A semi-supervised based astronomical time-frequency image target detection method and system

Abstract

The application discloses a kind of based on semi-supervised astronomical time-frequency image target detection method, belong to the field of astronomical image processing based on computer vision.The application includes: pre-training and verification are carried out to T-UNet student radio frequency interference detection model;The corresponding parameter file of the pre-trained T-UNet student radio frequency interference detection model is used to update the corresponding parameter in T-UNet teacher radio frequency interference detection model;The T-UNet teacher radio frequency interference detection model after updating is trained, and pseudo label is output;The astronomical time-frequency image corresponding to pseudo label is used as second labeled training set;The second labeled training set is used to expand, and the pre-trained T-UNet student radio frequency interference detection model is enhanced training and verification, and the enhanced T-UNet student radio frequency interference detection model after training is obtained.The application can significantly reduce the workload of labeling while ensuring the detection accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to a kind of based on semi-supervised astronomical time-frequency image target detection method and system, belong to based on computer vision astronomical image processing field. BACKGROUND

[0002] Modern wireless communication systems are facing increasingly complex radio frequency interference (RFI) challenges, which come not only from traditional man-made signal sources but also from new interference sources such as 5G communications and Internet of Things devices. These interferences exhibit strong time-varying and complex spectral characteristics, posing great challenges to traditional threshold-based and rule-based analysis methods. To address this issue, intelligent detection methods based on deep learning have shown unique advantages, enabling automatic learning of interference features from massive spectral data and achieving more accurate interference identification.

[0003] In the field of radio frequency interference detection and identification based on deep learning, fully supervised methods are the most common. While these methods can achieve good identification results, they have significant drawbacks, including high costs for manual label annotation. In view of this, semi-supervised learning provides an innovative solution for RFI detection. However, there is still a lack of effective solutions for radio frequency detection in the field of astronomy. Therefore, in the task of radio frequency detection, how to utilize a small amount of labeled data and a large amount of unlabeled data to reduce the dependence on labeled data for model training while ensuring the detection accuracy of the model is still a problem worth studying. SUMMARY

[0004] The present application provides a kind of based on semi-supervised astronomical time-frequency image target detection method, according to the way of semi-supervised learning, the training set of mixed labeled data and unlabeled data is constructed, further adopts teacher-student framework to carry out knowledge transfer, and after the optimization screening of pseudo label, original labeled training set is added again to be used for the enhanced training of model, and then the enhanced training based on T-UNet student radio frequency interference detection model is obtained, the model of this design can adapt to interference detection task under different signal-to-noise ratio conditions, can significantly reduce the workload of labeling while ensuring detection accuracy.

[0005] The technical solution of the present application is:

[0006] According to the first aspect of the present application, a kind of based on semi-supervised astronomical time-frequency image target detection method is provided, including the following steps:

[0007] S1, obtain astronomical time-frequency image, and is divided into training set, verification set and test set according to first preset proportion;The training set is divided into first training set and second training set according to second preset proportion;Radio frequency interference is taken as target, and real label is made for first training set, verification set and test set respectively, to obtain first labeled training set, labeled verification set and labeled test set;

[0008] S2, construct a T-UNet teacher radio frequency interference detection model and a T-UNet student radio frequency interference detection model;

[0009] S3, pre-training and verification are performed on the T-UNet student radio frequency interference detection model based on the first labeled training set and the labeled verification set, a pre-trained T-UNet student radio frequency interference detection model is obtained, and the corresponding parameters of the T-UNet teacher radio frequency interference detection model are updated using the parameter file corresponding to the pre-trained T-UNet student radio frequency interference detection model;

[0010] S4, the updated T-UNet teacher radio frequency interference detection model is trained based on the second training set and the labeled verification set, a trained T-UNet teacher radio frequency interference detection model is obtained, and the pseudo-labels of the second training set are output; the pseudo-labels of the second training set are filtered to obtain high-quality pseudo-labels; and the high-quality pseudo-labels and the corresponding astronomical time-frequency images are used as a second labeled training set.

[0011] S5, the first labeled training set is expanded using the second labeled training set to obtain an expanded labeled training set; and the pre-trained T-UNet student radio frequency interference detection model is enhanced and trained based on the expanded labeled training set and the labeled verification set to obtain an enhanced and trained T-UNet student radio frequency interference detection model.

