A progressive active learning weak supervision training method for dim and weak target detection

Abstract

The present application relates to the technical field of space target detection, and specifically provides a progressive active learning weak supervision training method for dim and weak target detection, comprising: obtaining sample data to form a sample set, dividing the sample set into an initial sample set and a to-be-evaluated group according to the signal-to-noise ratio of the target, training a model using the initial sample set to obtain a preliminarily trained model, and starting in-group iteration: selecting the to-be-evaluated group in turn in descending order of signal-to-noise ratio, dividing the target in the to-be-evaluated group into simple samples and difficult samples, training a model using the simple samples to obtain an updated model, re-dividing the difficult samples into simple samples and difficult samples, repeatedly training and dividing until a preset in-group iteration termination condition is met, completing in-group iteration, and selecting a next to-be-evaluated group in turn to repeatedly execute in-group iteration. The present application progressively improves the training difficulty according to the real-time capability of the model, improves the detection precision and training efficiency, and reduces the data labeling cost.

Description

Technical Field

[0001] This invention relates to the field of space target detection technology, specifically providing a progressive active learning weakly supervised training method for detecting dark and weak targets. Background Technology

[0002] Combining active learning with weakly supervised learning methods, this approach maximizes model performance while minimizing annotation costs by actively selecting samples to be labeled. In the field of spatial object detection, while deep learning methods have improved detection performance, they face four major bottlenecks: the lack of accurate real-world labels in actual observation data, limiting model generalization due to reliance on simulated data training; difficulties in transferring models across different telescopes due to heterogeneity; instability in target appearance features caused by dynamic observation scenarios; and high cost and susceptibility to subjective errors in spatial object data annotation. These issues highlight the urgent need to develop low-cost, high-efficiency weakly supervised learning. Single-point supervision, as a highly promising approach within weak supervision, only requires labeling one point on each target, but provides limited information, posing a significant challenge to model training.

[0003] Existing single-point supervised target detection methods have the following problems:

[0004] First, there is the problem of low-quality pseudo-labels and noise accumulation. The initial pseudo-labels generated based on single-point annotations are of low quality, resulting in limited model performance and high noise in the prediction results. Using low-quality predictions as new pseudo-labels creates a vicious cycle of "poor students teaching poor students," making it difficult to improve model performance or even causing it to degrade.

[0005] Second, it has a weak ability to learn from difficult samples. The difficulty distribution of the samples to be detected is uneven, and difficult samples are easily misidentified by the model in the early stages of training, generating incorrect gradient signals and interfering with the normal learning of the model. Existing methods treat all samples equally and send them into training, which affects the final detection accuracy.

[0006] Third, the training strategies are rigid and lack adaptability. Some methods introduce course learning, but the training order is fixed and cannot be dynamically adjusted according to the model's real-time capabilities; in weakly supervised scenarios, traditional active learning tends to judge difficult samples as "informative" in the early stages, leading to unstable training.

[0007] Furthermore, in the field of fine-grained small object detection under single-point supervision, although existing technologies have made some progress, several key drawbacks still exist. These drawbacks are interconnected and collectively restrict further improvements in model performance. The key drawbacks of existing technologies in the field of fine-grained small object detection under single-point supervision are as follows:

[0008] First, the inaccurate quantification of sample difficulty leads to a bias in the starting point of training. Existing weakly supervised methods lack precise physical quantification of the inherent difficulty of samples, either treating all samples the same or using coarse indicators for classification, resulting in low training efficiency, slow model convergence, and unstable final performance.

[0009] Second, the training strategies are rigid and lack adaptability to match the model's capabilities. Traditional course learning adopts a fixed sequence, and traditional active learning is prone to "blindly rushing forward" in weakly supervised scenarios, consuming a lot of training time and making it difficult to achieve efficient and stable knowledge accumulation.

[0010] Third, the improvement of pseudo-label quality is limited, which can easily lead to a vicious cycle of "less effective learning from less effective learning." Existing methods have a crude pseudo-label update mechanism, resulting in poor initial model performance, high pseudo-label noise, and a single standard for evaluating pseudo-label quality, leading to low self-refinement efficiency and limiting the final accuracy of the model.

