequency phase characteristics of real pulsars are manifested in the sub-band diagram as continuous bright vertical stripes throughout all frequency sub-bands.

[0130] Exemplarily, authenticity labeling processing of the frequency phase sample graph may be: performing binary classification labeling processing on the frequency phase sample graph, and the labeling content may be a binary classification label, such as "yes" or "no", where "yes" indicates that the frequency phase sample graph displays the frequency phase characteristics of a real pulsar; "no" indicates that the frequency phase sample graph does not display the frequency phase characteristics of a real pulsar.

[0131] Step 2: For each dispersion curve sample graph, the dispersion curve sample graph is annotated according to the dispersion curve characteristics of the real pulsar to obtain a labeled dispersion curve sample graph.

[0132] Among them, the dispersion curve characteristics of real pulsars are manifested as reaching a peak at a non-zero dispersion value, and there is obvious wavelength division, with an upward trend on the left side of the peak and a downward trend on the right side.

[0133] Exemplarily, the curve morphology matching degree annotation processing of the dispersion curve sample graph can be: performing grade classification annotation processing on the dispersion curve sample graph, and the annotation content can be a grade classification label, such as "high", "medium" or "low"; wherein, "high" indicates that the dispersion curve sample graph has a relatively obvious peak, an upward trend on the left side of the peak, a downward trend on the right side of the peak, and the dispersion value at the maximum fitting value is not 0; "medium" indicates that the dispersion curve sample graph is such as a monotonic upward trend as a whole, there is a peak but the dispersion value at the maximum fitting value is 0, the dispersion curve fluctuates frequently with 0 or more relatively obvious peaks, etc.; "low" indicates that the dispersion curve sample graph is a monotonic downward trend as a whole, or is a horizontal trend as a whole.

[0134] Step 3: Construct a labeled target subgraph dataset based on the multiple labeled frequency-phase sample graphs and the multiple labeled dispersion curve sample graphs.

[0135] In this example, by labeling the authenticity of frequency-phase sample images and the morphology matching of dispersion curve sample images, the interpretability and structure of the pulsar candidate image data were effectively improved, ensuring data quality and accuracy. The resulting labeled target sub-image dataset not only enhanced the model's ability to classify true and false pulsars, but also supported multimodal fusion modeling, improving the robustness and generalization of the automated recognition system.

[0136] In an exemplary embodiment, in step 1, authenticity labeling processing is performed on the frequency phase sample graph according to the frequency phase characteristics of a real pulsar to obtain the labeled frequency phase sample graph, including the following steps:

[0137] Step 1.1: Perform feature recognition on the frequency-phase sample image to obtain the corresponding candidate frequency-phase features.

[0138] The method for performing feature recognition on the frequency-phase sample graph may be implemented by using an existing feature recognition method, which is not specifically limited here.

[0139] The candidate frequency phase features include the stripe characteristics under all frequency sub-bands.

[0140] Step 1.2: Compare the frequency and phase characteristics of the candidate with those of the real pulsar to determine the target comparison result.

[0141] The similarity comparison method may be, but is not limited to, performing a quantitative comparison of two frequency-phase features through a mathematical method to evaluate the degree of matching between them, which is not specifically limited here.

[0142] In an exemplary embodiment, since some sub-band ranges may be blank after PRESTO interference removal, the comparison standard in this embodiment is set to whether there is a black vertical line covering most of the sub-band area.

[0143] Step 1.3: Determine the target comparison result as the first target label corresponding to the frequency-phase sample image.

[0144] Step 1.4: Obtain a labeled frequency-phase sample graph based on the first target label and the frequency-phase sample graph.

[0145] Exemplarily, for each frequency phase sample graph, feature recognition is performed on the frequency phase sample graph to obtain the corresponding candidate frequency phase feature; the candidate frequency phase feature is compared with the frequency phase feature of the real pulsar for similarity; if the candidate frequency phase feature meets the frequency phase feature of the real pulsar, the target comparison result is determined to be "yes"; if the candidate frequency phase feature does not meet the frequency phase feature of the real pulsar, the target comparison result is determined to be "no"; the target comparison result is further determined as the first target label corresponding to the frequency phase sample graph, and the labeled frequency phase sample graph is obtained according to the first target label and the frequency phase sample graph, and so on, all frequency phase sample graphs are labeled.