[0012] S6, the astronomical time-frequency images in the labeled test set are input into the enhanced and trained T-UNet student radio frequency interference detection model to obtain a prediction result of the astronomical time-frequency images.

[0013] Further, the semi-supervised astronomical time-frequency image target detection method further comprises: under a preset prediction number of times, repeatedly performing S3-S6 to predict the labeled test set each time, after each prediction, judging the result of the labeled test set according to a judgment index, saving an enhanced and trained T-UNet student radio frequency interference detection model with a judgment index higher than a preset value and the highest, and predicting the astronomical time-frequency images to be predicted according to the saved enhanced and trained T-UNet student radio frequency interference detection model.

[0014] Further, the pseudo-labels of the second training set are filtered to obtain high-quality pseudo-labels, specifically: the pseudo-labels of the second training set are adaptively filtered to obtain high-quality pseudo-labels using a Monte Carlo Dropout method.

[0015] Further, the second preset proportion is specifically: the quantity proportion of the first training set is less than the quantity proportion of the second training set.

[0016] Further, the quantity ratio of the first training set and the second training set is 1:4-5.

[0017] Further, the T-UNet teacher radio frequency interference detection model and the T-UNet student radio frequency interference detection model have the same architecture, and both include an encoder and a decoder; the encoder and the decoder are based on the encoder and the decoder of the U-net network, and a skip connection is introduced between the encoder and the decoder; the encoder adopts three down-sampling operations connected in sequence; the decoder includes two convolutional layers, three up-sampling operations, and a time-frequency attention module introduced after the fusion features of the first up-sampling operation and the second down-sampling operation.

[0018] Further, the time-frequency attention module specifically models the input double time sequence features through a parallel channel attention branch and a spatial attention branch.

[0019] According to a second aspect of the present application, a semi-supervised astronomical time-frequency image target detection system is provided, comprising:

[0020] A first obtaining module is configured to obtain an astronomical time-frequency image, and divide the image into a training set, a validation set and a test set according to a first preset ratio; divide the training set into a first training set and a second training set according to a second preset ratio; and regard radio frequency interference as a target to make real labels for the first training set, the validation set and the test set respectively, to obtain a first labeled training set, a labeled validation set and a labeled test set;

[0021] A constructing module is configured to construct a T-UNet teacher radio frequency interference detection model and a T-UNet student radio frequency interference detection model.

[0022] A pre-training module is configured to pre-train and validate the T-UNet student radio frequency interference detection model according to the first labeled training set and the labeled validation set, to obtain a pre-trained T-UNet student radio frequency interference detection model; and update the corresponding parameters in the T-UNet teacher radio frequency interference detection model by using the parameter file corresponding to the pre-trained T-UNet student radio frequency interference detection model.

[0023] A second obtaining module is configured to train the updated T-UNet teacher radio frequency interference detection model according to the second training set and the labeled validation set, to obtain a trained T-UNet teacher radio frequency interference detection model, and output the pseudo-labels of the second training set; filter the pseudo-labels of the second training set to obtain high-quality pseudo-labels; and use the high-quality pseudo-labels and the corresponding astronomical time-frequency images as a second labeled training set.

[0024] a third obtaining module configured to expand the first labeled training set by using the second labeled training set to obtain an expanded labeled training set; and perform enhanced training and verification on the pre-trained T-UNet-based student radio frequency interference detection model based on the expanded labeled training set and the labeled verification set to obtain an enhanced and trained T-UNet-based student radio frequency interference detection model;

[0025] a fourth obtaining module configured to input the astronomical time-frequency image in the labeled test set into the enhanced and trained T-UNet-based student radio frequency interference detection model to obtain a prediction result of the astronomical time-frequency image.

[0026] Further, the astronomical time-frequency image target detection system based on semi-supervision further comprises:

[0027] The screening and prediction module repeatedly performs the pre-training module, the second obtaining module, the third obtaining module and the fourth obtaining module to predict the labeled test set for a preset number of times, judges the result of the labeled test set based on the judging index after each prediction, saves the enhanced and trained T-UNet-based student radio frequency interference detection model with the highest judging index higher than a preset value, and predicts the astronomical time-frequency image to be predicted based on the saved enhanced and trained T-UNet-based student radio frequency interference detection model.