[0011] Fourth, data utilization is insufficient, and learning from difficult samples is inadequate. Existing methods lack mechanisms for handling difficult samples, wasting valuable information and causing the model to miss opportunities to attempt them again, resulting in low data utilization and insufficient improvement in model generalization ability and robustness. Summary of the Invention

[0012] To address the aforementioned problems, this invention provides a progressive active learning weakly supervised training method for detecting dark and weak targets. The method divides targets into simple and difficult samples based on the target signal-to-noise ratio and the current model's capabilities, and trains the model from simple to difficult samples, progressively increasing the training difficulty. This fully considers the model's capabilities at each time step, efficiently utilizes every sample data point, continuously optimizes the quality of pseudo-labels, and obtains a high-performance model.

[0013] The present invention provides a progressive active learning weakly supervised training method for detecting targets in dark and weak environments. The trained model is used to detect targets in dark and weak spaces, including:

[0014] S1: Acquire sample data to form a sample set. The sample data includes the target. Divide the sample set into an initial sample set and an initial sample set according to the target's signal-to-noise ratio. Groups to be evaluated ;

[0015] S2: Update the training pool using the initial sample set, train the model using the existing sample data in the training pool, and obtain the initially trained model;

[0016] S3: Select the groups to be evaluated in descending order of signal-to-noise ratio. Perform intra-group iterations:

[0017] The groups to be evaluated were divided according to the preset difficulty level criteria. The targets are divided into simple samples and difficult samples; , indicating the sequence number of the group to be evaluated;

[0018] The training pool is updated using simple samples, and the model is trained using the existing sample data in the training pool to obtain the updated model. Based on the difficulty distinction criteria and the updated model, the difficult samples are reclassified into simple and difficult samples. S3 is repeated until the preset intra-group iteration termination condition is met, completing the grouping process using the samples to be evaluated. Training the model;

[0019] S4: Perform inter-group iterations:

[0020] Select the group to be evaluated Repeat the iteration within the group until the resource is utilized. Each group to be evaluated has completed model training.

[0021] The preferred method for obtaining the signal-to-noise ratio of the target is as follows:

[0022] Based on the target's centroid, a two-dimensional Gaussian model of the target is fitted using the Levenberg-Marquardt optimization algorithm to obtain the target's FWHM value; the signal-to-noise ratio (SNR) of the target is then obtained using the FWHM value. The calculation formula is:

[0023] ;

[0024] in, , Indicates the measured aperture radius. Represents the empirical coefficient. Indicates the measuring aperture. Represents pixel value, Indicates the background estimate. Indicates the background standard deviation. This indicates the number of pixels within the aperture.

[0025] The preferred method for dividing the initial sample set and the groups to be evaluated is as follows:

[0026] Set a signal-to-noise ratio (SNR) threshold, select targets with an SNR higher than the SNR threshold from all sample data, and cover other targets in the sample data containing the target. The resulting sample data set is called the initial sample set.

[0027] Other targets outside the initial sample set are sorted in descending order of signal-to-noise ratio, and these targets are divided into... For each group to be evaluated, targets that do not belong to that group in the sample data are masked.

[0028] The preferred criterion for distinguishing between easy and difficult is:

[0029] ;

[0030] in, This indicates the results of dividing the sample data into easy and difficult categories. Represents sample data, Representing simple samples, Indicates a difficult sample. Indicates average recall rate. Indicates the recall threshold. This represents the lower limit threshold of the signal-to-noise ratio. This represents the average centroid loss. This represents the median loss of the centroids of all targets in the sample data. Indicates the signal-to-noise ratio. Indicates other situations.

[0031] Preferably, the iteration termination condition within the group is considered satisfied if any one of the following conditions is met:

[0032] The first item: All targets in the current evaluation group have been classified as simple samples, and the training pool has been updated for subsequent training of the model;

[0033] The second criterion is that no new simple samples are identified within a preset number of iteration update cycles.

[0034] Preferably, after the iteration within a group is completed, if there are any remaining difficult samples, they are added to the group to be evaluated. They will participate in the next group iteration.

[0035] Preferably, after completing the inter-group iterations, the remaining difficult samples are discarded.