[0146] In an exemplary embodiment, step 2 includes performing curve morphology matching annotation processing on the dispersion curve sample graph according to the dispersion curve characteristics of the real pulsar to obtain the annotated dispersion curve sample graph, which includes the following steps:

[0147] Step 2.1: Perform feature recognition on the dispersion curve sample graph to obtain the corresponding candidate dispersion curve features.

[0148] The method for performing feature recognition on the dispersion curve sample graph may be implemented by using an existing feature recognition method, which is not specifically limited here.

[0149] Among them, the dispersion curve characteristics of the candidate include the dispersion curve change trend and the peak maximum dispersion value.

[0150] Step 2.2: Compare the dispersion curve characteristics of the candidate with those of the real pulsar to obtain the target matching degree.

[0151] Among them, the curve matching comparison method can be, but is not limited to, comparing the dispersion curve characteristics of the candidate body with the dispersion curve characteristics of the real pulsar through mathematical or algorithmic means to quantify the similarity in their shapes, trends, parameters, etc.

[0152] Among them, the target matching degree is used to characterize the similarity between the dispersion curve characteristics of the candidate object and the dispersion curve characteristics of the real pulsar.

[0153] In step 2.3, the target matching degree is determined as the second target label corresponding to the dispersion curve sample graph.

[0154] Step 2.4: Obtain a labeled dispersion curve sample graph based on the second target label and the dispersion curve sample graph.

[0155] Exemplarily, for each dispersion curve sample graph, feature recognition is performed on the dispersion curve sample graph to obtain the corresponding candidate dispersion curve feature, and the curve matching degree of the candidate dispersion curve feature is compared with the dispersion curve feature of the real pulsar; if the candidate dispersion curve feature satisfies the first preset form, the target matching degree is "high"; if the candidate dispersion curve feature satisfies the second preset form, the target matching degree is "medium"; if the candidate dispersion curve feature satisfies the third preset form, the target matching degree is "low"; wherein, the first preset form indicates that the dispersion curve sample graph has a relatively obvious peak, an upward trend on the left side of the peak, a downward trend on the right side, and the dispersion value at the maximum fitting value is not 0; the second preset form indicates that the dispersion curve sample graph is monotonically rising as a whole, has a peak but the dispersion value at the maximum fitting value is 0, the dispersion curve fluctuates frequently with 0 or more relatively obvious peaks, etc.; the third preset form indicates that the dispersion curve sample graph is monotonically declining as a whole, or is horizontal as a whole. Then, the target matching degree is determined as the second target label corresponding to the dispersion curve sample graph; based on the second target label and the dispersion curve sample graph, a labeled dispersion curve sample graph is obtained; and similarly, all dispersion curve sample graphs are labeled.

[0156] It should be noted that the dispersion curves of most real pulsars exhibit distinct characteristics: a relatively distinct peak, an upward trend to the left and a downward trend to the right, and a non-zero dispersion value at the point of maximum fitting. However, a few real pulsar signals may exhibit anomalies such as monotonically increasing dispersion curves or less distinct peaks. To address this, the "Medium" category is introduced. When recall is important, ensuring that no real pulsar signals are missed, all candidates classified as "Yes" in the frequency phase diagram and "High" or "Medium" in the dispersion curve diagram are considered positive. When precision is important, ensuring that too many false positives are avoided, only candidates classified as "Yes" in the frequency phase diagram and "High" in the dispersion curve diagram are considered positive, thus achieving a flexible balance between recall and precision.