[0028] The present application has the advantages that: in view of the fact that the real observation data obtained by observing and recording the radio astronomical signals by the radio telescope has high data labeling cost and less labeled data, the present application constructs a training set with a small amount of labeled data and a large amount of unlabeled data and a labeled verification set with a small amount of labeled data based on the student model and the teacher model, realizes the training and verification based on the semi-supervised teacher-student radio frequency interference detection model, and effectively saves the huge cost caused by manual labeling of the training set samples; further, the T-UNet-based teacher radio frequency interference detection model and the T-UNet-based student radio frequency interference detection model constructed by the present application only introduce a time-frequency attention module based on the U-net to form an efficient semantic segmentation network, and the introduction of the time-frequency attention module can realize the dynamic enhancement and detection of the time-frequency features and improve the perception ability of the model to the key signals. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 is a flowchart of the present application;

[0030] Figure 2 is a radio frequency interference detection architecture diagram based on the T-UNet teacher and student radio frequency interference detection models;

[0031] Figure 3 is a structure diagram of the T-UNet-based teacher and student radio frequency interference detection models.

[0032] Figure 4 This is a time-frequency attention module diagram;

[0033] Figure 5 It is a radio astronomy time-frequency image from a labeled test set;

[0034] Figure 6 Based on Figure 5 The image with radio frequency interference mask is obtained by prediction using a UNet-based student radio frequency interference detection model that has been enhanced and trained.

[0035] Figure 7 Based on Figure 5 The radio frequency interference masked image is obtained by using the enhanced and trained T-UNet student radio frequency interference detection model of this invention for prediction. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be arbitrarily combined with each other.

[0037] Example 1: As Figures 1-7 As shown, according to a first aspect of the present invention, a semi-supervised astronomical time-frequency image target detection method is provided, comprising the following steps:

[0038] S1. Acquire astronomical time-frequency images and divide them into training set, validation set and test set according to a first preset ratio; divide the training set into first training set and second training set according to a second preset ratio; take radio frequency interference as the target, and create real labels for the first training set, validation set and test set respectively to obtain the first labeled training set, labeled validation set and labeled test set.

[0039] S2. Construct a teacher radio frequency interference detection model and a student radio frequency interference detection model based on T-UNet;

[0040] S3. Based on the first labeled training set and labeled validation set, pre-train and validate the T-UNet-based student radio frequency interference detection model to obtain the pre-trained T-UNet-based student radio frequency interference detection model; use the parameter file corresponding to the pre-trained T-UNet-based student radio frequency interference detection model and adopt the EMA strategy to update the corresponding parameters in the T-UNet-based teacher radio frequency interference detection model.

[0041] S4, training the updated T-UNet teacher radio frequency interference detection model based on the second training set and the labeled validation set to obtain a trained T-UNet teacher radio frequency interference detection model, and outputting pseudo-labels of the second training set; using a Monte Carlo Dropout method to adaptively filter low-quality pseudo-labels from the pseudo-labels of the second training set to obtain high-quality pseudo-labels; and using the high-quality pseudo-labels and corresponding astronomical time-frequency images as a second labeled training set;

[0042] S5, expanding the first labeled training set using the second labeled training set to obtain an expanded labeled training set; and enhancing training and verifying the pre-trained T-UNet student radio frequency interference detection model based on the expanded labeled training set and the labeled validation set to obtain an enhanced trained T-UNet student radio frequency interference detection model;

[0043] S6, inputting the astronomical time-frequency images in the labeled test set into the enhanced trained T-UNet student radio frequency interference detection model to obtain a prediction result of the astronomical time-frequency images.

[0044] Further, it further comprises: under a preset prediction number, repeatedly performing S3-S6 to predict the labeled test set each time, after each prediction, judging the result of the labeled test set according to a judgment index, saving the enhanced trained T-UNet student radio frequency interference detection model with the highest judgment index higher than a preset value; and predicting the astronomical time-frequency images to be predicted based on the saved enhanced trained T-UNet student radio frequency interference detection model.

[0045] Further, the prediction result of the astronomical time-frequency images and the true label of the astronomical time-frequency images are evaluated using an evaluation index.

[0046] Further, the second preset proportion is specifically: the quantity proportion of the first training set is less than the quantity proportion of the second training set, that is, the quantity of the first training set is set to be much less than the quantity of the second training set.

[0047] Further, the quantity proportion of the first training set and the second training set is 1:4-5.