[0036] Compared with the prior art, the present invention can achieve the following beneficial effects:

[0037] 1. Scientific and Reasonable Training Process: The "simplest" sample group with the highest signal-to-noise ratio is selected as the initial training set to ensure that the model establishes basic cognitive abilities in an "interference-free" environment. Then, a progressive training process, from easy to difficult, is adopted, strictly processing the samples in descending order of signal-to-noise ratio. Iterative learning is fully conducted within each group, and "residual difficult samples" are automatically and losslessly passed to the next more difficult group. Compared with traditional hybrid training, the grouping mechanism ensures a smooth and progressive learning process. Learning in descending order of signal-to-noise ratio ensures an orderly increase in knowledge difficulty. The "residual sample passing" forms a feedback loop, ensuring no information is lost and allowing difficult samples to be re-examined. The model learns appropriate samples at each stage, greatly improving the utilization rate of sample data and the upper limit of model performance, making the training process orderly and resilient.

[0038] 2. Dynamic Adaptive Difficulty Differentiation Criteria: By using multi-dimensional and dynamic difficulty differentiation criteria such as recall rate and centroid loss, the model's mastery of sample data of different difficulty levels is evaluated in real time, so that the training difficulty is dynamically matched with the model's ability. The high degree of adaptability is far superior to the traditional fixed course learning strategy.

[0039] 3. Introduction of target masking processing: The original approach introduces target masking processing to cover targets that do not belong to the current difficulty level and known interference sources when training specific groups of images. As an "attention-guided" method, it actively eliminates gradient interference from other targets in the same image, ensuring that the model focuses on learning knowledge of specific difficulty. Combined with the aforementioned methods, a more robust and pure initial model can be trained.

[0040] 4. Solving the challenges of weakly supervised learning: Effectively solves the common problems of "cold start" and "noise accumulation" in weakly supervised learning, avoids interference from noise information and continuously optimizes the quality of pseudo-labels. Under the condition of only single-point supervision, it significantly improves the detection accuracy of fine-grained small targets and performs outstandingly in handling complex and difficult samples.

[0041] The method of this invention only requires centroid or point labeling for weakly supervised training. It can dynamically select samples based on the model's real-time capabilities, progressively increase the training difficulty, and continuously optimize the quality of pseudo-labels. As an independent training strategy, it has good versatility and can be flexibly applied to various spatial object detection network architectures. At the same time, relying only on single-point labeling greatly reduces data labeling costs. Attached Figure Description

[0042] Figure 1 This is a flowchart of a progressive active learning weakly supervised training method for dark and weak target detection provided by an embodiment of the present invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and do not constitute a limitation thereof. Similar elements in different embodiments are referred to by associated similar element reference numerals. In the following embodiments, many details are described to facilitate a better understanding of the invention. However, those skilled in the art will readily recognize that some features may be omitted in different situations, or may be replaced by other elements, materials, or methods. In some cases, some operations related to the invention are not shown or described in the specification. This is to avoid obscuring the core parts of the invention with excessive description. For those skilled in the art, detailed description of these related operations is not necessary; they can fully understand the related operations based on the description in the specification and general technical knowledge in the art.

[0044] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined to form various implementations. Furthermore, the order of the steps or actions in the method description can be changed or adjusted in a manner readily apparent to those skilled in the art. Therefore, the various orders in the specification and drawings are merely for the clear description of a particular embodiment and do not imply a mandatory order, unless otherwise stated that a particular order must be followed.

[0045] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation on this invention. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0046] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0047] The invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0048] like Figure 1 As shown, this invention provides a weakly supervised model training method for training a model to detect faint targets in space. The specific model training method is as follows:

[0049] S1: Acquire sample data to form a sample set. The sample data includes the target. Divide the sample set into an initial sample set and an initial sample set according to the target's signal-to-noise ratio. Groups to be evaluated .

[0050] Training the model is a core step in machine learning and deep learning, aiming to optimize model parameters through data to enable predictive or classification capabilities. Before training, multiple sample data points are acquired, forming a sample set. Each sample may contain one or more targets. In this embodiment of the invention, the sample data is derived from images of space targets obtained by a ground-based telescope in tracking mode. In tracking mode, space targets appear as Gaussian dots, while stars, due to relative motion, appear as stripes of varying lengths. This significant morphological difference provides a physical basis for subsequently distinguishing targets from interference using morphological parameters.

