A method for constructing a low signal-to-noise ratio spatial point source target data set

Abstract

The present application relates to the technical field of image processing, and specifically provides a kind of construction method of low signal-to-noise ratio spatial point source target data set, including obtaining scientific image, according to scientific image, establish signal-to-noise ratio-exposure time physical model to obtain the optimal exposure time corresponding to target signal-to-noise ratio, according to optimal exposure time, collect multiple frames of images to form multiple groups of acquisition sequences, each group of acquisition sequences includes two long exposure frames and a short exposure frame, according to target in long exposure frame, obtain the target label position of short exposure frame by label migration, create annotation mask with target label position as center, and based on the imaging characteristics of short exposure frame and target neighborhood, perform half-height full-width anomaly detection, ellipticity test, background stability test and star orbit crossing detection on generated sample, obtain sample with quality level, and then form data set.The present application method can improve the labeling accuracy, physical authenticity and data set reliability of low signal-to-noise ratio spatial point source target sample.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically providing a method for constructing a low signal-to-noise ratio spatial point source target dataset. Background Technology

[0002] With the rapid development of deep learning technology, object detection networks such as YOLO and Faster R-CNN have demonstrated significant advantages in the field of spatial object detection. Compared with traditional manual feature extraction methods such as morphological filtering and inter-frame differencing, deep learning has a natural advantage in feature extraction and adaptability to complex scenes. However, the performance of deep learning models is highly dependent on the quality and signal-to-noise ratio (SNR) coverage of the training data. As the SNR of the target detection limit continues to decrease (e.g., for faint targets with an SNR < 3), there is a severe lack of low SNR target image datasets available for training, testing, and validation, which has become a core bottleneck restricting the improvement of algorithm performance. Therefore, how to construct a point-like spatial object dataset that covers any SNR, is accurately labeled, and is statistically realistic is a key technical problem that urgently needs to be solved.

[0003] Existing dataset construction techniques mainly fall into two categories: manual annotation and simulation generation, both of which have significant drawbacks.

[0004] In terms of manual annotation, when the target signal-to-noise ratio drops to near the critical detection threshold, the target signal is severely overwhelmed by noise, making it difficult for the human eye to reliably identify and accurately locate it. This results in low annotation efficiency, poor accuracy, and difficulty in ensuring consistency, as well as generating a large amount of mislabeled, missed, and positionally biased "label noise." Furthermore, manual annotation cannot provide a large-scale, finely segmented set of target samples with varying signal-to-noise ratios, making it difficult to construct a training set with a balanced distribution of signal-to-noise ratio and sample size, thus affecting the network's sensitivity to different signal-to-noise ratio ranges.

[0005] While simulations can obtain precise labeled locations, the simulated targets exhibit significant inter-domain differences from real observations in terms of point spread function morphology, noise statistics, and spatiotemporal distribution. Computer simulations struggle to characterize complex physical processes such as detector non-ideal responses (dark current, readout noise, gain inhomogeneity), background non-stationary characteristics (gradual changes in skylight, stellar afterimages), and atmospheric turbulence effects (dynamic changes in PSF due to seeing variations). Models trained with such "overly ideal" synthetic data suffer from feature distribution mismatches in real-world scenarios, severely insufficient generalization ability, and high false alarm rates.

[0006] Existing patented technologies have also failed to effectively solve the above problems. For example, semi-supervised learning schemes rely on the initial model's ability to identify targets with low signal-to-noise ratios, the confidence of false labels cannot be guaranteed, and there is a lack of quantitative modeling of exposure time and signal-to-noise ratio; Unreal Engine schemes are mainly for extended targets with distinguishable shapes, and there is insufficient statistical modeling of noise for point targets; the annotation accuracy of dedicated simulation schemes depends entirely on the accuracy of the simulation model, and there is a lack of statistical consistency verification mechanism with real data. Summary of the Invention

[0007] To address the aforementioned problems, this invention provides a method for constructing a low signal-to-noise ratio spatial point source target dataset. It utilizes the precise location obtained from long-exposure frames acquired using the target's signal-to-noise ratio to calculate the target's location in short-exposure frames, thereby forming training samples and a dataset. This method is low-cost, highly practical, and has high accuracy.

[0008] The method for constructing a low signal-to-noise ratio spatial point source target dataset provided by the present invention includes: S1: Acquire the original image and preprocess it to obtain the scientific image; preprocessing includes bias correction, dark field correction and flat field correction; S2: Based on scientific images, establish a physical model of signal-to-noise ratio and exposure time; S3: Preset multiple target signal-to-noise ratios, and use the signal-to-noise ratio-exposure time model to obtain multiple optimal exposure times; based on the multiple optimal exposure times, acquire multiple frames of images to form multiple acquisition sequences. Each acquisition sequence includes a first long exposure frame, a short exposure frame, and a second long exposure frame. The target signal-to-noise ratios corresponding to the first long exposure frame and the second long exposure frame are both higher than the target signal-to-noise ratio of the short exposure frame. The target in the images acquired in the first and second long exposure frames is obtained, and the centroid consistency test, photometric consistency test, and fitting residual anomaly detection are performed on the target to screen qualified acquisition sequences. S4: Use the first and second long exposure frames as references for short exposure frame label migration: Based on the target's position in the qualified acquisition sequence, calculate the coarse centroid position of the target in the image acquired by the short exposure frame using interpolation, and perform local matching filtering fine-tuning with the coarse centroid position as the center; obtain the fused target centroid position by fusing the centroid positions of the first and second long exposure frames, and correct the fused target centroid position to the time corresponding to the short exposure frame to obtain the target label position of the short exposure frame; S5: Create a label mask centered on the target label location, and perform full width at half maximum (FWHM) anomaly detection, ellipticity test, background stability test, and star trail crossing detection on the generated samples based on the imaging characteristics of short exposure frames and the target neighborhood, generating samples with quality grades. Multiple samples form a low signal-to-noise ratio spatial point source target dataset.

[0009] Preferably, the expression for the signal-to-noise ratio-exposure time physical model is: ; in, This represents the electron production rate of a spatial point source target. Indicates the exposure time. Indicates the signal-to-noise ratio. This represents the total number of photoelectrons generated by the target during the exposure time. Indicates the local sky background rate. Indicates dark current rate. Indicates readout noise. This represents the number of pixels equivalent to noise.