[0048] Further, the T-UNet teacher radio frequency interference detection model and the T-UNet student radio frequency interference detection model have the same architecture, both including an encoder and a decoder; the encoder and the decoder are based on the encoder and the decoder of the U-net network, and a skip connection is introduced between the encoder and the decoder; the encoder adopts three down-sampling operations connected in sequence; the decoder includes two convolutional layers, three up-sampling operations, and a time-frequency attention module introduced after the fusion features of the first up-sampling operation and the second down-sampling operation.

[0049] Further, the down-sampling operation in the encoder uses 3x3 convolution + Relu operation, the up-sampling operation in the decoder uses 3x3 deconvolution + Relu operation, and the two convolutional layers in the decoder use 1x1 convolution + Relu operation.

[0050] Further, the time-frequency attention TFAM module specifically models the input double time sequence features through a parallel channel attention branch and a spatial attention branch; in the channel attention branch, the features are subjected to average pooling and maximum pooling, then spliced in the channel dimension, and the channel attention weight is calculated through one-dimensional convolution and Softmax; in the spatial attention branch, the features are subjected to average pooling and maximum pooling, then spliced in the spatial dimension, and the spatial attention feature map is generated through two-dimensional convolution and Softmax. Finally, the outputs of the two branches are weighted and fused to obtain a feature map containing more rich feature information, realizing dynamic enhancement and detection of the time-frequency features, improving the perception ability of the model to key signals, and being helpful for subsequent interference detection.

[0051] According to a second aspect of the embodiment of the present application, a semi-supervised astronomical time-frequency image target detection system is provided, comprising: a first obtaining module configured to obtain an astronomical time-frequency image, and divide the image into a training set, a validation set and a test set according to a first preset ratio; divide the training set into a first training set and a second training set according to a second preset ratio; take radio frequency interference as a target, and make real labels for the first training set, the validation set and the test set respectively to obtain a first labeled training set, a labeled validation set and a labeled test set; a constructing module configured to construct a T-UNet teacher radio frequency interference detection model and a T-UNet student radio frequency interference detection model; a pre-training module configured to pre-train and validate the T-UNet student radio frequency interference detection model according to the first labeled training set and the labeled validation set to obtain a pre-trained T-UNet student radio frequency interference detection model; update corresponding parameters in the T-UNet teacher radio frequency interference detection model using a parameter file corresponding to the pre-trained T-UNet student radio frequency interference detection model; a second obtaining module configured to train the updated T-UNet teacher radio frequency interference detection model according to the second training set and the labeled validation set to obtain a trained T-UNet teacher radio frequency interference detection model, and output pseudo-labels of the second training set; filter the pseudo-labels of the second training set to obtain high-quality pseudo-labels; use the high-quality pseudo-labels and corresponding astronomical time-frequency images as a second labeled training set; a third obtaining module configured to expand the first labeled training set using the second labeled training set to obtain an expanded labeled training set; enhance train and validate the pre-trained T-UNet student radio frequency interference detection model according to the expanded labeled training set and the labeled validation set to obtain an enhanced trained T-UNet student radio frequency interference detection model; and a fourth obtaining module configured to input an astronomical time-frequency image in the labeled test set into the enhanced trained T-UNet student radio frequency interference detection model to obtain a prediction result of the image.

[0052] Further, the semi-supervised astronomical time-frequency image target detection system further comprises a screening prediction module configured to repeatedly execute the pre-training module, the second obtaining module, the third obtaining module and the fourth obtaining module to predict the labeled test set for a preset number of times, judge the result of the labeled test set according to a judgment index after each prediction, save an enhanced trained T-UNet student radio frequency interference detection model with a judgment index higher than a preset value and the highest, and predict a to-be-predicted astronomical time-frequency image according to the saved enhanced trained T-UNet student radio frequency interference detection model.

[0053] The various modules in the above semi-supervised astronomical time-frequency image target detection system can be realized by software, hardware, and combinations thereof, in whole or in part. The various modules can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so that the processor calls and executes the operations corresponding to the various modules.