[0051] The training method in this embodiment of the invention requires dividing the sample set according to the signal-to-noise ratio (SNR) of the target; therefore, it is necessary to obtain the SNR of each target. The specific method for obtaining the SNR is as follows:

[0052] First, the acquired targets are screened to remove anomalous targets, including those exhibiting star-crossing or occlusion phenomena. The telescope observes known space targets, therefore their star orbits are known. High-quality centroid annotations of space targets in the sample data are obtained through precise star orbits. Using the target's centroid, morphological parameters such as image moments and the ratio of eigenvalues ​​to the structure tensor are calculated. Combined with analysis of the temporal changes in target morphological features across multiple consecutive frames of sample data, a robust star-crossing and occlusion recognition mechanism can be established to effectively filter out these anomalous targets. For targets identified as star-crossing or occluded in the sample data, masking processing is performed to ensure that the model's attention is entirely focused on valid targets during training on that sample data, avoiding interference from worthless signals.

[0053] Using the selected valid targets, a Gaussian model is fitted and its full width at half maximum (FWHM) value is extracted. For each valid target, a sub-region of appropriate size is extracted centered on its centroid. Local statistical methods are used for background estimation and subtraction. Under physical constraints, the Levenberg-Marquardt optimization algorithm is applied to fit a Gaussian model to the target. The fitted two-dimensional Gaussian model is as follows:

[0054] ,

[0055] in, Represents pixel coordinates, Represents pixel coordinates Model strength at that location Indicates peak amplitude. Indicates the background estimate. The standard deviation in the X direction is represented by the standard deviation in the X direction. The standard deviation in the Y direction is represented by... Indicates the centroid coordinates. This represents an exponential function.

[0056] To ensure the accuracy of the full width at half maximum (FWHM) measurement, a multi-dimensional verification system can be designed, including signal-to-noise ratio (SNR) threshold screening, parameter rationality verification, FWHM range limitation, and axial ratio anomaly detection. This ensures the high reliability of the extracted FWHM value, and the SNR of the effective target is measured using the FWHM. Based on the aforementioned FWHM, the SNR measurement aperture is dynamically adjusted, and the accurate SNR of the effective target is measured. Specifically, the measurement aperture radius is set. The signal-to-noise ratio (SNR) calculation formula is:

[0057] ,

[0058] in, , This represents an empirical coefficient, which in this embodiment of the invention is set to 1.5. Indicates the measuring aperture. Indicates the background estimate. Indicates the background standard deviation. Indicates the number of pixels within the aperture. Represents pixel value.

[0059] Using the method described above, the signal-to-noise ratio (SNR) of each target in multiple sample data sets is calculated one by one. The sample data is then divided according to the SNR of each target, as follows:

[0060] A signal-to-noise ratio (SNR) threshold is set. Targets with an SNR higher than this threshold are selected from all sample data. Other targets in the same sample data containing these targets are then masked. This sample data containing targets with an SNR higher than the threshold is called the initial sample set, which is then added to the training pool. The targets in the initial sample set have relatively high SNRs and are considered relatively easy to learn from. All targets outside the initial sample set are then sorted in descending order of SNR. Based on this descending order, these targets are further divided into... Groups to be evaluated , and will Groups to be evaluated Place them in the preparation pool for selection.

[0061] To prevent interference between targets in different evaluation groups, targets that do not belong to a group in the grouped sample data are masked. This ensures that when the model is training on the sample data, its attention is fully focused on the target at the current difficulty level, avoiding interference from gradient signals from targets of other difficulty levels.

[0062] It should be noted that the segmentation process is based on the target. Each sample data may contain multiple targets, and the signal-to-noise ratios of these targets may differ. Different targets in the same sample data may belong to different evaluation groups. Therefore, after a target is assigned to an evaluation group, other targets in the sample data containing that target are still assigned to a certain evaluation group according to their signal-to-noise ratio. After the segmentation, all other targets in the sample data (including targets already assigned to an evaluation group) are masked.

[0063] S2: Update the training pool using the initial sample set, train the model using the existing sample data in the training pool, and obtain the initially trained model.

[0064] All sample data from the initial sample set are added to the training pool to update (initialize) the training pool. The model is then trained using the existing initial sample set in the training pool to obtain a pre-trained model. The sample data in the initial sample set are all high signal-to-noise ratio samples, which are "easy to learn" and ensure that the model establishes basic cognitive abilities in an "interference-free" environment. The initial training period is 20% of the total training rounds. The model that completes the initial training has a certain detection capability.