[0010] Preferably, before acquiring the acquisition sequence using the signal-to-noise ratio physical model, a nighttime baseline initialization procedure needs to be performed on the signal-to-noise ratio-exposure time physical model. The nighttime baseline initialization procedure includes: constructing a point spread function template, establishing a seeing baseline and verifying the background estimation algorithm, and obtaining the initial full width at half maximum (FWHM). Using the initial half-height and full width Calculate the initial estimate of the noise equivalent pixel count. : .

[0011] Preferably, the nighttime baseline initialization procedure also includes establishing a seeing benchmark and verifying the background estimation algorithm; the verification includes the spatial consistency of the background mean, the statistical characteristics of the background noise, and the ability to suppress star trail interference.

[0012] Preferred, optimal exposure time Represented as: ; in, This indicates the maximum allowed single-frame exposure time by the system. This indicates the maximum exposure time for the trail effect. Indicates the upper limit of exposure time for saturation. This represents the exposure time obtained by inputting the target signal-to-noise ratio into the signal-to-noise ratio-exposure time physical model.

[0013] The preferred method for verifying centroid consistency is as follows: Calculate the centroid offset between the first long exposure frame and the second long exposure frame. If the following conditions are met... or If the centroids are inconsistent, then it is determined that the centroids are not the same; among them, Represents the confidence coefficient. This indicates the centroid coordinates of the first long exposure frame. This indicates the uncertainty of the first long exposure frame. This indicates the centroid coordinates of the second long exposure frame. This indicates the uncertainty of the second long exposure frame. Indicates system redundancy; Photometric consistency detection involves calculating the target flux between the first long exposure frame and the second long exposure frame. and and relative differences : ; When the relative difference exceeds a preset threshold, it is determined that the luminance between the first long exposure frame and the second long exposure frame is inconsistent. The method for detecting fitting residual anomalies is as follows: structural analysis is performed on the point spread function fitting residual map to detect fitting residual anomalies. The detected abnormal features include linear structures and outlier pixel ratios.

[0014] Preferred, short exposure frames The approximate centroid position of the target in the middle Represented as: , ; in, and These represent the centroid positions of the target in the first and second long exposure frames, respectively. This indicates the exposure center moment corresponding to the optimal exposure time of the first long exposure frame. This indicates the exposure center moment corresponding to the optimal exposure time of the second long exposure frame. Indicates short exposure frame The exposure center moment corresponding to the optimal exposure time.

[0015] Preferably, the local matched filter fine-tuning method is: using the coarse centroid position Centered on a local region, extract the local region within the short-exposure frame, and within this local region, slide a spread function template to calculate the matched filter response; the expression for the matched filter response is: , in, Indicates the search offset. Indicates the first in the template Each pixel is offset relative to the center. This represents the estimated local background value. This represents the image matrix or pixel intensity function of a short-exposure frame. Represents the template of the normalized point spread function In the Weight values ​​at each pixel; Search offset for the maximum response value : ; rough centroid position After fine-tuning .

[0016] Preferably, the centroid position fusion method is to perform inverse variance weighted fusion on the centroid results of the first long exposure frame and the second long exposure frame: ; : in, This indicates the position of the centroid of the fusion target after fusion.

[0017] The preferred method for detecting full width at half maximum (FWHM) anomalies is to examine the FWHM of the minor axis measured in short-exposure frames. Is it within a reasonable range? ; in, Determined by the diffraction limit of the optical system, Determined by extreme seeing based on historical observations; samples exceeding this range are discarded or flagged. The method for testing ellipticity is to calculate the ellipticity of the point spread function. : ; If the ellipticity exceeds a preset threshold, it will be rejected. The background stability test method is as follows: calculate the time series variation coefficient of the background mean; test whether the background noise conforms to the expected Poisson-Gaussian statistical model. Star trail detection is performed using morphological analysis and gradient direction consistency: linear connected regions exceeding the background threshold are detected; the gradient direction of the trailing region is consistent along the trailing direction. If a star trail is detected crossing the target area, the sample is removed or the contaminated area is marked in the annotation mask.

[0018] Compared with the prior art, the present invention can achieve the following beneficial effects: Compared to traditional manual visual annotation and computer simulation generation techniques, this invention utilizes a "long-short-long" encircling acquisition strategy. Using high signal-to-noise ratio (SNR) long-exposure frames as the "teacher," it achieves centroid accuracy better than 0.2 pixels through PSF fitting. Then, through drift interpolation and local matched filtering fine-tuning algorithms, high-quality labels are stably transferred to short-exposure frames with SNR below 3. This enables sub-pixel-level automated annotation and pixel-level mask generation for faint targets, completely solving the bottleneck of "unlabelable" low SNR real data. At the data construction level, this invention establishes an analytical relationship between "exposure time and SNR" based on a physical model. It directly calculates the exposure time required to meet a specified SNR using an inverse formula. Combined with a closed-loop "acquisition-verification-reacquisition" process, it actively fills gaps in the SNR distribution, enabling a balanced dataset covering the entire SNR range to be acquired in a single observation, significantly improving acquisition efficiency and reducing observation costs. Regarding data authenticity, this invention directly generates training samples based on real observation data. Adaptive mask reconstruction technology ensures that the background noise statistical characteristics, PSF morphology, and optical path distortion are completely consistent with the actual physical environment. Furthermore, multi-dimensional quality verification eliminates anomalous samples, resulting in a higher detection rate and significantly lower false alarm rate in real nighttime observation scenarios. In practical applications, this invention forms a complete technical pipeline covering precise time synchronization, standard calibration, automated registration, parameter fitting, and quality verification. Primarily relying on software algorithms, it requires minimal hardware modifications to existing ground-based photoelectric telescopes, exhibiting high compatibility and portability. Full-process automation increases data production efficiency by tens of times, significantly reducing labor costs, demonstrating outstanding engineering practical value and promising prospects for large-scale deployment. Attached Figure Description

[0019] Figure 1 This is a flowchart of a method for constructing a low signal-to-noise ratio spatial point source target dataset according to an embodiment of the present invention. Detailed Implementation

[0020] 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.