[0054] Embodiment 2: The optional specific implementation process of the present application is described as follows in combination with experimental data:

[0055] I. According to the real pulsar observation data of the 40-meter aperture radio telescope of the Yunnan Observatory of the Chinese Academy of Sciences from September 2016 to October 2022, 6602 observation files were randomly selected from the real pulsar observation data; according to the observation files, they were converted into RGB radio time-frequency images and saved as original data sets. The original 6602 radio time-frequency images were divided into 5882 training sets, 360 validation sets, and 360 test sets (the images in the original data set are astronomical time-frequency images with different signal-to-noise ratios), and then the 5882 training sets were divided into a first training set and a second training set according to a 1:4 ratio, i.e. 1176 as the first training set and 4706 as the second training set. The corresponding labels of the 1176 astronomical time-frequency images in the first training data set, the 360 validation set time-frequency images, and the 360 test set astronomical time-frequency images were made and saved, obtaining a first labeled training set, a labeled validation set, and a labeled test set. The first labeled training set is used for learning and training; the validation set is used to verify the performance of the model; and the test set is used to evaluate the radio frequency interference detection effect of the model.

[0056] II. According to the first labeled training set and the labeled validation set, the T-UNet student radio frequency interference detection model is pre-trained and verified to obtain a pre-trained T-UNet student radio frequency interference detection model; the corresponding parameter file of the pre-trained T-UNet student radio frequency interference detection model is used to update the corresponding parameters in the T-UNet teacher radio frequency interference detection model, specifically: using the 1176 first labeled training set to train the T-UNet student radio frequency interference detection model and saving the parameters, and updating the parameters to the T-UNet teacher radio frequency interference detection model;

[0057] III. Training the updated T-UNet teacher radio frequency interference detection model based on the second training set and the labeled validation set, obtaining the trained T-UNet teacher radio frequency interference detection model, and outputting the pseudo-labels of the second training set. The Monte Carlo Dropout method is used to adaptively filter low-quality pseudo-labels and obtain high-quality pseudo-labels. The high-quality pseudo-labels and corresponding astronomical time-frequency images are used as the second labeled training set, specifically: the T-UNet teacher radio frequency interference detection model is trained using the second training set, i.e., 4706 unlabeled astronomical time-frequency images, and the Monte Carlo Dropout method is used for dynamic confidence screening to adaptively filter low-quality pseudo-labels and save high-quality pseudo-labels. The high-quality pseudo-labels and corresponding astronomical time-frequency images are used as the second labeled training set.

[0058] IV. The first labeled training set is expanded using the second labeled training set to obtain an expanded labeled training set. The expanded labeled training set with pseudo-labels is used for retraining of the pre-trained T-UNet student radio frequency interference detection model. In the learning and training process, the T-UNet student radio frequency interference detection model learns the interference information in the original time-frequency images. Then, the labeled validation set is used for verification. After verification, the labeled test set is used for testing to ensure the rationality and reliability of the model.

[0059] The precision, recall, and F2 score are used to evaluate the model in the verification and testing process.

[0060] In the radio frequency interference target detection task, the precision represents the proportion of pixels correctly predicted as radio frequency interference to all pixels predicted as radio frequency interference. The formula for precision is:

[0061] ;

[0062] Where TP (true positive) is the number of pixels correctly predicted as radio frequency interference by the model, and FP (false positive) is the number of pixels incorrectly predicted as radio frequency interference by the model.

[0063] In the radio frequency interference target detection task, the recall represents the proportion of pixels correctly predicted as radio frequency interference to all pixels actually predicted as radio frequency interference. The formula for recall is:

[0064] ;

[0065] Where TP is the number of pixels correctly predicted as radio frequency interference by the model, and FN (false negative) is the number of pixels incorrectly predicted as radio frequency interference by the model.

[0066] ​​In the radio frequency interference target detection task, The score is a weighted harmonic mean of precision and recall, and by adjusting The value can balance the relationship between precision and recall. The formula of the score is:

[0067] ;

[0068] Wherein, is a weight parameter for balancing precision and recall. In The score, The greater the value of , the greater the influence on recall. In the present application, the value of is set to 2, i.e. F2 score.

[0069] After the preset number of predictions on the labeled test set according to the method of the present application, the comprehensive index F2 score is used as the evaluation index, and the model with F2 score higher than 80% and the highest is saved as the enhanced training based on the T-UNet student radio frequency interference detection model. In the embodiment of the present application, the preset number of predictions is set to 20 times.