[0065] Based on the precise signal-to-noise ratio of the target obtained in S1, only the "simplest" sample with the highest signal-to-noise ratio is selected as the initial training set to ensure that the model establishes basic cognitive abilities in an "interference-free" environment.

[0066] S3: Select the groups to be evaluated in descending order of signal-to-noise ratio. Perform intra-group iterations:

[0067] The groups to be evaluated were divided according to the preset difficulty level criteria. The targets are divided into simple samples and difficult samples; , indicating the sequence number of the group to be evaluated;

[0068] The training pool is updated using simple samples, and the model is trained using the existing sample data in the training pool to obtain the updated model. Based on the difficulty distinction criteria and the updated model's evaluation of difficult samples, the difficult samples are reclassified into simple and difficult samples. S3 is repeated until the preset intra-group iteration termination condition is met, completing the grouping process using samples to be evaluated. Training the model.

[0069] Groups to be evaluated in the candidate pool Sort by target signal-to-noise ratio in descending order. , This indicates the sequence number of the group to be evaluated, i.e., the group to be evaluated within it. High signal-to-noise ratio, grouping to be evaluated The signal-to-noise ratio is low. Groups to be evaluated are selected sequentially in descending order of signal-to-noise ratio, and iterations are performed within each group:

[0070] First, the selected groups to be evaluated... To classify a target as easy or difficult, the following criteria are used in this embodiment of the invention: For a target to be classified as easy or difficult, the average recall and average centroid loss of all targets in the sample data are first calculated. When the average recall of all targets in the sample data is greater than or equal to a preset recall threshold, the target is initially classified as an easy sample. In this embodiment, the recall threshold is 0.8. Specifically, for targets with a signal-to-noise ratio (SNR) lower than 3, another condition must be met to classify the target as an easy sample: the average centroid loss of a target with a SNR lower than the lower threshold (3 in this embodiment) does not exceed the median of the centroid losses of all targets in its sample data group. Otherwise, the target is considered a difficult sample. This classification criterion can be expressed by the following formula:

[0071] ,

[0072] in, This indicates the results of dividing the sample data into easy and difficult categories. This represents the median loss of the centroids of all targets in the sample data. Indicate the goal, Representing simple samples, Indicates a difficult sample. This represents the average recall rate of the target in each sample data point. Indicates the recall threshold. Indicates the lower limit threshold. This represents the mean loss of the target centroid in each sample data point. Indicates the signal-to-noise ratio. Indicates other situations.

[0073] Similarly, in the sample data where the target is located after segmentation, other targets that do not meet the requirements to become simple samples are still masked, but these other targets will also continue to be evaluated as targets to be evaluated.

[0074] The model trained on the initial sample set already possesses preliminary detection capabilities. Then, the simple samples and their pseudo-labels selected in S3 are added to the training pool to update it. The model is then trained using all the sample data in the updated training pool to update the model, while difficult samples are retained. The pseudo-labels are generated during model training and are automatically generated by the model. In this embodiment of the invention, after the training pool is updated, the training cycle is 5 rounds. That is, after each test of difficult and simple samples, new simple samples are added to the training pool, and then training is performed for 5 rounds.

[0075] After a new round of training (training on simple samples), the updated model's detection capability has improved. Therefore, the difficult samples selected in S3 may not be "difficult" for the current model, necessitating intra-group iteration. The difficult samples are reclassified according to the simple / difficult sample classification method from S3. The newly classified simple samples are used to update the training pool again for subsequent model training. The model is updated again, while the difficult samples are retained. Using the updated model, the retained difficult samples are reclassified into simple and difficult samples according to the simple / difficult sample classification method from S3. This iteration continues until the intra-group iteration termination condition is met. Since this iterative process is performed within the same evaluation group, it is called intra-group iteration. The intra-group iteration termination condition is considered met if at least one of the following conditions is met:

[0076] The first condition is that all targets in the current evaluation group have been identified as simple samples and moved into the training pool for subsequent model training; the second condition is that no new simple samples are identified within a preset number of iteration update cycles. In this embodiment of the invention, the preset number of iterations is twice.

[0077] After the iteration within a group is completed, if there are any remaining difficult samples, they are added to the evaluation group in the preparation pool. They will participate in the next group iteration.

[0078] S4: Perform inter-group iteration: Select the groups to be evaluated Repeat the intra-group iteration until the desired result is achieved. One group of training models to be evaluated.