[0157] In one embodiment, Figure 9 As shown, Figure 9 The figure is a flowchart of the steps for constructing an initial pulsar candidate identification model in one embodiment. The method uses the labeled target subgraph dataset to fine-tune the preset multimodal large language model to obtain the initial pulsar candidate identification model, including the following steps:

[0158] Step S901: obtain a preset multimodal large language model, use the labeled target subgraph dataset to fine-tune the preset multimodal large language model, and save multiple first checkpoint models according to a first preset period during the fine-tuning training process until the number of iterations of the fine-tuning training meets the first preset iteration threshold.

[0159] Among them, the first preset period refers to a fixed number of rounds or iterations; the first preset period can be set according to actual training needs and is not specifically limited here.

[0160] The first checkpoint model refers to the preset multimodal large language model state saved every first preset period.

[0161] The first preset iteration threshold refers to the maximum number of iterations or training rounds in the entire fine-tuning training process, which controls the training termination condition.

[0162] Step S902 : determining the first checkpoint model with the smallest loss function value among the multiple first checkpoint models as the initial pulsar candidate identification model.

[0163] For example, assuming the first preset period is to save the model every 5 epochs and the first preset iteration threshold is 50 epochs, 10 first checkpoint models will be generated during the entire fine-tuning training process. The first checkpoint model with the smallest loss function value among the 10 first checkpoint models is determined as the initial pulsar candidate identification model.

[0164] In this embodiment, by saving multiple first-checkpoint models throughout the fine-tuning training cycle and selecting the best-performing one as the initial pulsar candidate recognition model, this ensures that the initial pulsar candidate recognition model has optimal generalization and prediction accuracy, helping to avoid premature training cessation or missing optimal model parameters. Furthermore, fine-tuning the pre-set multimodal large language model using the annotated target subgraph dataset allows the model to better adapt to specific task requirements (i.e., pulsar candidate recognition). This approach not only inherits the powerful feature extraction capabilities of the pre-set multimodal large language model, but also enhances the model's understanding and processing capabilities for this domain data through targeted fine-tuning.

[0165] In one embodiment, Figure 10 As shown, Figure 10 The figure is a flowchart of the steps for building a pulsar candidate recognition model in a specific embodiment. The initial pulsar candidate recognition model is subjected to intensive training using a preset reward function and a labeled target subgraph dataset to obtain a pulsar candidate recognition model, including the following steps:

[0166] In step S1001, the initial pulsar candidate recognition model is subjected to intensive training using a preset reward function and annotated target subgraph dataset, and multiple second checkpoint models are saved according to a second preset period during the intensive training process until the number of iterations of the intensive training meets a second preset iteration threshold.

[0167] The preset reward function is used to evaluate the model's recognition quality of the target subgraph and provides numerical feedback (e.g., a score) to guide the model's learning direction. The preset reward function can be, but is not limited to, set by domain experts based on actual task requirements and experience, or constructed based on data statistical characteristics. This is not specifically limited here. The core design concept of the preset reward function is to assign different penalty weights to different types of recognition errors in different tasks to reflect the importance and impact of each misjudgment within the task.

[0168] Among them, the second preset period refers to a fixed number of rounds or iterations; the second preset period can be set according to actual training needs and is not specifically limited here.

[0169] The second checkpoint model refers to the initial pulsar candidate identification model state saved every second preset period.

[0170] Among them, the second preset iteration threshold refers to the maximum number of iterations or training rounds of the entire reinforcement training process, which controls the training termination condition.

[0171] Step S1002 : determining the second checkpoint model with the smallest loss function value among the plurality of second checkpoint models as the pulsar candidate identification model.

[0172] In an exemplary embodiment, the intensive training includes the following steps:

[0173] Step 1: In each second preset period, a preset reward function is used to calculate the reward for the initial recognition result output by the current initial pulsar candidate recognition model and the label corresponding to the initial recognition result to obtain the reward result corresponding to the initial recognition result.

[0174] The current initial pulsar candidate identification model refers to the initial pulsar candidate identification model corresponding to the second preset period during the intensive training process.

[0175] The initial recognition result includes a first initial recognition result corresponding to the frequency-phase sample graph and a second initial recognition result corresponding to the dispersion curve sample graph.