[0070] Figure 5 To explore the performance advantage of the T-UNet student radio frequency interference detection model proposed in the present application, the U-Net network is replaced by the T-UNet network for experiment according to the method of the present application, and the performance index of the model proposed in the present application is compared. In the experiment, the U-Net teacher / student radio frequency interference detection model and the T-UNet teacher / student radio frequency interference detection model are trained and verified using the same experimental platform using the first labeled training set, the second labeled training set, and the labeled verification set. An enhanced training based on the U-Net student radio frequency interference detection model and an enhanced training based on the T-UNet student radio frequency interference detection model are obtained. A random selection of an astronomical time-frequency image from the labeled test set is shown in Figure 6 , Figure 7 The visualization results of the prediction images obtained by the enhanced training based on the U-Net student radio frequency interference detection model and the enhanced training based on the T-UNet student radio frequency interference detection model are shown in

[0071] Table 1

[0072] Model Precision Recall F2 score U-Net based student radio frequency interference detection model 0.9075 0.7963 0.8163 T-UNet based student radio frequency interference detection model 0.9146 0.8257 0.8421

[0073] It should be noted that for the radio frequency interference target detection task, both the accuracy of detection and the reduction of missed detection, that is, the improvement of comprehensiveness, are required. Therefore, the comprehensive index F2 number is selected as the final evaluation basis. It can be known from Table 1 that the effect of the radio frequency interference target detection in the manner of using the U-Net student radio frequency interference detection model for semi-supervised learning is acceptable, but the comprehensiveness needs to be improved, and there is still room for improvement. The T-UNet student radio frequency interference detection model applied in the present application makes the recall rate increase to more than 80%, which indicates that the model can more accurately and comprehensively detect radio frequency interference.

[0074] More importantly, the present application can effectively train and verify by only using the first training set with a proportion less than the second training set for label making, so that the present application significantly reduces the cost of manually making real labels. At the same time, it can be known from the above comparison of multiple simulations that the present application ensures the superior effect of model detection, which indicates that the present application ensures the advantage of the model in multiple performance indicators while reducing the labeling cost.

[0075] The specific embodiments of the present application are described in detail above in combination with the drawings, but the present application is not limited to the above embodiments, and various changes can be made within the knowledge of those skilled in the art without departing from the purpose of the present application.

Claims

1. A semi-supervised based astronomical time-frequency image target detection method, characterized in that, The method comprises the following steps: S1, acquiring an astronomical time-frequency image, and dividing the image into a training set, a validation set and a test set according to a first preset ratio; dividing the training set into a first training set and a second training set according to a second preset ratio; taking radio frequency interference as a target, and making true labels for the first training set, the validation set and the test set respectively to obtain a first labeled training set, a labeled validation set and a labeled test set; S2, constructing a T-UNet teacher radio frequency interference detection model and a T-UNet student radio frequency interference detection model; S3, pre-training and verifying the T-UNet student radio frequency interference detection model according to the first labeled training set and the labeled validation set to obtain a pre-trained T-UNet student radio frequency interference detection model; and updating corresponding parameters in the T-UNet teacher radio frequency interference detection model by using a parameter file corresponding to the pre-trained T-UNet student radio frequency interference detection model; S4, training the updated T-UNet teacher radio frequency interference detection model according to the second training set and the labeled validation set to obtain a trained T-UNet teacher radio frequency interference detection model, and outputting pseudo labels of the second training set; filtering the pseudo labels of the second training set to obtain high-quality pseudo labels; and taking the high-quality pseudo labels and corresponding astronomical time-frequency images as a second labeled training set; S5, expanding the first labeled training set by using the second labeled training set to obtain an expanded labeled training set; and enhancing training and verifying the pre-trained T-UNet student radio frequency interference detection model according to the expanded labeled training set and the labeled validation set to obtain an enhanced trained T-UNet student radio frequency interference detection model; S6, inputting the astronomical time-frequency images in the labeled test set into the enhanced trained T-UNet student radio frequency interference detection model to obtain a prediction result of the astronomical time-frequency images; The T-UNet teacher radio frequency interference detection model and the T-UNet student radio frequency interference detection model have the same architecture, and both comprise an encoder and a decoder; the encoder and the decoder are based on the encoder and the decoder of a U-net network, and a skip connection is introduced between the encoder and the decoder; the encoder adopts three down-sampling operations in sequence; the decoder comprises two convolutional layers, three up-sampling operations, and a time-frequency attention module introduced after the fusion features of the first up-sampling operation and the second down-sampling operation; The pseudo labels of the second training set are filtered to obtain high-quality pseudo labels, specifically: the pseudo labels of the second training set are adaptively filtered to remove low-quality pseudo labels by using a Monte Carlo Dropout method to obtain high-quality pseudo labels; The time-frequency attention module specifically models the input double-time sequence features through a parallel channel attention branch and a spatial attention branch.