[0079] After the intra-group iterations are completed, inter-group iterations are performed: the next group to be evaluated is selected from the preparation pool in descending order of signal-to-noise ratio. Repeat the same in-group iteration process as S3 to group the objects to be evaluated. Intra-group iteration, group to be evaluated After the intra-group iterations are completed, if there are remaining difficult samples, these remaining difficult samples are added to the evaluation groups in the preparation pool. Participate in the next group iteration. After each group to be evaluated underwent intra-group iteration, inter-group iteration was completed, thus completing the progressive model training and obtaining the trained model.

[0080] It should be noted that after completing the inter-group iterations, a very small number of difficult samples that cannot be identified as simple samples by the model at any stage are discarded to ensure that the final model is not affected by these extreme outliers.

[0081] Although embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

[0082] The specific embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any other corresponding changes and modifications made in accordance with the technical concept of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A progressive active learning weak supervision training method for dim and weak target detection, a model is trained for detecting dim and weak space targets, characterized in that, Comprise: S1: obtaining sample data to form a sample set, the sample data including a target, dividing the sample set into an initial sample set and a to-be-evaluated group according to a signal-to-noise ratio of the target ; S2: update the training pool with the initial sample set, train the model with the existing sample data in the training pool, and obtain a preliminary trained model; S3: select the to-be-evaluated groups in descending order of signal-to-noise ratio , perform intra-group iteration: According to a preset difficulty discrimination criterion, the target in the to-be-evaluated group is divided into a simple sample and a difficult sample. , represents a to-be-evaluated group serial number; the difficulty discrimination criterion is:​ ； wherein, represents the sample data difficulty classification result, represents the sample data, represents the simple sample, represents the difficult sample, represents the average recall rate, represents the recall rate threshold, represents the signal-to-noise ratio lower limit threshold, represents the average centroid loss, represents the median of all target centroid losses of the group sample data, represents the signal-to-noise ratio, represents other cases; The simple sample is used to update the training pool, the existing sample data in the training pool is used to train the model, and an updated model is obtained; the difficult sample is re-divided into a simple sample and a difficult sample according to the difficulty and distinguishability standard and the evaluation of the difficult sample by the updated model; and the step S3 is repeatedly executed until a preset in-group iteration termination condition is met, and the evaluation of the to-be-evaluated group is completed training the model; S4: perform inter-group iteration: Selecting the to-be-evaluated group The intra-group iteration is repeatedly performed until the model training is completed by using to-be-evaluated groups.

2. The progressive active learning weakly supervised training method for dim and weak target detection of claim 1, wherein, The signal-to-noise ratio acquisition method of the target is: According to the target centroid, a two-dimensional Gaussian model of the target is fitted by using a Levenberg-Marquardt optimization algorithm to obtain a FWHM value of the target; a signal-to-noise ratio of the target is obtained by using the FWHM value The calculation formula is: ； wherein, , denotes the measured aperture radius, denotes an empirical coefficient, denotes the measured aperture, denotes the pixel value, denotes the background estimate, denotes the background standard deviation, denotes the number of pixels within the aperture.

3. The progressive active learning weakly supervised training method for dim and weak target detection of claim 1, wherein, The initial sample set and the division method of the to-be-evaluated group are: Set the signal-to-noise ratio threshold, select the target with a signal-to-noise ratio higher than the signal-to-noise ratio threshold from all sample data, and cover other targets in the sample data where the target is located. The obtained sample data set is called the initial sample set; Other targets than the initial sample set are arranged in descending order of signal-to-noise ratio, and these targets are divided into groups to be evaluated, and targets not belonging to the group are masked in the grouped sample data.

4. The progressive active learning weakly supervised training method for dim and weak target detection of claim 1, wherein, If any of the following conditions is met, it is considered to meet the intra-group iteration termination condition: First: all targets in the current to-be-evaluated group have been determined to be simple samples, and the training pool is updated for subsequent training of the model; The second item: in the continuous preset number of iteration update periods, no new simple sample is identified.

5. The progressive active learning weakly supervised training method for dim and weak target detection of claim 1, wherein, After the end of the intra-group iteration, if there are remaining difficult samples, they are added to the evaluation group and participate in the next intra-group iteration.

6. The progressive active learning weakly supervised training method for dim and weak target detection of claim 1, wherein, After completing the inter-group iteration, the remaining difficult samples are discarded.

Images