[0021] 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.

[0022] 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.

[0023] 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.

[0024] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0025] like Figure 1 As shown, this invention provides a method for constructing a low signal-to-noise ratio spatial point source target dataset, comprising: S1: Acquire the original image and preprocess it to obtain the scientific image; the preprocessing includes bias correction, dark field correction and flat field correction.

[0026] Raw images of celestial bodies and other targets are acquired using image acquisition equipment such as telescopes. To eliminate systematic errors, this embodiment of the invention preprocesses the acquired raw images to obtain usable scientific images. The specific preprocessing method is as follows: First, offset correction is performed on the original image. Offset frames are pre-acquired. Offset frame The acquisition method is as follows: Before observation, multiple frames (typically 20-50 frames) are acquired under shading conditions with zero exposure time (or the shortest allowable exposure time) and the average value is taken to obtain the offset frame. The matrix corresponding to the original image. and pre-acquired bias frames As input, perform pixel-by-pixel subtraction; the output is the bias-corrected image. : .

[0027] Next, dark field correction is performed. Under shaded conditions, multiple frames are captured at scientifically observed exposure times (typically 10-30 frames), and the average value is taken, then the offset frames are subtracted to obtain the dark field template. With dark field template and the image after bias correction The image is used as input for dark-field correction and output as the dark-field corrected image. : , in, Indicates the exposure time of the current frame. This indicates the exposure time of the dark field template.

[0028] Finally, flat-field correction is performed. By imaging a uniform light source (such as twilight skylight or an integrating sphere), and after offset correction and dark-field correction, the image is normalized to a mean of 1 to obtain a flat-field template. With images Peace Field Template As input, pixel-by-pixel division is performed to obtain the preprocessed scientific image. : .

[0029] It should be noted that during flat field correction, if there are pixels with a flat field response lower than the flat field response threshold, the pixel is replaced using interpolation or directly removed. In this embodiment of the invention, the flat field response threshold is set to 0.5.

[0030] S2: Based on scientific images Establish a physical model of signal-to-noise ratio and exposure time.

[0031] Each frame of a scientific image possesses its own signal-to-noise ratio (SNR) and exposure time. Based on these SNR and exposure times, a physical model of SNR-exposure time is established to decouple system parameters from observation conditions, providing a theoretical foundation for subsequent adaptive exposure planning. The expression for the SNR-exposure time physical model is: ; in, The electron generation rate of a spatial point source target is the number of photoelectrons produced by the target on the detector per unit time, expressed in electrons per second. , Indicates the exposure time. Indicates the signal-to-noise ratio. The total number of photoelectrons generated by the target during the exposure time is called the signal term. This represents the local sky background rate, measured in electrons per second per pixel. It reflects the combined effects of lunar phase, sky region, and atmospheric conditions. Dark current rate, measured in electrons per second per pixel. Determined by the detector temperature. This represents readout noise, measured in electrons. This is the inherent noise of the detector's readout circuit. It represents the noise equivalent number of pixels, and the effective number of pixels covered by the target energy distribution. It is affected by atmospheric seeing and the shape of the system point spread function (PSF).

[0032] Solving for the values ​​of each parameter in the signal-to-noise ratio-exposure time physical model expression: In this embodiment of the invention, the parameters in the signal-to-noise ratio-exposure time physical model expression are divided into two categories according to their time-varying characteristics: system static parameters and time-varying environmental parameters. System static parameters include dark current rate. and readout noise These two parameters can be considered constant within a single observation night and are obtained through the calibration procedure before observation. Readout noise. Dark current rate is obtained by statistically analyzing the pixel standard deviation of multiple offset frames. The slope is obtained by linearly fitting a sequence of dark-field frames at different exposure times. This dark-field frame sequence refers to multiple sets of dark-field images acquired at different exposure times, used to estimate the dark current rate. Time-varying environmental parameters include the local sky background rate. Equivalent pixels to noise These two parameters need to be estimated in real-time or near real-time during the observation process. Within the target neighborhood (typically a 50×50 to 100×100 pixel area centered on the target), the sigma shearing iteration method is used to estimate the local background value, iteratively removing values ​​exceeding the specified range. The mean is recalculated after each pixel is counted, typically with 3-5 iterations, and then divided by the exposure time to obtain the local sky background rate. The noise equivalent number of pixels under the current observation conditions is... Full width at half maximum (FWHM) of the point spread function For estimation, for a two-dimensional Gaussian distribution point spread function, The number of noise equivalent pixels during a single exposure time calculation. It remains unchanged and is dynamically updated based on the measurement results of the point spread function in subsequent observed images.

[0033] After solving for the parameters, the dynamic parameters in the signal-to-noise ratio-exposure time physical model that change with the environment are initialized. In this embodiment of the invention, before the formal observation sequence begins, a nighttime baseline initialization procedure is executed to complete the cold start calibration of the dynamic parameters. The specific details are as follows: First, a point spread function template is constructed. Four to seven medium-brightness standard stars are selected within the field of view that evening as initial targets. These standard stars must have a sufficiently high signal-to-noise ratio and not be saturated. Then, multiple frames of images are acquired using a relatively long exposure time (an empirical value, adjusted according to the target brightness). A two-dimensional Gaussian function or Moffat function is fitted to each standard star to obtain the following parameters: one of which is a normalized point spread function template. The fitting result is normalized to make its integral value 1. The point spread function template is equivalent to the standard of the target on the image plane, used for all subsequent matched filtering and metering; the second is the peak energy ratio. The ratio of the response value of the center pixel of the point spread function to the total response value, and the peak energy percentage are important criteria for determining whether the target will saturate; thirdly, the initial half-height and full width. The initial half-height full width serves as the baseline value for seeing that night, directly reflecting the baseline level of atmospheric seeing that night, and is the initial value for subsequent calculation of the noise equivalent pixel number.

[0034] Next, a seeing benchmark is established and the background estimation algorithm is validated. This is done using the initial full width at half maximum (FWHM) values ​​obtained in the previous step. The initial estimate of the noise equivalent number of pixels was calculated. : .