[0176] The labels corresponding to the initial recognition results include the first target labels corresponding to the frequency-phase sample graph and the second target labels corresponding to the dispersion curve sample graph.

[0177] Among them, the reward result is a numerical feedback signal used to guide the direction of reinforcement learning.

[0178] Exemplarily, the method of calculating the reward for the initial recognition result output by the current initial pulsar candidate recognition model and the label corresponding to the initial recognition result using a preset reward function can be: for a frequency-phase sample diagram with a first target label of "yes", if the first initial recognition result is "yes", 1 point is obtained, otherwise 1 point is deducted; for a frequency-phase sample diagram with a first target label of "no", if the first initial recognition result is "no", 1 point is obtained, otherwise 0.5 points are deducted. For the dispersion curve sample graph with the second target label of "high", if the second initial recognition result is "high", 1 point is obtained, if the second initial recognition result is "medium", no point is obtained, and if the second initial recognition result is "low", 1 point is deducted; for the dispersion curve sample graph with the second target label of "medium", if the second initial recognition result is "high", 0.5 points are obtained, if the second initial recognition result is "medium", 1 point is obtained, and if the second initial recognition result is "low", 0.5 points are deducted; for the dispersion curve sample graph with the second target label of "low", if the second initial recognition result is "high", 1 point is deducted, if the second initial recognition result is "medium", 0.5 points are deducted, and if the second initial recognition result is "low", 1 point is obtained.

[0179] Step 2: Update the model weight of the current initial pulsar candidate identification model according to the reward result, and save the updated current initial pulsar candidate identification model as the second checkpoint model.

[0180] It is understandable that updating the model weights of the current initial pulsar candidate identification model based on the reward results can improve its ability to make better judgments in the next round.

[0181] In this embodiment, during each second preset period, a preset reward function is used to compare and analyze the initial recognition results output by the current initial pulsar candidate recognition model with their corresponding labels, and a reward result reflecting the model's discrimination quality is calculated. This mechanism enables the model to continuously adjust its decision-making strategy within a reinforcement learning framework, accurately capturing the subtle differences between real pulsars and non-candidates. It exhibits particularly strong adaptability when dealing with boundary samples and complex interference samples, significantly improving overall recognition accuracy and robustness, and providing technical support for more generalized automated pulsar screening systems.

[0182] In one embodiment, Figure 11 As shown, a method for identifying pulsar candidates is provided. This embodiment uses the method applied to a terminal as an example for illustration. It is understandable that the method can also be applied to a server, or to a system including a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0183] Step S1101 : obtaining a candidate image corresponding to a pulsar candidate to be identified, performing sub-image segmentation processing on the candidate image to be identified, and obtaining a corresponding sub-image to be identified.

[0184] The pulsar candidates to be identified may be, but are not limited to, those generated by the tool PRESTO, and are not specifically limited here.

[0185] It's worth noting that PRESTO is one of the most widely used open-source software tools in the field of astronomical pulsar searches. This toolset is widely used to process time-series data from radio telescope observations, offering efficient data preprocessing, de-dispersion, period search, and candidate generation capabilities.

[0186] After completing data processing for an observation mission, PRESTO automatically generates a set of potential pulsar candidates and records relevant metadata in a designated mission status file, including but not limited to the mission identifier; data processing completion time; output storage path; and the directory structure of candidate signature files. This embodiment periodically polls this status file to detect newly generated candidate missions in real time and automatically trigger the subsequent identification process, achieving a seamless transition from data processing to intelligent identification.

[0187] PRESTO generates pulsar candidates by processing time series data observed by radio telescopes. The process involves removing radio frequency interference (RFI) from the radio spectrum, de-dispersion processing the observed data to correct for time delays between signals of different frequencies, and then performing frequency domain analysis of the time series using a fast Fourier transform (FFT) to extract periodic signals. Significant periodic signals are then screened through methods such as period doubling, frequency doubling search, and pulse profile optimization. Ultimately, key features are output as candidate data, including cumulative profile plots, time phase plots, frequency phase plots, and dispersion curves.