2. The method according to claim 1, wherein, The semi-supervised astronomical time-frequency image target detection method further comprises: repeatedly performing S3-S6 on the labeled test set for a preset number of times of prediction, judging the results of the labeled test set according to the evaluation index after each prediction, saving the enhanced trained T-UNet student radio frequency interference detection model with an evaluation index higher than a preset value and the highest one; and predicting the astronomical time-frequency image to be predicted according to the saved enhanced trained T-UNet student radio frequency interference detection model.

3. The method of claim 1, wherein the method is based on semi-supervised astronomical time-frequency image target detection. The second preset proportion is specifically that the number proportion of the first training set is less than the number proportion of the second training set.

4. The method according to claim 3, wherein, The number proportion of the first training set and the second training set is 1:4-5.

5. A semi-supervised based astronomical time-frequency image target detection system, characterized in that, It comprises: A first obtaining module is configured to obtain an astronomical time-frequency image and divide the astronomical time-frequency image into a training set, a validation set and a test set according to a first preset proportion; The training set is divided into a first training set and a second training set according to a second preset proportion; radio frequency interference is taken as a target to make real labels for the first training set, the validation set and the test set respectively, and a first labeled training set, a labeled validation set and a labeled test set are obtained; A constructing module is configured to construct a T-UNet teacher radio frequency interference detection model and a T-UNet student radio frequency interference detection model; A pre-training module is configured to pre-train and validate the T-UNet student radio frequency interference detection model according to the first labeled training set and the labeled validation set, obtain a pre-trained T-UNet student radio frequency interference detection model, and update the corresponding parameters in the T-UNet teacher radio frequency interference detection model by using the parameter file corresponding to the pre-trained T-UNet student radio frequency interference detection model; A second obtaining module is configured to train the updated T-UNet teacher radio frequency interference detection model according to the second training set and the labeled validation set, obtain a trained T-UNet teacher radio frequency interference detection model, and output the pseudo labels of the second training set; filter the pseudo labels of the second training set to obtain high-quality pseudo labels; and use the high-quality pseudo labels and the corresponding astronomical time-frequency images as a second labeled training set; A third obtaining module is configured to use the second labeled training set to expand the first labeled training set to obtain an expanded labeled training set; and enhance train and validate the pre-trained T-UNet student radio frequency interference detection model according to the expanded labeled training set and the labeled validation set to obtain an enhanced trained T-UNet student radio frequency interference detection model; A fourth obtaining module is configured to input the astronomical time-frequency images in the labeled test set into the enhanced trained T-UNet student radio frequency interference detection model to obtain the prediction results of the astronomical time-frequency images. The architecture of the T-UNet-based teacher radio frequency interference detection model and the T-UNet-based student radio frequency interference detection model is same, and each includes an encoder and a decoder; the encoder and the decoder are in the framework of the encoder and the decoder of a U-net network, a skip connection is introduced between the encoder and the decoder, a three-time downsampling operation is adopted in the encoder in sequence; the decoder includes two convolutional layers, three-time upsampling operations, and a time-frequency attention module is introduced after the fusion features of the first upsampling operation and the second downsampling operation; The pseudo labels of the second training set are filtered to obtain high-quality pseudo labels, specifically: the pseudo labels of the second training set are adaptively filtered by a Monte Carlo Dropout method to filter low-quality pseudo labels and obtain high-quality pseudo labels. The time-frequency attention module specifically models the input double-time sequence features through a parallel channel attention branch and a spatial attention branch.

6. The system for semi-supervised based astronomical time-frequency image target detection according to claim 5, wherein, Further comprising: A screening prediction module, under a preset prediction number, repeatedly executes the pre-training module, the second obtaining module, the third obtaining module and the fourth obtaining module to predict the labeled test set each time, judges the result of the labeled test set according to the evaluation index after each prediction, saves the enhanced and trained T-UNet-based student radio frequency interference detection model with an evaluation index higher than a preset value and the highest, and predicts the astronomical time-frequency image to be predicted according to the saved enhanced and trained T-UNet-based student radio frequency interference detection model.

Images