[0035] The initial estimate of the noise equivalent pixel count serves as the starting point for subsequent exposure time planning and signal-to-noise ratio calculation. In actual observations, it is continuously and dynamically updated based on measurements of background stars in newly acquired images. Simultaneously, several blank areas without prominent stars are selected within the field of view, and the local background estimation module is run to verify the following three metrics: Spatial consistency of background mean: Whether the relative deviation of the estimated background mean in different regions is less than a preset threshold, which is preset to 5% in this embodiment; Statistical characteristics of background noise: Whether the pixel residuals of background noise approximately follow a Gaussian distribution to confirm the absence of systematic bias; Star trail interference suppression capability: Whether the Sigma shearing algorithm can effectively suppress interference in regions with star trails. Only when all three verification indicators pass can it be said that the background estimation algorithm for that night is reliable and can be used for subsequent target processing.

[0036] Before formally constructing the dataset, the photon flux level of the target is accurately estimated using a single detection frame, providing a physical basis for subsequently calculating the exposure time required for arbitrary signal-to-noise ratios. The process is as follows: First, for each new target to be observed, a detection frame with a carefully designed exposure time is acquired. Two conditions must be met simultaneously, one of which is , This indicates the saturation upper limit exposure time, which is the maximum allowable exposure time while ensuring that the center pixel of the target does not exceed the full-well capacity of the detector. Exposure Time The exposure time needs to be short enough to ensure the target is not saturated; secondly, the exposure time... The exposure time needs to be long enough to ensure the target signal-to-noise ratio is greater than the detection threshold (typically between 5 and 10) for reliable detection. Too long an exposure time will cause the target to saturate, making metering impossible. If the exposure time... If the exposure time is too short, the target may be completely submerged in noise and cannot be detected. The purpose of this detection frame is to find a balance between these two constraints, aiming to obtain an initial observation that is "clear" but not "overwhelmed" by the target. In practice, a conservative initial exposure time can be estimated based on the target's predicted magnitude and the current system status. If the target is not detected, the exposure time is gradually increased and the data is collected again.

[0037] Secondly, after obtaining the probe frame image, a pre-constructed normalized point spread function template is used. and local background estimation values The point spread function weighted photometry method is used to estimate the photon flux of the target. The calculation formula is as follows: , in, Indicates pixel index, The template for the normalized point spread function is shown at the th... Weight values ​​at each pixel Indicates the first The signal value of each pixel, that is, the preprocessed digital signal or the converted number of electrons. Indicates the first Local background estimation value at each pixel Indicates the first The noise variance of each pixel can be calculated according to... Estimate the value of the noise variance. This indicates the system gain, measured in electrons per digital unit (e- / ADU), and is used for the conversion between the detector output digital signal and the number of electrons.

[0038] The summation range is centered on the target and has a radius of 2. A circular or rectangular region with a half-height to full width of 3. The estimated target electron production rate is: .

[0039] Target electron production rate estimate Used as the target electron generation rate in the signal-to-noise ratio-exposure time physical model The estimated value, i.e., the value used when planning subsequent exposure times. Or with As Measurement estimates.

[0040] Finally, the estimated value of the calculated photon flux was analyzed. Cross-validation is performed to ensure its reliability. The specific content of cross-validation includes: first, comparing the results with aperture measurement results. Independent aperture measurement is performed using a fixed aperture. If the relative deviation between the results of the two methods (point spread function weighted photometry and fixed aperture photometry) is less than a preset threshold, the flux estimation is considered reliable. That is, the point spread function weighted photometry result is compared with the fixed aperture photometry result; if the relative deviation is less than a preset threshold, the flux estimation is considered reliable. In this embodiment of the invention, the fixed aperture value is... The first step is to check the full width at half maximum (FWHM) of the photon flux, with a preset threshold of 10%. The second step is to check for linear structures (indicating star trails) in the residual plot of the spread function fitting at the checkpoint, or for abnormal bright spots in the background estimation region. If these are found, the frame is marked as a suspicious sample requiring manual review. If the problem persists after review, it is confirmed as an abnormal sample and removed. After removal, the sample may be re-acquired. Although the signal-to-noise ratio of the probe frame is relatively high, it can still be affected by various interference factors. If the accuracy of the photon flux estimation is low, it will cause errors in all subsequent exposure times calculated based on that photon flux estimate. Therefore, abnormal samples are removed to avoid incorrect photon flux estimates affecting subsequent calculations.

[0041] Given target signal-to-noise ratio Substitute the signal-to-noise ratio-exposure time physical model to solve for the required exposure time. Target signal-to-noise ratio Substitute the signal-to-noise ratio-exposure time physical model: .

[0042] Expanding and rearranging it yields the following expression: .

[0043] The positive root of this quadratic equation is used as the exposure time. .

[0044] The exposure time obtained by solving This is only a theoretical value; saturation and trailing constraints need to be calculated to obtain the final exposure time. Saturation upper limit exposure time. The calculation method is as follows: , in, This represents a safety factor, typically ranging from 0.7 to 0.8, used to provide a margin to account for errors in the estimated photon flux and momentary brightening. This represents the percentage of peak energy in the point spread function, obtained from calibration during the initialization phase. This represents the detector's full-well capacity, i.e., the maximum number of electrons a single pixel can hold, determined by the detector specifications. Saturation constraint calculations can prevent saturation of the target's center pixel.

[0045] Calculate the maximum exposure time for trailing effect : , in, This represents the allowable trailing length as a percentage of the full width at half height, typically ranging from 0.3 to 0.5. This represents the residual angular velocity of the target on the image plane, measured in pixels per second. It needs to be calculated based on the target's predicted orbit and the telescope's tracking mode. Specifically, when the telescope's tracking mode is set to star tracking mode... This is equal to the angular velocity of the target relative to the stellar background, when the telescope's tracking mode is target tracking mode. This is equivalent to tracking the residual. This represents the full width at half maximum (FWHM) of the point spread function under the current seeing conditions. Calculating the maximum exposure time for the trailing effect ensures that motion blur of the target on the image plane does not exceed an acceptable range.

[0046] After the above saturation constraint calculation and trailing constraint calculation, the final exposure time is obtained. Final exposure time The minimum value of all constraints is expressed as: , in, This indicates the maximum single-frame exposure time allowed by the system, which is determined by the telescope control system or storage bandwidth limitations.