[0188] It should be noted that the method of performing sub-image segmentation processing on the candidate body image to be identified to obtain the corresponding sub-image to be identified is the same as the principle of the sub-image segmentation processing method described in the above embodiment, and will not be repeated here.

[0189] Step S1102 : inputting the subgraph to be identified into the pulsar candidate identification model described in any of the above embodiments to perform pulsar candidate identification, and obtaining a target identification result corresponding to the pulsar candidate to be identified.

[0190] The sub-graph to be identified includes a frequency phase graph to be identified and a dispersion curve graph to be identified; the target identification result includes a first identification result corresponding to the frequency phase graph to be identified and a second identification result corresponding to the dispersion curve graph to be identified.

[0191] In this embodiment, pulsar candidate identification is performed on the subgraph to be identified based on the pulsar candidate identification model, thereby improving the accuracy of the identification result.

[0192] In one embodiment, the pulsar candidate identification method further includes the following steps:

[0193] According to the first recognition result corresponding to the frequency phase diagram to be identified and the second recognition result corresponding to the dispersion curve diagram to be identified, the pulsar candidate to be identified is classified to obtain a classification result corresponding to the pulsar candidate to be identified.

[0194] It should be noted that, based on different combinations of the first identification result corresponding to the frequency phase diagram to be identified and the second identification result corresponding to the dispersion curve diagram to be identified, the pulsar candidate to be identified is classified, and the classification result corresponding to the pulsar candidate to be identified is one of the following six types:

[0195] Type 1: The frequency-phase diagram of the candidate shows the frequency-phase characteristics corresponding to a real pulsar, and the dispersion curve of the candidate closely matches the dispersion curve characteristics of a real pulsar. Pulsar candidates corresponding to this type are very likely to be real pulsar signals and should be verified first.

[0196] Type 2: The frequency phase diagram of the candidate shows the frequency phase characteristics corresponding to a real pulsar, and the dispersion curve diagram of the candidate shows a moderate match with the dispersion curve characteristics of the real pulsar. Pulsar candidates corresponding to this type are likely to be real pulsar signals and should be verified with the second highest priority.

[0197] Type 3: The frequency-phase diagram of the target pulsar shows the frequency-phase characteristics corresponding to a real pulsar, and the dispersion curve of the target pulsar has a poor match with the dispersion curve characteristics of the real pulsar. A very small number of real pulsar signals have no peaks in their dispersion curves. Further investigation is recommended after completing the first two types of verification.

[0198] Type 4: The frequency-phase diagram of the target pulsar does not display the frequency-phase characteristics corresponding to a real pulsar, but the dispersion curve of the target pulsar closely matches the dispersion curve of a real pulsar. A very small number of real pulsar signals have very shallow vertical lines in their frequency-phase diagrams, which can easily lead to misidentification. Further investigation is recommended after completing the first two types of verification.

[0199] Type 5: The frequency-phase diagram of the candidate does not show the frequency-phase characteristics corresponding to a real pulsar, but the dispersion curve diagram of the candidate has a moderate match with the dispersion curve characteristics of a real pulsar. The pulsar candidate corresponding to this type is unlikely to be a real pulsar signal and generally does not require verification.

[0200] Type 6: The frequency-phase diagram of the target pulsar does not show the frequency-phase characteristics corresponding to the real pulsar, but the dispersion curve diagram of the target pulsar has a low match with the dispersion curve characteristics of the real pulsar, so no verification is required.

[0201] Through the above steps, automatic classification of pulsar candidates can be achieved, thus providing reliable basic data support for subsequent in-depth analysis.

[0202] In one exemplary embodiment, a pulsar candidate identification system is provided. The system integrates monitoring capabilities. For example, the system outputs monitoring information to a log file every second and presents it in a visual format. The monitoring information includes the following: system CPU usage; system CPU memory usage; system GPU usage; system GPU memory usage; system processing efficiency (i.e., the number of images processed per second); and the time required for the system to process a pulsar candidate in each observation.