[0047] The final exposure time output after constraint calculation is the optimal exposure time, along with the expected signal-to-noise ratio corresponding to the optimal exposure time. Specifically, when... If the signal-to-noise ratio cannot be reached due to saturation or tailing constraints, then record the limiting factor and consider adjusting the observation strategy, such as using a target tracking mode to reduce the residual angular velocity.

[0048] S3: Preset multiple target signal-to-noise ratios, and use the signal-to-noise ratio-exposure time model to obtain multiple optimal exposure times; based on the multiple optimal exposure times, acquire multiple frames of images to form multiple acquisition sequences. Each acquisition sequence includes a first long exposure frame, a short exposure frame, and a second long exposure frame. The target signal-to-noise ratios corresponding to the first long exposure frame and the second long exposure frame are both higher than the target signal-to-noise ratio of the short exposure frame; acquire the targets in the images acquired by the first long exposure frame and the second long exposure frame, and perform centroid consistency checks, photometric consistency checks, and fitting residual anomaly detection on the targets to screen qualified acquisition sequences.

[0049] To obtain multiple optimal exposure times for different target signal-to-noise ratios, multiple frames of images are acquired to form multiple acquisition sequences. Each acquisition sequence consists of three frames, which are arranged in chronological order as the first long exposure frame. Short exposure frames Second long exposure frame In this process, the first and second long exposure frames have the same exposure parameters to reduce the impact of exposure parameter differences on the consistency check of the two long exposure frame images, thus facilitating subsequent consistency checks. Specifically, the first long exposure frame... The corresponding optimal exposure time is expressed as The start time of data collection is expressed as Short exposure frames The corresponding optimal exposure time is expressed as The start time of data collection is expressed as Second long exposure frame The corresponding optimal exposure time is expressed as The start time of data collection is expressed as Optimal exposure time and These are obtained according to the method in S2, and the two correspond to the signal-to-noise ratio (SNR) settings for high SNR targets and low SNR targets, respectively. In this embodiment of the invention, the SNR of the high SNR target... Not less than 30 Signal-to-noise ratio of low signal-to-noise ratio targets Less than 3. Three frames should be acquired consecutively within the shortest possible time interval to minimize inter-frame field-of-view drift. The inter-frame interval is primarily determined by detector readout time, data transmission time, and shutter response time, typically ranging from 0.1 to 2 seconds. During acquisition, the telescope's pointing remains constant, without re-alignment or guide correction. Simultaneously, the following time information must be accurately recorded: exposure start time (system timestamp of shutter opening or integration start); exposure end time (system timestamp of shutter closing or integration end); and exposure center time. , This is used for subsequent drift interpolation calculations.

[0050] Multiple consistency and anomaly checks are performed on the first and second long exposure frames, including centroid consistency check, photometric consistency check, and fitting residual anomaly detection. The specific process is as follows: First, robust point spread function fitting is performed on the first and second long exposure frames respectively to obtain the target in the first and second long exposure frames, and the centroid coordinates and their uncertainties are calculated. The centroid coordinates of the first long exposure frame are expressed as follows: The uncertainty is expressed as The centroid coordinates of the second long exposure frame are represented as follows: Uncertainty representation The fitting method can be a two-dimensional Gaussian function least squares fitting or an iterative centroid method. Uncertainty is obtained from the fitted covariance matrix or through Monte Carlo simulation.

[0051] Centroid inconsistency can occur due to factors such as telescope jitter, target obstruction, and failure to fit the point spread function. Therefore, a centroid consistency check is necessary: ​​calculate the centroid offset between the first and second long exposure frames and compare it with a statistical significance threshold. If the following conditions are met... or If the centroids are inconsistent, then it is determined that the centroids are not the same. This represents the confidence coefficient, typically with a value of 3, corresponding to a confidence level of approximately 99.7%. This represents system redundancy, used to absorb systematic deviations such as timestamp errors and tracking jitter; a typical value is 0.05-0.1 pixels. Samples with inconsistent centroids should be discarded or marked for review. If the review still fails, the sample should be discarded.

[0052] When the target is obscured by thin clouds, has a stellar trail crossing the target area, or exhibits changes in its own light intensity, inconsistencies in light intensity may occur. Therefore, light intensity consistency detection is necessary: ​​calculate the target flux between the first long exposure frame and the second long exposure frame. and and relative differences : .

[0053] When calculated When the intensity exceeds a preset threshold, it is determined that the luminance between the first long exposure frame and the second long exposure frame is inconsistent. In this embodiment of the invention, the preset threshold value is... .

[0054] Structural analysis was performed on the residual plot of the point spread function to detect anomalies in the fitting residuals. Detected anomalies included linear structures and outlier pixel proportions. Linear structures were identified by the presence of continuous positive or negative residual bands along a certain direction, indicating that stellar tails traversed the target region. Outlier pixel proportions were identified by residuals exceeding [a certain threshold]. If the pixel ratio exceeds a preset threshold, it is marked as abnormal. The method for detecting abnormal features in this embodiment of the invention is to use Hough transform to detect straight lines or perform one-dimensional projection analysis along different directions, and set the preset threshold to 5%.

[0055] Only when the first and second long exposure frames simultaneously pass all the above checks is the acquired sequence considered a qualified acquisition sequence and eligible for subsequent label migration processing. Otherwise, appropriate handling measures are taken based on the specific cause of failure. If the abnormal feature is temporary, such as due to momentary jitter, re-acquisition is attempted; if the abnormal feature is persistent, the abnormality type is recorded for subsequent dataset quality analysis. The output after the above checks is the set of valid sequences that have passed the gating and the quality assessment parameters for each sequence group, i.e., the qualified acquisition sequences.

[0056] S4: Use the first and second long exposure frames as references for short exposure frame label migration: Based on the target's position in the qualified acquisition sequence, calculate the coarse centroid position of the target in the image acquired by the short exposure frame using interpolation, and perform local matching filtering fine-tuning with the coarse centroid position as the center; obtain the fused target centroid position by fusing the centroid positions of the first and second long exposure frames, and correct the fused target centroid position to the time corresponding to the short exposure frame to obtain the target label position of the short exposure frame.