[0203] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above can include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0204] In an exemplary embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as shown in FIG. Figure 12 As shown. The computer device includes a processor, memory, an input / output (I / O) interface, and a communication interface. The processor, memory, and I / O interface are connected via a system bus, and the communication interface is connected to the system bus via the I / O interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the computer device is used to store data related to the construction of a pulsar candidate identification model and pulsar candidate identification. The I / O interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with external terminals via a network connection. When executed by the processor, the computer program implements a method for constructing a pulsar candidate identification model and a method for identifying pulsar candidates.

[0205] Those skilled in the art will understand that Figure 12 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0206] In one embodiment, a computer device is further provided, including a memory and a processor. The memory stores a computer program, and the processor implements the steps in the above method embodiments when executing the computer program.

[0207] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0208] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in the above method embodiments when executed by a processor.

[0209] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0210] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. In particular, any reference to memory, database, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The databases involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the various embodiments provided herein may be, but are not limited to, general-purpose processors, central processing units (CPUs), graphics processing units (GPUs), digital signal processors (DSPs), programmable logic devices (PLDs), quantum computing-based data processing logic devices, artificial intelligence (AI) processors, and the like.

[0211] The technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0212] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. A method for constructing a pulsar candidate identification model, characterized in that: The method comprises: Obtaining a preset candidate data set, performing sub-image segmentation processing on each pulsar candidate sample image in the preset candidate data set to obtain a target sub-image data set; Performing labeling processing on the target subgraph dataset to construct a labeled target subgraph dataset; Fine-tuning a preset multimodal large language model using the labeled target subgraph dataset to obtain an initial pulsar candidate recognition model; The initial pulsar candidate recognition model is subjected to intensive training using a preset reward function and the labeled target subgraph dataset to obtain a pulsar candidate recognition model.

2. The method according to claim 1, characterized in that The step of obtaining a preset candidate dataset and performing sub-image segmentation processing on each pulsar candidate sample image in the preset candidate dataset to obtain a target sub-image dataset includes: For each of the pulsar candidate sample images, according to the target size information of the pulsar candidate sample image, determining a target segmentation template that matches the target size information from a preset segmentation template library; Using the target segmentation template, performing sub-image segmentation processing on the pulsar candidate sample image to obtain a corresponding target sub-image; A target subgraph dataset is constructed according to the multiple target subgraphs.

3. The method according to claim 2, characterized in that The step of determining, based on the target size information of the pulsar candidate sample image, a target segmentation template that matches the target size information from a preset segmentation template library includes: determining a target aspect ratio corresponding to the pulsar candidate sample image according to target size information of the pulsar candidate sample image; According to the target aspect ratio, a target segmentation template matching the target aspect ratio is determined from the preset segmentation template library.

4. The method according to claim 3, characterized in that The method of performing sub-image segmentation processing on the pulsar candidate sample image by using the target segmentation template to obtain a corresponding target sub-image includes: Obtaining reference size information corresponding to the reference image where the target segmentation template is located; determining a scaling factor according to a proportional relationship between the target size information and the reference size information; performing scaling mapping processing on the target segmentation template according to the scaling factor to obtain target cropping region coordinates corresponding to the pulsar candidate sample image; The target cropping region coordinates are used to perform sub-image segmentation processing on the pulsar candidate sample image to obtain a corresponding target sub-image.

5. The method according to claim 1, wherein The target subgraph dataset includes a plurality of target subgraphs; the target subgraphs include a frequency phase sample graph and a dispersion curve sample graph; and the target subgraph dataset is annotated to construct the annotated target subgraph dataset, including: For each of the frequency-phase sample graphs, authenticity annotation processing is performed on the frequency-phase sample graph according to the frequency-phase characteristics of a real pulsar to obtain an annotated frequency-phase sample graph; For each dispersion curve sample graph, performing curve morphology matching annotation processing on the dispersion curve sample graph according to the dispersion curve characteristics of a real pulsar to obtain an annotated dispersion curve sample graph; A labeled target subgraph data set is constructed according to the plurality of labeled frequency-phase sample graphs and the plurality of labeled dispersion curve sample graphs.

6. The method according to claim 5, characterized in that The authenticity marking process is performed on the frequency phase sample graph according to the frequency phase characteristics of the real pulsar to obtain the marked frequency phase sample graph, including: Performing feature recognition on the frequency-phase sample image to obtain corresponding candidate frequency-phase features; Comparing the frequency and phase characteristics of the candidate with those of the real pulsar to determine the target comparison result; Determining the target comparison result as a first target label corresponding to the frequency-phase sample graph; A labeled frequency-phase sample graph is obtained according to the first target label and the frequency-phase sample graph.

7. The method according to claim 5, characterized in that The step of performing curve morphology matching annotation processing on the dispersion curve sample graph according to the dispersion curve characteristics of the real pulsar to obtain the annotated dispersion curve sample graph includes: Performing feature recognition on the dispersion curve sample graph to obtain corresponding candidate dispersion curve features; Comparing the dispersion curve characteristics of the candidate object with the dispersion curve characteristics of the real pulsar to obtain the target matching degree; Determining the target matching degree as a second target label corresponding to the dispersion curve sample graph; A labeled dispersion curve sample graph is obtained according to the second target label and the dispersion curve sample graph.

8. The method according to claim 1, characterized in that The method of fine-tuning a preset multimodal large language model using the labeled target subgraph dataset to obtain an initial pulsar candidate recognition model includes: Obtaining a preset multimodal large language model, performing fine-tuning training on the preset multimodal large language model using the labeled target subgraph dataset, and saving a plurality of first checkpoint models according to a first preset period during the fine-tuning training process until the number of iterations of the fine-tuning training meets a first preset iteration threshold; The first checkpoint model with the smallest loss function value among the multiple first checkpoint models is determined as the initial pulsar candidate identification model.

9. The method according to claim 1, characterized in that The method of using a preset reward function and the labeled target subgraph dataset to perform intensive training on the initial pulsar candidate recognition model to obtain a pulsar candidate recognition model includes: Performing intensive training on the initial pulsar candidate recognition model using a preset reward function and the labeled target subgraph dataset, and saving a plurality of second checkpoint models according to a second preset period during the intensive training process until the number of iterations of the intensive training meets a second preset iteration threshold; The second checkpoint model with the smallest loss function value among the plurality of second checkpoint models is determined as the pulsar candidate identification model.

10. The method according to claim 9, characterized in that The intensive training includes: In each of the second preset periods, using the preset reward function, a reward calculation is performed on the initial recognition result output by the current initial pulsar candidate recognition model and the label corresponding to the initial recognition result to obtain a reward result corresponding to the initial recognition result; According to the reward result, the model weight of the current initial pulsar candidate identification model is updated, and the updated current initial pulsar candidate identification model is saved as a second checkpoint model.

11. A method for identifying pulsar candidates, characterized in that: The method comprises: Obtaining a candidate image corresponding to the candidate pulsar to be identified, and performing sub-image segmentation processing on the candidate image to be identified to obtain a corresponding sub-image to be identified; The subgraph to be identified is input into the pulsar candidate identification model according to any one of claims 1 to 8 to perform pulsar candidate identification, and a target identification result corresponding to the pulsar candidate to be identified is obtained.

12. The method according to claim 11, characterized in that The sub-graph to be identified includes a frequency-phase graph to be identified and a dispersion curve graph to be identified; the target identification result includes a first identification result corresponding to the frequency-phase graph to be identified and a second identification result corresponding to the dispersion curve graph to be identified; the method further includes: The pulsar candidate to be identified is classified according to the first recognition result corresponding to the frequency phase diagram to be identified and the second recognition result corresponding to the dispersion curve diagram to be identified, so as to obtain a classification result corresponding to the pulsar candidate to be identified.

13. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 12 are implemented.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 12 are implemented.

15. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 12 are implemented.

Images