[0057] Extract the following parameters from the timestamps of the aforementioned records: First long exposure frame The center moment of the corresponding optimal exposure time Second long exposure frame The center moment of the corresponding optimal exposure time Short exposure frames The center moment of the corresponding optimal exposure time Due to the acquisition time of the acquisition sequence, the above parameters satisfy... .because The value of is extremely small, typically lasting only a few seconds or tens of seconds. Since telescope tracking errors are approximately uniform over short periods, and atmospheric refraction changes are negligible, while the target's own motion (such as satellite orbital motion) can be linearized over short periods, the field-of-view drift can be approximated as linear within this timescale. For high-precision applications, the applicability of the linear assumption can be verified by analyzing the drift residuals in historical data. If the residuals significantly exceed the positioning accuracy requirements, a quadratic polynomial or spline interpolation model can be considered.

[0058] Based on the aforementioned linear field-of-view drift, interpolation is used to calculate the short exposure frame. The approximate centroid position of the target in the middle : , .

[0059] in, and These represent the centroid positions of the target in the first and second long exposure frames, respectively.

[0060] The coarse centroid position obtained by interpolation is estimated using the error propagation formula. uncertainty : , in, Indicates the interpolation weight factor. , and Let represent the standard uncertainty of the target centroid's x-coordinate in the two long exposure frames, and let their squares represent the corresponding variances. When the accuracy of the target centroid localization results in the first and second long exposure frames is comparable, for typical center interpolation cases, , Because the first and second long exposure frames have a high signal-to-noise ratio, their centroid positioning accuracy can reach sub-pixel level (typical value 0.02). 0.1 pixels (depending on the signal-to-noise ratio and full width at half maximum), so the interpolation position also has subpixel accuracy.

[0061] Interpolation calculates a coarse centroid position, and its accuracy is improved through fine-tuning with local matched filtering and position fusion. The local matched filtering fine-tuning process is as follows: With a rough center of mass position Centered on a local region, a local region is extracted within a short-exposure frame. Within this local region, a sliding point spread function template is used to calculate the matched filter response. In this embodiment of the invention, the size of the local region is... Pixel to Between pixels. The expression for the matched filter response is: , in, Indicates the search offset. Indicates the first in the template Each pixel is offset relative to the center. This represents the estimated local background value. This represents the short-exposure frame image matrix. Represents the coordinates in a short exposure frame Pixel value at that location, Represents the template of the normalized point spread function In the Weight values ​​at each pixel.

[0062] Within the allowed search range (±1 pixel in this embodiment), find the search offset corresponding to the maximum response value. : .

[0063] rough centroid position After fine-tuning This fine-tuning process is equivalent to maximum likelihood estimation guided by high-confidence priors, which can correct minor system biases while maintaining interpolation accuracy. If the signal-to-noise ratio of short-exposure frames is too low, resulting in no obvious peak in the response map, the original interpolation position is retained without fine-tuning.

[0064] To further improve positioning accuracy, the centroid results of the first and second long exposure frames are fused using inverse variance weighting: , , The position after fusion The centroid position of the fusion target is used as the reference point. When the accuracy of the centroid positioning results in the first and second long exposure frames is comparable, this fusion strategy can improve the positioning accuracy by approximately [missing information]. times.

[0065] A preset position error threshold, such as 0.05-0.1 pixels, is used to determine whether time correction should be performed based on the time deviation between the exposure center time of the short exposure frame and the average exposure center time of two long exposure frames, as well as the target's movement speed on the image plane. If the value is greater than the position error threshold, time correction should be performed; if the value is less than the position error threshold, time correction is not required.

[0066] If it is necessary to correct the centroid position of the fused target to the time corresponding to the short exposure frame, rather than the average time of the long exposure frame, then time correction should be performed according to the following formula: , .

[0067] The above correction is especially important when the target is moving at a high speed. The corrected target label position is shown in the output short-exposure frame. and its uncertainty estimation ,in, and These represent the uncertainties in the x and y directions of the target label position, respectively, which can be estimated by the centroid fitting error of the two long exposure frames, the time interpolation error, the local matched filter fine-tuning error, and the system redundancy error.

[0068] S5: Create a label mask centered on the target label location, and perform full width at half maximum (FWHM) anomaly detection, ellipticity test, background stability test, and star trail crossing detection on the generated samples based on the imaging characteristics of short exposure frames and the target neighborhood, generating samples with quality grades. Multiple samples form a low signal-to-noise ratio spatial point source target dataset.

[0069] Target label location While these are merely point locations, the low signal-to-noise ratio spatial point source target dataset generated in this embodiment of the invention is used for model training; therefore, its final annotation should be regions, not points. An annotation mask is created centered on the target label location. Because the point spread function of long-exposure frames is affected by the cumulative effect of atmospheric turbulence over time, its full width at half maximum (FWHM) is usually larger than that of short-exposure frames, a phenomenon known as the "blurring effect." Therefore, the scale of the annotation mask is determined by the instantaneous seeing of the short-exposure frame. The target's neighborhood (typically within 100 km of the target) is defined in the short-exposure frame. 100 pixels to 200 Within a 200-pixel area, select 3-5 background stars. These background stars must meet the following conditions: a sufficiently high signal-to-noise ratio (SNR) to ensure reliable fitting (in this embodiment, the SNR is higher than 20); unsaturated (pixel value far below the full-well capacity); and isolated (no other bright sources interfering with the background star). Perform a two-dimensional Gaussian fit on the selected background stars, measure their full width at half maximum (FWHM), and take the mean as the instantaneous seeing estimate for the short exposure frame. .

[0070] If the background stars are not perfectly circular but elliptical, it indicates a trailing effect, i.e., a tracking error exists. In this embodiment of the invention, when the ratio of the major axis to the minor axis of the fitted background star is greater than 1.2, the fitted result is considered to be a distinct ellipse. In this case, the full width at half maximum (FWHM) along the minor axis is taken as the estimated seeing value to reflect the true contribution of atmospheric seeing rather than the tracking error.

[0071] Ellipse fitting can be achieved using a two-dimensional elliptic Gaussian function: , in, Representing coordinates Pixel intensity at that location This represents the magnitude of the elliptic Gaussian function. Indicates local background offset. Let represent the standard deviations of the elliptic Gaussian function along the major and minor axes, respectively. This represents the fitted center coordinates of the background stars.

[0072] Then the half-height and full width of the minor axis .

[0073] The fused position is obtained by inverse variance weighted fusion of the centroid results of two long-exposure frames. , in the position after fusion As the center of the annotation mask, a two-dimensional Gaussian soft mask matrix is ​​generated. : , in, This represents the pixel coordinates in the labeled mask matrix or short-exposure frame image. This represents the standard deviation of the labeled mask. The size of the soft mask matrix is ​​[value to be filled in]. Or larger, to ensure the main energy distribution of the spread function at the coverage point. A soft mask is a form of labeled mask whose matrix elements are continuous values, usually normalized to [0,1].

[0074] The annotation mask is normalized, and the output is an annotation mask matrix paired with the short exposure frame. Annotation mask normalization can be implemented using any of the following methods: Peak value normalization: This makes the center pixel value of the annotation mask 1; Integral normalization: , so that the sum is 1; Binarization: , The threshold (in this embodiment, the value is half of the peak value) is used to generate a hard mask.

[0075] To evaluate the quality of samples generated from short-exposure frame images, target label locations, and annotation masks, the generated samples underwent full-width / half-height anomaly detection, ellipticity testing, background statistical stability testing, and star trail crossing detection. Specifically: Test the full width at half maximum (FWHM) of the minor axis measured in short exposure frames. Is it within a reasonable range? , in, Determined by the diffraction limit of the optical system, The extreme seeing values ​​are determined by historical observation statistics. Samples exceeding this range should be discarded or flagged. Here, "sample" refers to a complete sample record in the low signal-to-noise ratio (FWHM) spatial point source target dataset, not a single labeled mask. Each complete data record typically includes a pre-processed short-exposure frame image, target label location, labeled mask, uncertainty, exposure time, acquisition time, seeing estimate, and quality grade. If the FWHM is abnormal, the entire complete data record should be discarded or flagged.

[0076] Calculate the ellipticity of the point spread function : , If the ellipticity exceeds the preset threshold (typically 0.3), it indicates a serious image quality anomaly, including excessive tracking error, optical system misalignment, etc., and should be rejected.

[0077] Samples with unstable backgrounds may be affected by factors such as cloud cover and stray light, reducing the reliability of annotation. Therefore, a background statistical stability test is required. The background statistical stability test verifies whether the background estimate of the target neighborhood is stable: the time series coefficient of variation of the background mean (multiple observations of the same target) is calculated, and the time series coefficient of variation is compared with a preset stability threshold. If the coefficient of variation does not exceed the preset stability threshold, the background is considered stable; otherwise, the background is considered to have abnormal fluctuations, and the sample quality level is reduced or the sample is removed. The test also verifies whether the background noise conforms to the expected Poisson-Gaussian statistical model: the statistical results of the background noise in the target neighborhood are compared with the preset Poisson-Gaussian statistical model; if the deviation does not exceed the preset noise threshold, the background noise is considered normal; otherwise, the background noise is considered abnormal, and the sample quality level is reduced or the sample is removed depending on the degree of abnormality.

[0078] The system detects the presence of star trails crossing the target neighborhood. Detection methods include: morphological analysis: detecting linear connected regions and identifying them as candidate star trail tail regions; where exceeding a background threshold indicates that the grayscale response of the linear connected region is significantly higher than normal background fluctuations; gradient direction consistency: calculating the principal directions of each local segment of the candidate star trail tail region; when the deviation between the principal directions of each local segment does not exceed a preset direction threshold, it indicates that the candidate star trail tail region extends continuously along a fixed direction and is determined to be a star trail tail; otherwise, it is determined to be random noise, stray light, or other non-star trail interference. If a star trajectory is detected crossing the target region, the sample is either removed or the contaminated area is marked in the annotation mask, depending on the severity.

[0079] Based on the above quality control results, each generated sample was assigned a quality level, including Level A, Level B, Level C, and Level D. Level A indicates that all tests were passed and the sample can be directly used for model training; Level B indicates that there are minor anomalies that do not affect the reliability of the labels and the sample can be used for training but requires monitoring in post-processing; Level C indicates that there are significant anomalies and the sample can only be used for testing or requires manual review; Level D indicates severe anomalies and the sample is discarded.

[0080] The results of all the aforementioned processing steps, including preprocessed short-exposure frame images, target centroid coordinates and uncertainties obtained through cross-exposure label migration, and soft / hard mask matrices generated based on the instantaneous seeing of the short exposure, along with metadata such as exposure time, acquisition time, seeing estimate, and quality grade, are organized in a structured format to form each traceable and valid sample record. Subsequently, following standard machine learning procedures, all samples are divided into training, validation, and test sets (typically in a 70%:15%:15% ratio) based on a balanced distribution of key parameters such as signal-to-noise ratio and seeing. This ensures that observations of the same target at different times do not simultaneously appear in both the training and test sets to avoid data leakage. Finally, a standardized low signal-to-noise ratio spatial point source target dataset with quality grade labels is output, which can be directly used for training deep learning target detection models.

[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 method for constructing a low signal-to-noise ratio spatial point source target dataset, characterized in that, include: S1: Acquire the original image and preprocess the original image to obtain a scientific image; the preprocessing includes bias correction, dark field correction and flat field correction; S2: Based on the scientific images, establish a signal-to-noise ratio-exposure time physical model; S3: Preset multiple target signal-to-noise ratios, and use the signal-to-noise ratio-exposure time model to obtain multiple optimal exposure times; based on the multiple optimal exposure times, acquire multiple frames of images to form multiple acquisition sequences, each acquisition sequence including a first long exposure frame, a short exposure frame and a second long exposure frame, the target signal-to-noise ratios corresponding to the first long exposure frame and the second long exposure frame are both higher than the target signal-to-noise ratio of the short exposure frame; The target in the images acquired by the first long exposure frame and the second long exposure frame is obtained, and the centroid consistency test, photometric consistency test and fitting residual anomaly detection are performed on the target to screen qualified acquisition sequences; S4: Using the first long exposure frame and the second long exposure frame as references for the label migration of the short exposure frame, perform label migration: Based on the position of the target in the qualified acquisition sequence, calculate the coarse centroid position of the target in the image acquired by the short exposure frame using interpolation, and perform local matching filtering fine-tuning with the coarse centroid position as the center; obtain the fused target centroid position by fusing the centroid positions of the first long exposure frame and the second long exposure frame, and correct the fused target centroid position to the time corresponding to the short exposure frame to obtain the target label position of the short exposure frame; S5: Create a label mask centered on the target label location, and perform full width at half maximum (FWHM) anomaly detection, ellipticity test, background stability test, and star trail crossing detection on the generated samples based on the imaging characteristics of the short exposure frame and the target neighborhood, generating samples with quality grades. Multiple samples form a low signal-to-noise ratio spatial point source target dataset.

2. The method for constructing a low signal-to-noise ratio spatial point source target dataset as described in claim 1, characterized in that, The expression for the signal-to-noise ratio-exposure time physical model is: ； in, This represents the electron production rate of a spatial point source target. Indicates the exposure time. Indicates the signal-to-noise ratio. This represents the total number of photoelectrons generated by the target during the exposure time. Indicates the local sky background rate. Indicates dark current rate. Indicates readout noise. This represents the number of pixels equivalent to noise.

3. The method for constructing a low signal-to-noise ratio spatial point source target dataset as described in claim 1, characterized in that, Before acquiring the acquisition sequence using the aforementioned signal-to-noise ratio physical model, a nighttime baseline initialization procedure needs to be performed on the signal-to-noise ratio-exposure time physical model. This nighttime baseline initialization procedure includes: constructing a point spread function template, establishing a seeing baseline and validating the background estimation algorithm, and obtaining the initial full width at half maximum (FWHM). Using the initial half-height full width Calculate the initial estimate of the noise equivalent pixel count. : 。 4. The method for constructing a low signal-to-noise ratio spatial point source target dataset as described in claim 3, characterized in that, The nighttime baseline initialization procedure also includes establishing a seeing benchmark and verifying the background estimation algorithm; the verification includes the spatial consistency of the background mean, the statistical characteristics of background noise, and the ability to suppress star trail interference.

5. The method for constructing a low signal-to-noise ratio spatial point source target dataset as described in claim 1, characterized in that, The optimal exposure time Represented as: ； in, This indicates the maximum allowed single-frame exposure time by the system. This indicates the maximum exposure time for the trail effect. Indicates the upper limit of exposure time for saturation. This represents the exposure time obtained by inputting the target signal-to-noise ratio into the signal-to-noise ratio-exposure time physical model.

6. The method for constructing a low signal-to-noise ratio spatial point source target dataset as described in claim 1, characterized in that, The centroid consistency verification method is as follows: Calculate the centroid offset between the first long exposure frame and the second long exposure frame. If the following conditions are met... or If the centroids are inconsistent, then it is determined that the centroids are not the same; among them, Represents the confidence coefficient. This indicates the centroid coordinates of the first long exposure frame. This indicates the uncertainty of the first long exposure frame. This indicates the centroid coordinates of the second long exposure frame. This indicates the uncertainty of the second long exposure frame. Indicates system redundancy; The photometric consistency detection involves calculating the target flux between the first long exposure frame and the second long exposure frame. and and relative differences : ； When the relative difference exceeds a preset threshold, it is determined that the luminance between the first long exposure frame and the second long exposure frame is inconsistent. The method for detecting fitting residual anomalies is as follows: structural analysis is performed on the point spread function fitting residual map to detect fitting residual anomalies. The detected abnormal features include linear structures and outlier pixel ratios.

7. The method for constructing a low signal-to-noise ratio spatial point source target dataset as described in claim 1, characterized in that, The short exposure frame The approximate centroid position of the target in the middle Represented as: ， ； in, and These represent the centroid positions of the target in the first and second long exposure frames, respectively. This indicates the exposure center time of the optimal exposure time corresponding to the first long exposure frame. This indicates the exposure center time of the optimal exposure time corresponding to the second long exposure frame. Indicates short exposure frame The exposure center moment corresponding to the optimal exposure time.

8. The method for constructing a low signal-to-noise ratio spatial point source target dataset as described in claim 1, characterized in that, The local matched filter fine-tuning method is as follows: using the coarse centroid position... Centered on a local region, extract the local region within the short-exposure frame, and within this local region, slide a spread function template to calculate the matched filter response; the expression for the matched filter response is: ， in, Indicates the search offset. Indicates the first in the template Each pixel is offset relative to the center. This represents the estimated local background value. This represents the image matrix or pixel intensity function of a short-exposure frame. Represents the template of the normalized point spread function In the Weight values ​​at each pixel; Search offset for the maximum response value : ； The approximate centroid location After fine-tuning .

9. The method for constructing a low signal-to-noise ratio spatial point source target dataset as described in claim 1, characterized in that, The centroid position fusion method involves performing inverse variance weighted fusion on the centroid results of the first long exposure frame and the second long exposure frame. ； ： in, This indicates the position of the centroid of the fusion target after fusion.

10. The method for constructing a low signal-to-noise ratio spatial point source target dataset as described in claim 1, characterized in that, The full width at half maximum (FWHM) anomaly detection method is as follows: Examine the full width at half maximum (FWHM) of the minor axis measured in the short exposure frame. Is it within a reasonable range? ； in, Determined by the diffraction limit of the optical system, Determined by extreme seeing based on historical observations; samples exceeding this range are discarded or flagged. The ellipticity test method is as follows: calculate the ellipticity of the point spread function. : ； If the ellipticity exceeds a preset threshold, it will be rejected. The background stability test method is as follows: calculate the time series variation coefficient of the background mean; test whether the background noise conforms to the expected Poisson-Gaussian statistical model; The star trail detection is performed using morphological analysis and gradient direction consistency: linear connected regions exceeding the background threshold are detected; the gradient direction of the trailing region is consistent along the trailing direction. If a star trail is detected crossing the target area, the sample is removed or the contaminated area is marked in the annotation mask.

Images